Who publicly predicted what and when about the future of developers and work with AI — with sources, deadlines and verification rules.
Hall navigator
Upcoming — the stated deadline has not yet arrived.
Passed — the deadline has arrived; this does not automatically produce a verdict.
No deadline — the author gave no verifiable deadline.
01
The deadline has passed, no verdict has been reached
Exhibit No. 004Deadline has passed · metric not defined
Dario Amodei
CEO and co-founder of Anthropic
Recorded wording
In 3–6 months, AI will write about 90% of the code, and in about a year, almost all the code.
Announced
Deadline
What we measure
Code share
Source
Official transcript of the event
The outcome review has not yet been prepared. A passed deadline alone is not treated as proof or disproof.
Editorial analysis
Meaning, limitations and future assessment protocol
What it means
The forecast concerns code writing, but does not determine the denominator: lines, changes, new code, accepted suggestions, or completed tasks.
Interpretation boundary
Even 90% of the generated lines do not show who set the task, checked the result, integrated the change and is responsible for the failure.
Verification protocol
A pre-agreed metric is needed for a representative sample of companies. Without it, after the deadline, you can only check individual teams, but not the industry.
Exhibit No. 005Deadline has passed · criterion not specified
Mark Zuckerberg
CEO Meta
Recorded wording
Probably, as early as 2025, Meta and other companies will have AI capable of being “something like a mid-level engineer” and writing code.
Announced
Deadline
What we measure
Level of engineering tasks
Source
Full podcast recording
The outcome review has not yet been prepared. A passed deadline alone is not treated as proof or disproof.
Editorial analysis
Meaning, limitations and future assessment protocol
What it means
This is a verifiable period, but there was no public set of responsibilities and tests for the “middle level” in the statement.
Interpretation boundary
The ability of a model to write a separate module is not equal to the independent work of an engineer in a live code base.
Verification protocol
Record a set of tasks for a mid-level engineer: unclear requirement, changes in the repository, tests, code review, deployment and responsibility for the result.
Exhibit No. 076Deadline has passed · criterion not specified
Sam Altman
CEO OpenAI
Recorded wording
In 2025, the first AI agents could “enter the workforce” and significantly change the output of companies.
Announced
Deadline
What we measure
The emergence of agents in the work process and material changes in companies' output
Source
Author's publication
The outcome review has not yet been prepared. A passed deadline alone is not treated as proof or disproof.
Editorial analysis
Meaning, limitations and future assessment protocol
What it means
The formulation connects two events: the agent must participate in the actual operation of the company, and the change in its output must be significant.
Interpretation boundary
The launch of a product called "agent", a pilot, or a single task successfully completed does not prove a material change in the company's output.
Verification protocol
Determine the autonomy and duration of the agent, the release baseline and the minimum effect size. We need company data going back to 2025, not just supplier demos.
Founder and head of Stability AI at the time of the announcement
Recorded wording
“In five years there will be no programmers” - in the full conversation the author clarifies: “in the usual sense.”
Announced
Deadline
What we measure
People and profession
Source
Initial video interview
Editorial analysis
Meaning, limitations and verification criterion
What it means
The wording is deliberately hyperbolic. Verifiable strong version - by the specified date, the separate paid programmer role should practically disappear or radically change.
Interpretation boundary
The growth of AI code in itself does not mean the disappearance of people who design, test and are responsible for the system.
Verification protocol
Compare the number of employed developers and vacancies in the summer of 2028 with 2023. Proactively separate in-house engineers from the people who sometimes create programs using AI.
The median forecast gave a 50% probability of machine superiority in all tasks by 2047, but full automation of all professions only by 2116.
Announced
Deadline
What we measure
Probability of opportunities and automation of professions
Source
Great Scientific Poll
Editorial analysis
Meaning, limitations and verification criterion
What it means
The survey shows a large gap between the ability of a machine to complete all tasks and the economic automatability of all professions. The 10% threshold for these events is 2027 and 2037.
Interpretation boundary
The median year is not a consensus promise and does not mean that the probability of an event occurring before it is zero. Expert long-range forecasts may be poorly calibrated.
Verification protocol
Maintain the original question definitions and response distributions, update with repeat surveys, and do not declare a probabilistic forecast false after one missed deadline.
By 2030, there will be 57.5 million Indian developers and 54.7 million US developers on GitHub; India will come out on top.
Announced
Deadline
What we measure
GitHub developer accounts by country
Source
Official platform report
Editorial analysis
Meaning, limitations and verification criterion
What it means
GitHub used the average of five growth models. The forecast describes the geography of a particular platform's users and complements a broader museum exhibit about a possible billion software creators.
Interpretation boundary
A GitHub user is not necessarily a programmer. The number of accounts cannot be directly translated into employment, salaries, or the number of full-time developers in the country.
Verification protocol
In 2030, compare GitHub's published India and US figures to the original 57.5 million and 54.7 million, ensuring that the definition of developer and the way country is defined remain constant.
The singularity will occur in 2045; The non-biological intelligence created by this moment will be a billion times more powerful than the total human intelligence of the beginning of the 21st century.
Announced
Deadline
What we measure
The onset of the author’s “singularity” and the stated ratio of the power of non-biological and human intelligence
Source
Copyright materials for the book
Editorial analysis
Meaning, limitations and verification criterion
What it means
The forecast connects a calendar date with a pattern of accelerating technological progress. This is a much stronger statement than the appearance of one human level in a separate problem.
Interpretation boundary
The growth of computing power, the result of a single benchmark, or the popularity of AI do not prove either singularity or billion-dollar superiority of intelligence.
Verification protocol
Before 2045, identify in advance observable signs of loss of ability for an unamplified person to track technological progress and a method for comparing combined human and machine intelligence.
Computing technologies must reach a state where they do not need to be programmed in the traditional way: human language becomes the programming language, and everyone becomes a programmer.
Announced
Deadline
No calendar deadline
What we measure
Percentage of people who can create working programs in natural language without knowledge of a traditional programming language
Source
Official recording of a public speech
Editorial analysis
Meaning, limitations and verification criterion
What it means
The phrase expands the concept of a programmer to any person who formulates a task for a machine. She predicts a change in the interface for creating programs, and not literally everyone becoming a developer.
Interpretation boundary
Generating a short script upon request does not mean that the user can independently design, test, secure and maintain a full-fledged system.
Verification protocol
Determine a typical finished software product and check what proportion of people without training are able to create and safely modify it only in natural language.
There is about a 50% chance that DeepMind's definition of AGI will appear in the next 5–10 years. Such an AGI must have the full range of human cognitive abilities.
Announced
Deadline
What we measure
Probabilistic forecast of the emergence of a system with the entire set of human cognitive abilities
Source
Direct interview
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is not a promise that AGI will definitely arrive by 2035, but a 50/50 estimate for that window. An important part of the forecast is the author’s definition through a full set of cognitive abilities.
Interpretation boundary
Passing individual exams, being fast at coding, or excelling at a few tasks does not prove the full range of abilities.
Verification protocol
Before the start of the window, publish a set of independent tests in different cognitive areas and rules for accessing the system. At the end of the window, evaluate both the event and the probability calibration without turning 50% into a promise.
Within 1-5 years, AI could eliminate half of entry-level office positions and push unemployment in the US to 10-20%.
Announced
Deadline
What we measure
Entry office positions in the USA
Source
Direct interview with the author
Editorial analysis
Meaning, limitations and verification criterion
What it means
The forecast applies to entry-level office professional positions in the U.S.—including technology, finance, legal and consulting—not half of all jobs.
Interpretation boundary
A decline in job vacancies by itself does not indicate the cause: hiring is simultaneously influenced by the economic cycle, rates, previous overhiring and changing demands.
Verification protocol
Compare the share of entry-level vacancies and actual hires, employment of young professionals, U-3 and U-6. Separately check whether AI was introduced before the role was reduced.
By 2030, 170 million jobs could be created and 92 million displaced; the final balance will be plus 78 million.
Announced
Deadline
What we measure
Formal jobs around the world
Source
Official report
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is the scenario for all major market changes - technology, demographics, economics and the green transition. Software and applications developers are among the growing categories.
Interpretation boundary
A positive world balance does not mean that displaced people will be given new jobs, but that the professions, countries and skills required will be the same.
Verification protocol
Count created and displaced places separately, then compare the total with the part of formal employment covered by the report and the original methodology.
The number of software developers in the United States will grow by 10%—approximately 174,700 jobs—from 2025 to 2035.
Announced
Deadline
What we measure
Busy Software Developers in the USA
Source
State forecast
Editorial analysis
Meaning, limitations and verification criterion
What it means
The official US baseline expects a broader engineering role to grow, despite the automation of some tasks and the spread of AI.
Interpretation boundary
The forecast does not claim that AI will create all new places. The ten-year estimate may be subject to revision, and openings include replacing departing workers.
Verification protocol
Reconcile employment by occupation code 15-1252 in 2035 with the 2025 base and retain all interim forecast revisions.
The individual profession of computer programmer will shrink by 7%—by about 8,100 people employed—from 2025 to 2035.
Announced
Deadline
What we measure
Busy computer programmers in the USA
Source
State forecast
Editorial analysis
Meaning, limitations and verification criterion
What it means
BLS simultaneously anticipates a reduction in the role focused on writing and testing code and an increase in the number of software developers. This is a forecast of a change in the structure of professions, and not the disappearance of the entire development.
Interpretation boundary
Minus 7% does not mean that these people will be fired by AI: some of the tasks and workers can move into developer, analyst or platform roles.
Verification protocol
Review 15-1251 employment, occupational category transitions, and classification methodology changes by 2035.
By 2030, at least 20 million ICT specialists should work in the EU; in 2025 they were estimated at about 10.4 million.
Announced
Deadline
What we measure
Number of employed ICT specialists in the EU
Source
An established political goal
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is not a neutral forecast, but an official growth target for a broad group of ICT professionals that includes more than just software developers.
Interpretation boundary
The goal of 20 million cannot be passed off as the most likely result. The gap between the current estimate and the target shows the magnitude of the acceleration needed.
Verification protocol
Use the same Eurostat methodology, check the actual number of employees and gender balance, without replacing people with vacancies or graduates.
Exhibit No. 017Intentions · exact date not specified
Microsoft Work Trend Index
Global Jobs Report and AI 2025
Recorded wording
33% of executives are considering downsizing and 78% are considering hiring for new AI roles; both strategies can occur simultaneously.
Announced
Deadline
Not named
What we measure
Managers' hiring and layoff intentions
Source
Official global report
Editorial analysis
Meaning, limitations and verification criterion
What it means
Companies can remove some features and create others at the same time. The report also expects training and management of agents within teams.
Interpretation boundary
“Considering” does not mean that a decision has been made. The percentages cannot be subtracted from each other: one respondent could choose both strategies.
Verification protocol
Compare intentions with actual reductions and hiring by function, and also count the companies that implemented both scenarios.
Due to the efficiencies from AI, Amazon's overall corporate headcount is expected to decline over the next few years.
Announced
Deadline
"The Next Few Years"
What we measure
Total Amazon Corporate Staff
Source
CEO Initial Letter to Employees
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is the intention of one company: there will be fewer people in some functions and more in others, but the CEO called the final expected sign negative.
Interpretation boundary
The letter does not specify the amount, exact date or share of engineers. Reductions can occur through layoffs, natural attrition, and uncreated vacancies.
Verification protocol
Monitor corporate headcount in Amazon reporting, separating office roles from warehouses, contractors and acquired companies, and documenting new hires.
IBM plans to triple hiring for entry-level positions in the US in 2026, while reshaping the content of entry-level roles around AI.
Announced
Deadline
What we measure
Actual number of entry-level hires by IBM in the US
Source
Official publication of the company
Editorial analysis
Meaning, limitations and verification criterion
What it means
It's a counterexample to its complete rejection of newbies: IBM says hiring is up, but the roles span software development, consulting, infrastructure, marketing and other functions.
Interpretation boundary
“Tripling” does not show the original base and does not mean tripling the entire staff. At the same time, other old junior roles could disappear.
Verification protocol
Get actual 2025 and 2026 hiring numbers, role composition, employee retention, and developer share without replacing hiring numbers with job postings.
AI security researcher, professor at University of Louisville
Recorded wording
In about five years, unemployment may reach not 10%, but 99%: first computer work will be automated, then physical work.
Announced
Deadline
What we measure
Unemployment throughout the economy
Source
Full recording of the video interview
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is an extreme bet on the almost complete disappearance of paid work across the entire economy, not a separate forecast for the number of programmers.
Interpretation boundary
The technical ability to automate a task does not mean instant implementation, dismissal of a person, or the absence of new forms of employment.
Verification protocol
Check global employment, unemployment and employment-to-population ratio in 2030. Literal implementation requires the almost complete disappearance of the regular labor market.
About a quarter of global employment is in occupations with some exposure to GenAI, but job transformation is more likely than complete replacement.
Announced
Deadline
Not named
What we measure
Exposure of tasks rather than dismissals
Source
Official international index
Editorial analysis
Meaning, limitations and verification criterion
What it means
The index combines almost 30 thousand professional tasks, expert verification and employment data. In rich countries, exposure is higher and roles in software development are more prominent than in the 2023 version.
Interpretation boundary
25% cannot be translated as “every fourth person will be fired”: the exhibition shows the potential change in at least part of the tasks with full access to technology.
Verification protocol
Compare the index with actual implementation, changes in task composition, employment and quality of work for the same occupational classifications.
Over the next five years, net job losses due to AI are likely to be limited—less than 2–4% rather than tens of percent.
Announced
Deadline
What we measure
Net change in employment in the entire economy
Source
Full transcript of the discussion
Editorial analysis
Meaning, limitations and verification criterion
What it means
The assessment is based on the models' capabilities slowly becoming robust applications for large employers. The author calls Coding a possible exception with a faster effect.
Interpretation boundary
The range applies to the entire economy and the net total. In some occupations, the cuts may be greater, and new jobs may mask major reallocations.
Verification protocol
By July 2031, compare employment to the counterfactual trend, separate layoffs from uncreated positions, and check whether robust AI applications have been implemented.
Exhibit No. 025Live forecast · cut to August 30, 2026
Metaculus
Labor Automation Hub Community Aggregate Forecast
Recorded wording
At the August 30, 2026 cutoff, the community expected a 7.4% decline in software developer employment by 2030 relative to 2025.
Announced
Deadline
What we measure
Employment of software developers relative to 2025
Source
Live Aggregated Forecast
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is a collective probabilistic estimate that is updated with new data. It is directly contrasted with the official BLS rising baseline, but uses a different period.
Interpretation boundary
The current number is not a promise from Metaculus and may change. It cannot be compared to the BLS without aligning the base year, horizon, and occupation definition.
Verification protocol
Save the value and cut-off date, then compare actual employment in 2030 with the level in 2025 and separately show the history of changes in the collective forecast.
Australian Jobs 2026 - official employment forecast
Recorded wording
Employment of Software and Applications Programmers in Australia will grow by 15.7% over the five years to May 2030.
Announced
Deadline
What we measure
Busy Software and Applications Programmers
Source
State forecast
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is a national growing baseline, the opposite of the disappearing profession scenario. Base - about 195,400 employed in November 2025.
Interpretation boundary
The forecast does not prove that entry into the profession will become easier or that growth will be evenly distributed across regions, experiences, and specialties.
Verification protocol
Compare official employment in May 2030 with the base and all revisions, separately checking vacancies, migration and hiring patterns by experience.
About 40% of the world's jobs will be potentially replaced by AI in the 15-25 year horizon; this is not a forecast of 40% unemployment.
Announced
Deadline
What we measure
Share of global jobs potentially displaced by AI
Source
Full editorial transcript of the interview
Editorial analysis
Meaning, limitations and verification criterion
What it means
In the full conversation, Lee corrects the word "suppressed" to "suppressed" and gives a 15-25 year window. This is about the potential for automation of physical and office roles, not the guaranteed disappearance of all these positions.
Interpretation boundary
40% of tasks with high exposure to AI, using an assistant, or changing responsibilities does not equal 40% of jobs actually disappearing. The forecast does not take into account new roles and worker transitions.
Verification protocol
Determine in advance “displacement” and a unified international classification of professions. In 2034–2044, separately count automated tasks, disappeared roles and actual transitions of people.
DSIT and Warwick Institute for Employment Research
Experimental model of demand for AI skills in the UK
Recorded wording
The number of jobs directly involving AI could grow from about 158,000 in 2024 to 3.9 million by 2035—about 12% of the country's current workforce.
Announced
Deadline
What we measure
Jobs with AI activities in the adjusted technology scenario
Source
Officially published study
Editorial analysis
Meaning, limitations and verification criterion
What it means
The model overlays data on job openings and patents onto the Working Futures scenario. The 3.9 million includes both new jobs and existing roles that will integrate AI work.
Interpretation boundary
3.9 million is not the number of AI researchers, nor is it a net increase in employment, nor is it an estimate of displaced workers. Research exploratory; its findings are not declared government policy.
Verification protocol
In 2035, replicate the comparison of occupations and actual activities with AI, check the share of vacancies and employment by the original definition, and separately count new and transformed roles.
European model of employment, professions and replacement needs
Recorded wording
Between 2023 and 2035, employment in the EU27 will grow by more than 4%, and employment of ICT specialists and technicians by an average of 2.5% per year.
Announced
Deadline
What we measure
Average annual growth in employment of ICT specialists and technicians in the EU-27
Source
Official update of the European model
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is a forecast of actual employment by harmonized categories and not an EU policy target for the number of ICT specialists. The average hides a range from 1.3% in Germany and Italy to 7.8% in Romania.
Interpretation boundary
The forecast does not attribute all of the growth to artificial intelligence. The EU average cannot be transferred to an individual country, and replacement vacancies due to worker departures do not equal net employment growth.
Verification protocol
At the end of 2035, compare harmonized employment with the 2023 base, calculate the comparable average annual rate and separately show countries and replacement needs.
Stanford University's Centennial Study of Artificial Intelligence
Recorded wording
In the report's 2030 horizon, AI is more likely to replace individual tasks rather than entire jobs, while simultaneously creating new types of jobs.
Announced
Deadline
What we measure
The ratio of automated tasks, completely disappeared jobs and new types of work
Source
Official research panel report
Editorial analysis
Meaning, limitations and verification criterion
What it means
The report suggests looking at work as a set of tasks and limits the scenario to a typical North American city. The focus is on specialized AI, rather than the assumption of imminent general artificial intelligence.
Interpretation boundary
Automating one task does not automatically save a job, and the emergence of new job titles does not prove a positive employment balance. There are no percentages or a strict comparison base in the resume.
Verification protocol
Select a comparable sample of professions and cities, record the composition of tasks in 2016, and by the end of 2030 separately measure disappeared tasks, entire jobs and new types of paid work.
Labor Economy White Paper 2022, based on the METI model
Recorded wording
By 2030, Japan may lack about 160 thousand IT specialists with low market growth, 450 thousand with average growth and 790 thousand with high growth.
Announced
Deadline
What we measure
Scenario gap between model demand and supply of IT personnel in Japan
Source
Official State Report
Editorial analysis
Meaning, limitations and verification criterion
What it means
The range arises from three scenarios for the growth of the IT market. The average estimate - about 450 thousand - is not a point with a known error, but corresponds to a separate set of premises.
Interpretation boundary
A “shortage” does not equal the number of open positions and does not mean a reduction in employment. The model describes demand minus supply and is not a forecast of the impact of generative AI.
Verification protocol
In 2030, replicate the IT workforce, demand, supply and productivity definitions from the METI model. Compare the fact separately with the low, medium and high scenarios.
By 2030, the share of workers with skills in using AI technologies should grow to at least 80%, up from 5% in 2022.
Announced
Deadline
What we measure
Share of Russian workers with skills in using AI technologies
Source
Official regulatory document
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is a government goal for training and skills distribution, not a forecast of layoffs. Between the base 5% and the target 80% lies a sixteen-fold change in share.
Interpretation boundary
Taking a short course, accessing a chatbot, or self-assessing “I use AI” can produce very different results. Without a common definition, comparison with the 2022 base becomes meaningless.
Verification protocol
Retain skill definition, sampling method, and methodology for 2022; repeat a comparable measurement in 2030. Separately show basic, applied and professional skills.
By 2030, the total number of employees in the Russian IT industry should grow from almost one million to 1.4 million people; at least 250 thousand students must undergo training with the participation of business.
Announced
Deadline
What we measure
The number of employees in the Russian IT industry and the number of students trained with the participation of business
Source
Official publication of the Government
Editorial analysis
Meaning, limitations and verification criterion
What it means
The indicator describes the growth of sectoral employment simultaneously with the expansion of personnel training. It serves as a Russian counterpoint to predictions of a rapid decline in the profession due to AI.
Interpretation boundary
IT workers are not the same as developers: the metric may include sales, support, management, and other roles. The number of trained students is also not equal to the number of employed ones.
Verification protocol
In 2030, use the same definition of the IT industry and employment, which in the base value is “almost a million.” The number of employees and the training of 250 thousand students should be checked as two separate metrics.
AI could add up to $15.7 trillion to the global economy in 2030: about 6.6 trillion from productivity and $9.1 trillion from consumer effects.
Announced
Deadline
What we measure
Additional global economic output in 2030 attributable to AI in the model, separated by productivity and demand
Source
Economic forecast
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is an upper estimate of the model's contribution, not a promise of revenue for AI companies. Much of the amount depends on the spread of technology, changes in product quality and additional demand.
Interpretation boundary
The size of the software market, investment in data centers, or the growth of technology stocks cannot be directly compared to the contribution of AI to global output.
Verification protocol
Fix base year currencies and prices, actual world GDP, a counterfactual scenario without AI, and a method for separating AI from other technologies. Check performance and consumer effects separately.
Employers expect 39% of workers' core skills to change by 2030; out of a nominal 100 workers, 59 will need training, and for 11 it may remain unavailable.
Announced
Deadline
What we measure
Expected proportion of core skills changed and proportion of workers requiring training, retraining or transfer to another role
Source
International Employer Survey
Editorial analysis
Meaning, limitations and verification criterion
What it means
We are talking primarily about changing the set of skills within a job, and not about the disappearance of 39% of professions. The training indicator divides workers into skills enhancement, retraining with transfer, and unavailable training.
Interpretation boundary
A new tool in a company, the number of certifications issued, or a change in job title does not indicate what proportion of core skills actually changed or whether the training helped perform the job.
Verification protocol
Repeat a comparable survey of employers and supplement it with employee data and actual job requirements. Maintain definitions of core skill, training, and transfer to another role.
By 2030, the program aims to provide hands-on AI skills to ten million UK workers, including at least two million SMEs.
Announced
Deadline
What we measure
Number of unique workers who have mastered a verifiable set of basic practical AI skills, and a separate number of SME workers
Source
Official government goal
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is a program to prepare people for AI jobs, not a forecast of job losses. At the time of the goal expansion, one million courses were reported to have already been completed.
Interpretation boundary
Ten million registrations, open classes or short course completions do not necessarily equal ten million unique workers with a sustainable, hands-on skill set.
Verification protocol
Publish number of unique participants, deduplication rules, skill testing, and completion rate. The indicator of employees of small and medium-sized businesses should be counted separately and not mixed with the total number of courses completed.
41% of surveyed employers expect to reduce the number of employees by 2030 where AI can automate tasks; at the same time, 77% plan to upgrade employee skills, and 47% plan to move people out of affected roles.
Announced
Deadline
What we measure
Percentage of employers who have actually implemented downsizing, training, and internal transfers as a response to AI task automation
Source
International Employer Survey
Editorial analysis
Meaning, limitations and verification criterion
What it means
The indicator refers to companies' strategy, not the share of jobs lost. One employer can simultaneously cut some tasks, train people, and hire specialists for other roles.
Interpretation boundary
If 41% of companies cut at least one affected position, this does not mean that 41% of workers will be cut. Answers about plans also cannot be passed off as dismissals that have already occurred.
Verification protocol
Repeat a comparable survey and supplement it with actual data on headcount, reasons for layoffs, training and internal transfers. Company percentages do not translate into employee percentages.
A flexible system of computers, sensors and control mechanisms should lead to factories without workers; a critical situation with industrial employment, according to Wiener, could arise in ten to twenty years.
Announced
Deadline
What we measure
The emergence of factories without workers and large-scale industrial unemployment caused by computer control systems in the named twenty-year window
Source
Transcription of an archived letter
The outcome review has not yet been prepared. A passed deadline alone is not treated as proof or disproof.
Editorial analysis
Meaning, limitations and future assessment protocol
What it means
This is an early forecast of not only mechanization, but also programmable production control based on data from sensors. At the same time, Wiener made a strong social conclusion about the rate of displacement of labor.
Interpretation boundary
A single automatic line, an increase in productivity, or a reduction in people in one operation is not the same as a factory without workers and does not confirm large-scale unemployment in a specified period.
Verification protocol
Compare the level of automation, number of employees and output in industry by 1959 and 1969. The reasons for changes in employment should be separated from economic cycles, production relocations and changes in demand.
Survey of 31,000 workers and managers in 31 countries
Recorded wording
81% of executives surveyed expected AI agents to be moderately or heavily embedded into their company's AI strategy within 12 to 18 months.
Announced
Deadline
What we measure
Percentage of respondent organizations where agents are actually moderately or highly embedded in an established and implemented AI strategy
Source
Official global report
Editorial analysis
Meaning, limitations and verification criterion
What it means
The card records the expectations of managers over a short horizon. It complements a separate museum card about the long road to Frontier Firm, but measures a narrower embedding of agents in strategy.
Interpretation boundary
Intent, pilot, purchased license, or mention of agents in a presentation does not equal moderate or broad integration. 81% applies to the executives surveyed, not all companies in the world.
Verification protocol
Repeat the survey on a comparable sample and determine levels of integration through actual processes, budgets, responsibilities and regular use. Expectations for 2025 should be compared with reality, not with new expectations.
The global number of new installations of industrial robots should exceed 700,000 units per year by 2028; for 2025, 575,000 installations were expected.
Announced
Deadline
What we measure
Number of new industrial robots installed in the world per calendar year according to IFR statistical definition
Source
Official summary of the statistical report
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is a measurable forecast for the spread of industrial automation. It counts units supplied and installed, not the number of humanoids, workers, or fully autonomous factories.
Interpretation boundary
The fleet of already operating robots, orders, market revenue and software robots are not included in this indicator. An increase in installations does not automatically translate into an equal number of job cuts.
Verification protocol
Use the resulting IFR series with the same industrial robot definition, check annual installations 2025–2028 and data revisions. Analyze employment and productivity in separate series.
The total addressable market for humanoid robots could reach $38 billion by 2035; the base case separately assumes more than 250,000 deliveries in 2030, almost entirely to industry.
Announced
Deadline
What we measure
Global addressable market in dollars and annual shipments of devices meeting the study's definition of a humanoid robot
Source
Official Study Review
Editorial analysis
Meaning, limitations and verification criterion
What it means
The card stores the commercial forecast of the physical embodied AI. The main benchmark is addressable market size, with 2030 deliveries providing a separate intermediate point.
Interpretation boundary
A single manufacturer's revenue, investment, and installed fleet value are not equal to the global addressable market. Industrial manipulators without a humanoid form should not be included in the supply.
Verification protocol
In 2030, check annual deliveries for a unified definition of a humanoid robot. In 2035, separately estimate actual sales and addressable market, currency and research methodology.
By 2050, more than a billion humanoid robots could be used in the world, about 90% in industry and commerce; the market, including supply chains and services, could exceed $5 trillion.
Announced
Deadline
What we measure
The number of operating humanoid robots by purpose and the volume determined by market research, including supplies, repairs and maintenance
Source
Official Study Review
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is a long-term scenario for mass embodied AI, different from the Goldman Sachs forecast and the IFR timeline. It considers the fleet in operation and the expanded market, not just annual installations or deliveries.
Interpretation boundary
A billion units produced does not equal a billion active robots. The $5 trillion market includes connected chains and service, so it cannot be compared one-to-one with sales of the devices themselves.
Verification protocol
By 2050, determine the global operating fleet, exclude decommissioned and non-humanoid systems, check the distribution by sphere. Break down the monetary value into devices, components, repairs and maintenance.
By the end of 2027, generative AI will require 80% of the engineering workforce to be retrained.
Announced
Deadline
What we measure
Share of engineers who require upskilling
Source
Official analytical forecast
Editorial analysis
Meaning, limitations and verification criterion
What it means
The rate relates to changing skills and the emergence of new engineering roles. Gartner itself expects human expertise to continue and demand for skilled engineers to grow.
Interpretation boundary
The number 80% does not mean that 80% of engineers will lose their jobs. Without defining upskilling, the forecast can be formally fulfilled even in a short course.
Verification protocol
Define retraining early as a change in role requirements, paid training, or validated new skill, and measure the proportion of engineers who do so.
Study Measuring AI Ability to Complete Long Software Tasks
Recorded wording
If the observed trend continues and transfers to real work, by around 2030 AI will be able to automate many development tasks that last about a month.
Announced
Deadline
What we measure
50% task length horizon
Source
Research paper NeurIPS 2025
Editorial analysis
Meaning, limitations and verification criterion
What it means
The authors measure the duration of tasks that the model has a 50% probability of completing, rather than the number of lines written. The original work explicitly addressed extrapolation and limited transferability.
Interpretation boundary
A month's task with half of successful attempts is not equal to an autonomous replacement of an engineer: failures can be expensive, and the benchmark already contains part of the setting and verification.
Verification protocol
Update the trend slope annually, separately calculate 50%, 80% and 90% reliability and check migration to live repositories with ambiguous requirements.
FutureScape: Worldwide Developer and DevOps Predictions
Recorded wording
By 2030, 70% of developers will work alongside autonomous AI agents, shifting to planning, design and orchestration.
Announced
Deadline
What we measure
Share of developers working with agents
Source
Official analytical forecast
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is a forecast of collaboration, not the disappearance of the developer: the person is left with setting tasks, checking the result, architecture, security and agent management.
Interpretation boundary
Using autocomplete or a chatbot does not automatically count as working with an autonomous agent, and 70% users does not mean 70% automated tasks.
Verification protocol
Define an autonomous agent through its ability to act in the repository and pipeline, then measure developer coverage, human interventions, failures, and level of authority.
In the scenario, the year 2027 becomes the moment of the emergence of a superhuman programmer capable of accelerating AI research.
Announced
Deadline
What we measure
The capabilities of a superhuman coding agent
Source
Author's script and model
Editorial analysis
Meaning, limitations and verification criterion
What it means
The authors describe a specific possible course of events. After the updates, they retain 2027 as a serious possibility, but their median estimates are later.
Interpretation boundary
This is not an industry consensus or a promise that the event will occur exactly in 2027. The script cannot be read as the authors' only forecast.
Verification protocol
Check whether the system can perform any coding tasks of the best AI laboratory engineers faster and cheaper, and also autonomously contribute to AI-R&D.
By 2027, models may be able to do the work of an AI researcher/engineer and thereby speed up their own development.
Announced
Deadline
What we measure
Full cycle of work AI researcher/engineer
Source
Author's research essay
Editorial analysis
Meaning, limitations and verification criterion
What it means
The rate relates to a narrow frontier role: setting up experiments, engineering implementation, interpreting results and improving subsequent models.
Interpretation boundary
A high result on the coding benchmark does not equal replacement of an AI researcher, and automation of this role does not mean automation of the entire development market.
Verification protocol
Look for autonomous research results, impact on the training cycle and the number of meaningful human interventions, not just the speed of code generation.
By 2028, 90% of enterprise software engineers will use AI coding assistants, up from less than 14% at the beginning of 2024.
Announced
Deadline
What we measure
Share of enterprise engineers using assistants
Source
Official analytical forecast
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is a prediction of widespread use of the tool and a shift in work from implementation to orchestration, design and quality control.
Interpretation boundary
An issued license, enabled feature, or one-time use does not mean permanent work with an assistant and does not directly say anything about the size of the team.
Verification protocol
Consider active use in repositories and IDEs, session frequency and impact on workflow, separating access to the product from meaningful adoption.
By the end of 2027, more than 40% of agentic AI projects will be canceled due to cost, unclear value, or poor risk control.
Announced
Deadline
What we measure
Share of canceled agentic AI projects
Source
Official analytical forecast
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is a counter-forecast to the idea of the inevitable rapid implementation of agents: some projects will not reach production or will be stopped after the pilot.
Interpretation boundary
Canceling a project does not mean the company abandons AI: it can be relaunched, renamed, or replaced with a simpler assistant or automation.
Verification protocol
Define project and cancellation in advance, separately consider closure, freezing, downscaling and relaunch under a different name.
CEO of Microsoft AI at the time of the announcement
Recorded wording
Within 12 to 18 months, AI will reach human levels in most professional computing tasks, and many of them will be fully automated.
Announced
Deadline
What we measure
Professional tasks at the computer
Source
Live video interview
Editorial analysis
Meaning, limitations and verification criterion
What it means
The author spoke about the tasks of lawyers, accountants, managers and marketers, and later separately emphasized: individual tasks are not equal to entire professions.
Interpretation boundary
A good model response or a single successful job does not prove reliable autonomous operation with all data, exceptions, and responsibilities.
Verification protocol
By August 2027, select representative workflows and measure the proportion of tasks completed entirely without humans, cost, error rate, and actual adoption.
Former Harvard professor, entrepreneur and ACM TechTalk author
Recorded wording
Large language models can make the source code a secondary artifact: the natural language becomes the main interface, and the model becomes a new virtual machine.
Announced
Deadline
Not named
What we measure
The role of source code and programming languages
Source
Official technical report
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is a prediction about a change in the underlying level of abstraction, not about the number of jobs: specifications and queries may become more important than manually writing programs.
Interpretation boundary
The popularity of code generation does not prove that repositories, languages, tests, code review and reproducible builds are no longer the contract of industrial systems.
Verification protocol
Track what is the durable source of truth in production: the code and tests, or the specification from which the system is reliably regenerated.
By June 2026, the heaviest internal users were regularly running more than 60 hours of agent activity per day—in parallel, across multiple agents.
Announced
Deadline
Not named
What we measure
Estimated time for parallel operation of agents
Source
Official economic study
Editorial analysis
Meaning, limitations and verification criterion
What it means
In the extreme case, the engineer moves from the sequential execution of one task to the formulation, distribution and verification of several streams of agent work.
Interpretation boundary
60 agent hours do not equal 60 saved human hours: this is runtime, not a measurement of the accepted result, quality or staff savings.
Verification protocol
Calculate the percentage of accepted results, fixes, incidents, and actual task time reduction, separate from the agent startup duration.
If agents make an engineer 2 to 10 times more productive, previously unprofitable software projects will become viable and the demand for engineers will spread across all functions of the business.
Announced
Deadline
Not named
What we measure
New projects and demand for engineering labor
Source
Public interview
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is the development version of Jevons' paradox: lowering the cost of creating software can open up more demand than productivity can remove.
Interpretation boundary
Range 2–10x scenario. Releasing more software does not guarantee staff growth if demand increases slower than productivity.
Verification protocol
Track the number of funded projects, engineering budget, staff and the emergence of embedded builders outside IT after the development cost has been reduced.
The distinct role of a software engineer may eventually become dissolved among generalists on products, problems, and systems; The creation of programs will not disappear.
Announced
Deadline
Not named
What we measure
Professional role and boundaries of the profession
Source
Public speaking with a direct quote
Editorial analysis
Meaning, limitations and verification criterion
What it means
The forecast is not about a world without software development, but about the erosion of a separate profession: the intention is formulated by different specialists, and the code becomes an intermediate product.
Interpretation boundary
The disappearance of a job title does not equal the disappearance of employment. Architecture, Validation, and Operations may continue under other roles.
Verification protocol
Keep track of professional classifiers, vacancies and actual responsibilities: who designs, tests and is responsible for production systems.
The autonomy of the agent should depend on the risk: 60–70% of low-risk pull requests can be automatically checked and merged, but even two lines in the authentication module can be left to a person.
Announced
Deadline
Not named
What we measure
Proportion of low risk pull request and change risk level
Source
Official conference report
Editorial analysis
Meaning, limitations and verification criterion
What it means
The engineer's value shifts from testing every change equally to risk classification, setting the boundaries of autonomy and responsibility for sensitive areas.
Interpretation boundary
60-70% refers to low-risk pull requests in individual companies, and not to the entire development cycle, working hours or possible staff reductions.
Verification protocol
Weigh pull requests by complexity and risk, then compare review times, incidents, rollbacks, and the percentage of changes that actually did not require human intervention.
In agent-based development, human control remains necessary, and newcomers need a preceptor model: real work and safe mistakes instead of disappearing entry into the profession.
Announced
Deadline
Not named
What we measure
Developing Newbies and Maintaining Engineering Control
Source
Microsoft Build Official Post
Editorial analysis
Meaning, limitations and verification criterion
What it means
If simple tasks are taken over by AI, training has to be designed intentionally: the senior not only writes complex code, but also creates a safe environment for the next generation to grow.
Interpretation boundary
The analogy with medical residency does not yet prove that such a model will pay off or be accessible to small and poor companies.
Verification protocol
Compare formal mentoring programs and conventional training on time to independent duty, quality, retention, and progression from entry-level to senior level.
By the end of 2029, the computer will pass the pre-agreed strict Turing test: Kurzweil set it to “yes”, Kapor - to “no”.
Announced
Deadline
What we measure
Two-hour test with three judges and three human participants
Source
Primary description of a public bet
Editorial analysis
Meaning, limitations and verification criterion
What it means
The rules are noticeably stricter than a short conversation with a chatbot: each of the three judges conducts two-hour interviews with a computer and three people; the machine must pass both a recognition test and a comparative ranking.
Interpretation boundary
A spectacular demonstration, short dialogue, taking a different version of the test or pre-selected topics does not mean winning this bet. Even official success will show compelling textual behavior in this protocol, not consciousness or physical versatility.
Verification protocol
By the end of 2029, conduct a test according to published rules: the computer must fool at least two judges and obtain a median rank of at least two out of three people.
Cognitive Science Researcher; emeritus professor at NYU
Recorded wording
In 2029, the same AI system will fail at least three of the five proposed tests of general intelligence.
Announced
Deadline
What we measure
At least three out of five tests with one system
Source
Primary author's text with criteria
Editorial analysis
Meaning, limitations and verification criterion
What it means
The five tests cover understanding a movie, reading a novel, being a cook in an unfamiliar kitchen, generating more than 10,000 lines of error-free code, and translating mathematical proofs into testable symbolic form.
Interpretation boundary
Three different specialized products do not equal one overall system. Demonstrations without pre-set tasks, independent judges and rules for successful completion are also not considered.
Verification protocol
Before testing, formalize each test protocol and reliability criteria, then in 2029, have one unchanged system pass all five tests. The bet was proposed, but Elon Musk did not accept it.
By March 2028, a significant portion of OpenAI's research could be performed by AI systems in collaboration with the company's researchers.
Announced
Deadline
What we measure
Proportion of OpenAI research work carried out by AI in collaboration with humans
Source
Official publication of the company
Editorial analysis
Meaning, limitations and verification criterion
What it means
We are talking specifically about the research process inside OpenAI: testing ideas, finding errors, searching through alternatives and iterations next to a person. This is no longer just help when writing ordinary application code.
Interpretation boundary
The number of queries to the model, the amount of text generated, or the presence of a research agent do not indicate how much of the research was actually performed by the system.
Verification protocol
OpenAI must disclose the unit of account, the baseline, the share of projects or effort, and the role of the person. The word “significant” needs to be translated into a numerical threshold before a verdict can be reached.
In 2026, systems capable of obtaining new scientific insights are likely to appear, and in 2027, robots that perform tasks in the physical world may appear.
Announced
Deadline
What we measure
Independently verified new scientific findings and robots performing useful tasks outside of laboratory demonstration
Source
Author's publication
Editorial analysis
Meaning, limitations and verification criterion
What it means
These are two careful calendar hypotheses in one short route: first the contribution to the discovery of new knowledge, then the transfer of agency from the digital environment to the physical.
Interpretation boundary
A correct guess from a model, a retelling of a famous article, or a staged video of a robot does not equal a new conclusion and stable performance in a real environment.
Verification protocol
Point 2026 requires a conclusion previously unknown to specialists and independently reproduced. For point 2027 - a predetermined physical task, long-term battery life and error statistics.
Within 2–5 years, every organization will begin the path to a “Frontier Firm”—a company where work is built around hybrid teams of humans and AI agents.
Announced
Deadline
What we measure
Percentage of organizations that have undertaken verifiable process redesign around teams of people and agents
Source
Official analytical report
Editorial analysis
Meaning, limitations and verification criterion
What it means
The wording does not speak of the completed restructuring of each company, but of the beginning of the journey. The report divides the journey into stages: assistant, digital colleague and human-led agent process.
Interpretation boundary
Purchasing a license, a single pilot, or using a chatbot by employees does not yet prove that the organization is restructuring processes and creating hybrid teams.
Verification protocol
Prior to the audit, identify a minimum set of organizational changes and a sample of companies, including small businesses and the public sector. Count the share of those who started the journey separately from the share of those who completed the transformation.
Within five years, people will no longer have to choose a separate application for each task: the user will explain the goal in ordinary language to a personal AI agent, and agents will become the next computing platform.
Announced
Deadline
What we measure
Share of common digital tasks that a personal agent performs on a natural request between multiple services without manually selecting and switching applications
Source
Author's publication
Editorial analysis
Meaning, limitations and verification criterion
What it means
The forecast refers to a change in interface and software platform, and not just the emergence of chatbots. The agent must remember the context, act between applications and help bring the task to a result.
Interpretation boundary
Chat within one application, voice search or generation of instructions does not mean that applications have ceased to be the main way of working. Actions between services and a stable user context are required.
Verification protocol
Make a list of daily tasks and measure what percentage of people complete one call to a personal agent without selecting an application. Separately check errors, permissions, cancellation of actions and real mass availability.
By 2028, at least 15% of everyday work decisions will be made autonomously using agent-based AI, up from 0% in the base year of 2024.
Announced
Deadline
What we measure
Proportion of everyday work decisions that agent-based AI makes autonomously
Source
Official analytical forecast
Editorial analysis
Meaning, limitations and verification criterion
What it means
This forecast is not about the share of employees or the number of companies. Its unit is the individual working decision, and the word “autonomous” suggests action without prior human choice in each case.
Interpretation boundary
A model's advice, an automatically completed document, or a decision that a person approves every time cannot be considered without reservation as an autonomously made decision.
Verification protocol
Gartner must disclose the definition of solution, autonomy, operating environment, and calculation method. Using a comparable sample, measure the number of autonomous solutions and separate them from recommendations, drafts and automation according to strict rules.
86% of employers surveyed expect AI and information technologies to transform their businesses by 2030.
Announced
Deadline
What we measure
Proportion of employers where AI and information processing have led to significant changes in processes, products, task structure or employee requirements
Source
International Employer Survey
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is a company's expectation about the impact of technology, not a prediction that 86% of jobs will disappear or be fully automated.
Interpretation boundary
Buying a subscription to an AI service, experimenting with one department, or using a chatbot for some employees does not mean business transformation.
Verification protocol
Before 2030, set a threshold for significant transformation and re-examine a comparable sample of companies. Separately measure change in processes, products, tasks and occupancy.
By 2030, at least 75% of EU businesses must use one or more of three technologies: cloud computing, big data or artificial intelligence.
Announced
Deadline
What we measure
Share of EU businesses using cloud services, big data analytics or AI, as measured by official indicator definitions
Source
Official EU goal
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is a measure of technology adoption by companies, not the share of automated work. It shows the extent to which AI and its associated digital infrastructure have become embedded in business as usual.
Interpretation boundary
Paid cloud email, a one-time AI pilot, or storage of large amounts of data may formally bring the indicator closer, but do not necessarily mean deep process transformation or productivity growth.
Verification protocol
Use official Eurostat surveys, retain the definitions of the three technologies and count enterprises without double counting. The depth of implementation and economic effect should be analyzed separately.
Market forecast for enterprise AI agents for development
Recorded wording
By 2027, more than 65% of engineering teams using agent-based programming will consider the traditional integrated development environment optional and will move management, control and verification to automated platforms.
Announced
Deadline
What we measure
Share of teams among agent programming users for whom the IDE is no longer a mandatory center for their main work
Source
Official analytical forecast
Editorial analysis
Meaning, limitations and verification criterion
What it means
The forecast is not about 65% of all developers. The denominator is already limited to teams that use agent programming; inside them the main workplace and the method of code control change.
Interpretation boundary
Having a terminal agent, background code generation, or a web interface does not mean that the team is truly no longer dependent on the IDE to view, debug, test, and release changes.
Verification protocol
First, identify the teams that regularly use agent programming, then measure the proportion of those where the IDE is optional in the main process. Check where reviews, tests, access control and rollback are moved.
FutureScape: Worldwide Developer and DevOps Predictions 2026
Recorded wording
By 2028, with the number of agent deployments growing tenfold, 60% of G2000 companies will use a separate agent development lifecycle to scale and control agents.
Announced
Deadline
What we measure
Growth in the number of deployed agents and the share of G2000 companies with a formalized Agent Development Life Cycle
Source
Official analytical forecast
Editorial analysis
Meaning, limitations and verification criterion
What it means
The forecast describes the emergence of a separate engineering discipline: the agent needs to be created, tested, observed, constrained and constantly updated, and not just plugged into the model once.
Interpretation boundary
An internal document called ADLC, a few chatbots, or 10x growth from a very small base do not show that the company manages agents as full-fledged changing software systems.
Verification protocol
Fix the base number of agents, deployment unit, and composition of the G2000. ADLC requires repeatable phases of agent development, evaluation, acceptance, monitoring, versioning, and decommissioning.
Survey of more than 700 CIOs on the composition of IT work
Recorded wording
By 2030, AI will impact all IT work: CIOs expect 75% of work will be done by AI-enhanced people, 25% by AI alone, and the proportion of work done by people without AI will drop to zero.
Announced
Deadline
What we measure
Distribution of the entire volume of IT work between three categories: human without AI, human with AI and AI without human
Source
Official analytical forecast for the survey
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is a forecast about the composition of the work performed, and not about the number of positions. The AI DIY category can exist within a process that is still managed by people and the organization.
Interpretation boundary
A license for an AI tool, rare use of a chatbot, or the presence of an automatic step does not mean that AI has actually touched every IT task and completed a quarter of the work.
Verification protocol
Define the IT work unit and the rules of the three categories, then measure the actual effort and results achieved. Do not translate job share directly into employee or vacancy share.
By 2029, 60% of organizations will widely use smaller development teams, up from 15% in 2026. Such teams typically consist of four to five people, with two to three-person teams expected to become more common as the AI's skills and capabilities increase.
Announced
Deadline
What we measure
Proportion of organizations using scaled-down engineering teams and the actual size distribution of such teams
Source
Official analytical forecast
Editorial analysis
Meaning, limitations and verification criterion
What it means
Gartner describes a change in the development organization unit, rather than a general reduction in engineers: there may be more small teams, and platform units should provide them with common tools and control.
Interpretation boundary
One pilot out of three strong engineers, staff reduction or renaming of the scrum team does not prove the large-scale implementation of the model. The forecast does not directly promise a decrease in overall demand for developers.
Verification protocol
Define early on what counts as a reduced team and “at scale” implementation. Count organizations and teams separately, record platform engineering support, role composition, release and product quality.
By 2030, 65% of enterprises will integrate AI agents into DevOps and DevSecOps pipelines so that agents execute development and security workflows as an ongoing part of software delivery.
Announced
Deadline
What we measure
Share of enterprises where AI agents actually perform development or security tasks within production DevOps/DevSecOps pipelines
Source
Official analytical forecast
Editorial analysis
Meaning, limitations and verification criterion
What it means
We are talking about performing actions inside the pipeline, and not just about prompting the programmer in the editor. Therefore, the forecast affects access rights, logging, security checks, and release responsibilities.
Interpretation boundary
A chatbot, standalone test generation, or external assistant without pipeline access is not considered an embedded agent. The presence of integration also does not prove the regular execution of the production process.
Verification protocol
Review production pipelines, types of self-executed activities, frequency of runs, authority, logging, and human supervision. Determine in advance the denominator of enterprises and the minimum threshold of use.
Global electricity consumption by data centers will increase from approximately 415 TWh in 2024 to 945 TWh in 2030; AI will be the main but not the only source of this growth, and the demand for AI-focused data centers will more than quadruple.
Announced
Deadline
What we measure
Global annual electricity consumption of all data centers and individual AI-optimized facilities, using comparable IEA methodology
Source
Official model forecast
Editorial analysis
Meaning, limitations and verification criterion
What it means
The forecast shows the physical infrastructure of digital growth. He does not claim that all the energy in data centers is spent on AI, and specifically emphasizes the uncertainty of adoption, efficiency and network connectivity.
Interpretation boundary
The growth in capacity of constructed data centers is not equal to actual consumption. The indicator of one operator or country cannot be passed off as a global result, and the entire workload of data centers cannot be presented as an AI workload.
Verification protocol
Compare actual global consumption in TWh for 2030 with the baseline scenario, maintaining the accounting boundaries. Separately evaluate the share of AI, equipment load, efficiency and delayed connections.
By 2033, the global AI market could grow from $189 billion in 2023 to approximately $4.8 trillion, becoming the largest emerging technology and capturing about 30% of the total market for such technologies.
Announced
Deadline
What we measure
Revenue from sales of AI products and services in 2033 and its share in the comparable advanced technology market as defined by the report
Source
UN official report
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is a forecast of market revenue, not an economy's value added or an estimate of productivity. The report also draws attention to the concentration of research, infrastructure and benefits in a small number of countries and companies.
Interpretation boundary
AI companies' market capitalization, investment volume, or stated economic benefits are not a substitute for revenue from products and services. Nominal growth without a common currency and methodology is also incomparable.
Verification protocol
In 2033, reproduce the list of segments and definition of market revenue from the report, normalize the estimates to the same currency, and check the share of AI in the same set of advanced technologies.
Official recommendations of the company to the US government
Recorded wording
Powerful AI systems could emerge in late 2026 or early 2027: matching or surpassing Nobel laureates in most disciplines, operating through digital interfaces, and autonomously solving complex problems over hours, days, or weeks.
Announced
Deadline
What we measure
Simultaneously execute the full set of powerful AI properties from the document: wide expert level, work through interfaces and long autonomy
Source
Official publication of the company
Editorial analysis
Meaning, limitations and verification criterion
What it means
This isn't a vague bet on a "smart model": Anthropic listed several strong features of the system and a short calendar window. For testing, their totality is important, including long-term battery life.
Interpretation boundary
One expert-level result, a successful demonstration of computer control, or a long running agent does not validate the entire package. The company's assessment of its own technological direction cannot be considered an independent consensus.
Verification protocol
By the end of the window, prepare independent tests across multiple disciplines, digital interfaces and multi-hour tasks, measuring human intervention and reliability. The verdict is made based on a set of criteria.
Forecast of agent-based AI in enterprise applications
Recorded wording
By the end of 2026, up to 40% of enterprise applications will contain built-in task-specific agents, up from less than 5% in 2025.
Announced
Deadline
What we measure
Share of enterprise applications with a built-in agent that can independently perform a complex end-to-end task, as determined by Gartner
Source
Official analytical forecast
Editorial analysis
Meaning, limitations and verification criterion
What it means
Gartner draws a line between an assistant who waits for every request and an agent who performs an end-to-end task. Therefore, the card does not measure the presence of any AI feature, but rather the adoption of a specific class of applications.
Interpretation boundary
Marketing “agent”, in-product chat, or automated prompts should not be included in the numerator. The metric relates to applications, not the share of companies, users or work completed.
Verification protocol
Capture a sample of enterprise applications and test products against the Gartner functional test: complete a complex task yourself. Consider applications, not licenses or vendor statements.
Exhibit No. 130Horizon without an exact upper limit
Sam Altman
CEO OpenAI
Recorded wording
Superintelligence may appear several thousand days after publication; it may take longer, but the author is confident that it will be created.
Announced
Deadline
"Several thousand days", with a caveat of possible delay
What we measure
A system that is reasonably recognized as superintelligence for a predetermined wide range of scientific, engineering and other intellectual tasks
Source
Initial author's publication
Editorial analysis
Meaning, limitations and verification criterion
What it means
The phrase sets an approximate horizon, but not a calendar deadline: “several” is not translated by the author into a number, and the clause allows for a longer path. The card therefore does not assign an arbitrary year of 2030 or 2033.
Interpretation boundary
A new model, AGI by a company's internal definition, or superiority in one benchmark is not necessarily super-intelligence. The growth of prosperity and scientific discoveries in the essay are separate consequences, and not a criterion for date.
Verification protocol
Prior to assessment, agree on an operational definition of superintelligence, an independent set of closed tasks, and requirements for autonomy, resilience, and skill transfer. The date is fixed only after independent confirmation.
Exhibit No. 131Conditional window after the appearance of powerful AI
Dario Amodei
CEO and co-founder of Anthropic
Recorded wording
Powerful AI could speed up discoveries by at least tenfold and compress the next 50-100 years of progress in biology and medicine into 5-10 years after its introduction.
Announced
Deadline
What we measure
Pace of major biological discoveries, clinically validated methods, and healthcare outcomes relative to trajectory without powerful AI
Source
Primary author's essay with formulated assumptions
Editorial analysis
Meaning, limitations and verification criterion
What it means
The countdown begins not with the publication of the essay, but with the appearance of a powerful AI determined by the author. The 2031–2036 date is shown only as a consequence of his early 2026 scenario, not as a promise in its own right.
Interpretation boundary
More publications, patents, or drug candidates do not equal 50–100 years of progress. It is necessary to separate the contributions of AI from investments, new laboratories, regulatory changes and normal scientific growth.
Verification protocol
First, independently record the moment of passing the threshold of powerful AI. Then compare the rate of major reproducible discoveries, development time, and clinical outcomes with the preconstructed baseline trajectory.
By 2029, agent-based AI will independently, without human intervention, resolve 80% of typical customer support requests, which will lead to a reduction in operating costs by 30%.
Announced
Deadline
What we measure
Proportion of typical problems completely resolved by an agent without human intervention, and comparable change in support service operating costs
Source
Official analytical forecast
Editorial analysis
Meaning, limitations and verification criterion
What it means
The forecast refers to completed solutions to common problems, and not to the proportion of chats where the bot responded first. The second part being tested is a 30% reduction in help desk operating costs.
Interpretation boundary
Automatic greeting, knowledge base retelling, routing, or replying to an operator are not considered standalone solutions. Reducing costs for other reasons does not confirm the influence of agents.
Verification protocol
Using a stable sample of requests, determine in advance a “typical problem” and a successful solution, take into account repeated requests and the hidden work of people. Compare costs for the same quality, volume and complexity of service.
John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon
Recorded wording
The authors suggested that any aspect of learning and intelligence could in principle be described by a machine, and ten researchers could make significant progress in one summer.
Announced
Deadline
What we measure
Significant progress in the listed AI tasks
Source
Archive of the original proposal
The outcome review has not yet been prepared. A passed deadline alone is not treated as proof or disproof.
Editorial analysis
Meaning, limitations and future assessment protocol
What it means
The document set the initial agenda for AI: language, abstractions, solving human problems, and self-improvement. The same words continue to be heard in today's job forecasts.
Interpretation boundary
“Significant progress” had no quantitative criterion, and a later archive note states that there was no final overall project report.
Verification protocol
Compare each of the stated sub-points with the published results of the participants, without declaring the entire field passed or failed with one verdict.
The authors estimated the reduction in coding time against assembly to be about 4 to 20 times and showed that 47 lines of FORTRAN replaced about 1,000 machine instructions.
Announced
Deadline
Not named
What we measure
Encoding time and volume of machine instructions
Source
Original conference paper
Editorial analysis
Meaning, limitations and verification criterion
What it means
One manual layer was indeed disappearing: the translator took over indexing, bookkeeping, and issuing machine instructions, and programming was elevated to a higher level.
Interpretation boundary
Speeding up coding did not predict the disappearance of programmers. A new language, a compiler, debugging of the source program, and specialists creating the translator itself were required.
Verification protocol
Comparing identical tasks, learning time, coding and debugging, as well as the quality and speed of the result is exactly what the authors of the article tried to do.
In 1960, Simon wrote that within twenty years machines would technically be able to do any human job—though economically, humans would retain a comparative advantage.
Announced
Deadline
What we measure
Technical ability to perform any human work
Source
Digitized original book
The outcome review has not yet been prepared. A passed deadline alone is not treated as proof or disproof.
Editorial analysis
Meaning, limitations and future assessment protocol
What it means
This is an early separation of capability and employment: even with broad technical substitutability, the market outcome depends on the cost and comparative advantage of people.
Interpretation boundary
Literal universal ability had not been shown by 1980. But the deadline miss does not refute Simon's economic framework of a moving task boundary.
Verification protocol
For each profession, separately check the completeness of tasks, price, reliability and the ability to work without hidden human help, and not one spectacular benchmark.
Posted by Application Development Without Programmers
Recorded wording
If productivity doesn't improve, Martin wrote, within ten years the industry will need 93.1 times more programmers—about 28 million—with a 41-fold increase in the number of applications.
Announced
Deadline
What we measure
Number of programmers and applications with zero productivity growth
Source
Official NIST report with book citation
The outcome review has not yet been prepared. A passed deadline alone is not treated as proof or disproof.
Editorial analysis
Meaning, limitations and future assessment protocol
What it means
The book's title promised development without programmers, but the numerical argument was about explosive pent-up demand. 4GL and end-user development were proposed as an answer to the shortage.
Interpretation boundary
28 million is a conditional counterfactual, not an unconditional promise. Productivity growth specifically changed the initial assumption.
Verification protocol
Compare 1992 occupational statistics to a benchmark and separately estimate how much of application growth was absorbed by tools and users.
Forecast Analysis: Low-Code Development Technologies
Recorded wording
By 2025, 70% of new enterprise applications were expected to use low-code or no-code, up from less than 25% in 2020.
Announced
Deadline
What we measure
Share of new applications with low-code or no-code
Source
Saved Gartner Forecast Content
The outcome review has not yet been prepared. A passed deadline alone is not treated as proof or disproof.
Editorial analysis
Meaning, limitations and future assessment protocol
What it means
This is a prediction of the spread of the development method, not the disappearance of professionals: low-code could be used within an application along with regular code.
Interpretation boundary
Without a published definition of "use" and a comparable final sample, 70% cannot be fairly declared as completed. Even execution doesn't measure headcount.
Verification protocol
We need a final summary for 2025 with the same denominator, list of platforms and the rule for accounting for hybrid applications, which was used in the base 25%.
Author of the essay Why software developers might be obsolete by 2030
Recorded wording
Despite the alarming headline, the author did not predict the disappearance of developers by 2030, but a renewal of the role: less routine, more planning, creativity and leadership.
Announced
Deadline
What we measure
Preserving the profession and changing the composition of tasks
Source
Dated author's essay
Editorial analysis
Meaning, limitations and verification criterion
What it means
The card shows why the museum stores not only the headline: within the text, the question of replacement turns into the opposite bet on strengthening the role of the developer.
Interpretation boundary
“They will become leaders” is a direction, not a professional classification or a numerical forecast. Increased responsibility may come with fewer positions.
Verification protocol
By 2030, compare employment and the share of vacancies where the developer is responsible for discovery, architecture, product solutions and automation management.
Scientist and engineer; essay writer No Silver Bullet
Recorded wording
In the next ten years, there will not be one technology or management technique that, on its own, promises a tenfold improvement in the productivity, reliability and ease of software development.
Announced
Deadline
What we measure
One technology and tenfold improvement in software development
Source
Initial technical report
The outcome review has not yet been prepared. A passed deadline alone is not treated as proof or disproof.
Editorial analysis
Meaning, limitations and future assessment protocol
What it means
What Brooks was denying was not progress at all, but the emergence of one “silver bullet.” He separately admitted that the sequential application of many improvements can give a large cumulative gain.
Interpretation boundary
An increase in the productivity of an individual team, the acceleration of one stage, or the cumulative effect of several technologies do not refute the thesis that there is no single universal solution.
Verification protocol
For a retrospective verdict, determine the baseline in advance and check whether one technology in 1986–1996 produced an order of magnitude at the same time in terms of its stated characteristics. Without such a base, the card remains open and does not automatically “come true.”
Economist; author of the essay “Economic Opportunities for Our Grandchildren”
Recorded wording
In a hundred years, the standard of living in progressive countries will be 4–8 times higher; the remaining work can be distributed in three-hour shifts, or 15-hour weeks.
Announced
Deadline
What we measure
Increase in living standards and actual working hours
Source
Primary text of historical essay
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is a composite forecast. A four to eightfold increase in living standards is framed as a straightforward expectation, and a 15-hour week is framed as a possible way to distribute work once basic economic needs are met.
Interpretation boundary
An increase in GDP or income per person does not in itself prove a shorter week. Conversely, shorter hours in individual countries do not confirm the claimed increase in living standards across the entire selected range of countries.
Verification protocol
Before the assessment, fix the set of “progressive countries”, an indicator of the real standard of living and the 1930 base. In 2030, separately compare the growth of the indicator and the average actual working time; do not reduce the two parts to one verdict.
Mathematician, writer and essayist on the technological singularity
Recorded wording
Within thirty years the technological means to create superhuman intelligence will be available; soon after this, according to the author's forecast, the human era will end.
Announced
Deadline
What we measure
Availability of technological means for superhuman intelligence according to a predetermined test
Source
Primary text of the report in the NASA archive
The outcome review has not yet been prepared. A passed deadline alone is not treated as proof or disproof.
Editorial analysis
Meaning, limitations and future assessment protocol
What it means
This is not a forecast about a specific model or profession, but an early formulation of a threshold: the emergence of intelligence beyond human intelligence should irreversibly change subsequent history. The said thirty-year window has already ended.
Interpretation boundary
High AI results in individual tests, widespread adoption of chatbots, or rapid improvement of models do not in themselves define “superhuman intelligence,” much less “the end of the human era.”
Verification protocol
Before a verdict, a published set of general cognitive and agential tests, a rule of comparison with people, and independent evidence of an irreversible historical transition are needed. So far there is no such initial protocol in the statement.
Neural network researcher, director of AI at Tesla at the time of publication
Recorded wording
Neural network weights are becoming Software 2.0: instead of explicitly writing all the instructions, people specify the data, the goal, and the architecture, and optimization finds a program that works.
Announced
Deadline
Not named
What we measure
Share of tasks where explicit code is replaced by learning parameters and data-driven development
Source
Author's program essay
Editorial analysis
Meaning, limitations and verification criterion
What it means
The forecast concerns a change in the medium of part of the logic: behavior is increasingly determined by data and the evaluation function, and not just by hand-written branches.
Interpretation boundary
Karpathy did not promise the disappearance of programmers. In the same text, he leaves it to people to work with data, learning infrastructure, analytics, visualization and tools.
Verification protocol
For a selected class of systems, compare the proportion of behavior specified by explicit code, trainable parameters, and data. Do not consider the size of a model file to be equivalent to lines of source code.
Psychoacoustician and interactive computing pioneer
Recorded wording
In 10–15 years, a “thinking center” should appear, combining the functions of a library, storing and retrieving information, computing and a network of similar centers for remote users.
Announced
Deadline
What we measure
A functioning network center combining digital library, search, computing and interactive access
Source
Primary scientific article
The outcome review has not yet been prepared. A passed deadline alone is not treated as proof or disproof.
Editorial analysis
Meaning, limitations and future assessment protocol
What it means
Licklider did not describe independent AI, but close cooperation: a person sets goals and evaluates the result, a computer takes care of routine operations. The “thinking center” was an infrastructural condition for such a regime.
Interpretation boundary
One terminal, electronic catalog or remote session alone does not yet form the entire system described. In the source text there is no exact boundary after which the center is considered created.
Verification protocol
To evaluate, compare the existing systems by March 1975 according to four criteria at once: a network of centers, remote access, full-text collections and interactive computing. Do not retroactively transfer the modern Internet into the past.
A home information terminal with a keyboard and screen will access a shared computer through the telephone network and provide instant access to books, newspapers, catalogs, schedules and personal files.
Announced
Deadline
No calendar deadline
What we measure
Massive home screen access to remote computing, shared information collections and personal data
Source
Primary author's article
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is an early, seamless image of a home network service: not only reading documents, but also working with a remote computer, personal files, purchases and messages through one terminal.
Interpretation boundary
The coincidence of individual details with a smartphone or the web does not mean complete execution. The article describes many functions, but mass distribution, price and exact deadline are not determined.
Verification protocol
Divide the forecast into functions and for each, record the first technical implementation and the moment of mass distribution. A general verdict is possible only after choosing a mass threshold.
Mathematician, author of notes on the description of the Analytical Engine
Recorded wording
If the relationships of musical sounds can be expressed by formal rules, the Analytical Engine will be able to compose complex scientifically constructed musical works of any complexity and duration.
Announced
Deadline
No calendar deadline
What we measure
The ability of a universal computing system to work with formalized relationships not only of numbers and generate musical material
Source
Primary historical text
Editorial analysis
Meaning, limitations and verification criterion
What it means
Lovelace separated the computational operation from concrete numerical material. This is an early vision of the general purpose programmable machine and the machine creation of symbolic content.
Interpretation boundary
A modern melody generator by itself does not prove that a machine understands music. The initial formulation is conditional: first musical relations must be expressed in a form accessible to calculation.
Verification protocol
Check the literal part against a working system that receives a formal representation of musical relationships and creates a work from it. The issue of creativity and understanding should be taken into account separately.
The future personal Memex will allow a person to store books, records and messages, quickly find them, link materials into associative routes and transfer such routes to other people.
Announced
Deadline
No calendar deadline
What we measure
Personal knowledge system with quick search, stable connections between materials, notes and the ability to share collected routes
Source
Author's publication
Editorial analysis
Meaning, limitations and verification criterion
What it means
The mechanics on microfilm did not literally come true, but the actions described are reminiscent of personal computers, hyperlinks, bookmarks, knowledge bases, and collaborative exploration of information.
Interpretation boundary
Conventional full-text search is not the whole idea of Memex: personal connections, long reading paths, comments, and transfer of structure to another person are essential.
Verification protocol
Compare not the appearance of the device, but the functions: personal storage, quick access, associative links, annotations and transfer of routes. Separately note which functions have become widespread.
Computer tools should not just automate individual operations, but enhance human intelligence: speed up and deepen understanding of complex situations, help find faster and better solutions.
Announced
Deadline
No calendar deadline
What we measure
Changing the speed and quality of solving complex professional problems by a person using a connected system of methods, language, interfaces and computing tools
Source
Primary research report
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is the precursor to the current idea of human-AI collaboration. The person remains the center, and the result is assessed by solving a complex problem as a whole, and not by the spectacular response of the program.
Interpretation boundary
An increase in typing speed, a single hint, or a demonstration of a new interface does not prove increased intelligence. Changes are needed in understanding, quality of solution, and ability to work with a previously too complex problem.
Verification protocol
Determine in advance the complex task, the basic process without a system, the quality of the solution, the time and number of errors. Compare the effect across multiple levels of experience and test whether the effect persists across teamwork.
Xerox PARC researcher, author of the DynaBook concept
Recorded wording
A personal computer should belong to the user, be portable, cost no more than a television, and serve simultaneously as a medium of expression, a set of modifiable tools, a book, a notebook, and a communication channel.
Announced
Deadline
No calendar deadline
What we measure
Mass availability of a portable personal computing device for reading, writing, programming, learning, creativity and communication
Source
Primary scientific publication
Editorial analysis
Meaning, limitations and verification criterion
What it means
The concept anticipated laptops and tablets not only in form. More important is the idea of a personal programmable environment, where a child or adult does not simply consume ready-made content, but creates their own models and tools.
Interpretation boundary
The mere presence of a thin screen or a cheap device does not confirm the entire forecast. Modern technology can be identical in form, but remain a closed environment for consuming other people's applications.
Verification protocol
Check separately for portability, personal ownership, affordability, versatility of media, the ability to create your own programs and use by children without the constant help of specialists.
The ultimate development of a computer display should be an environment in which the machine controls not only the image, but also the perceived properties of objects: a virtual chair should feel like a chair, and the world created by the program should feel like a space that can be entered.
Announced
Deadline
No calendar deadline
What we measure
The ability of a computing environment to convincingly reproduce the visual, auditory, and physical properties of interactive objects
Source
Primary scientific publication
Editorial analysis
Meaning, limitations and verification criterion
What it means
The text anticipated virtual and augmented reality, spatial interfaces and physical feedback. Sutherland was not describing a screen with fancy graphics, but a programmable interaction environment.
Interpretation boundary
A virtual reality helmet or a three-dimensional game does not support the strong version of the idea: the image of an object does not yet have its physical properties and does not replace the real environment.
Verification protocol
Separately test image realism, spatial audio, motion tracking, haptic feedback, and the ability to safely interact with virtual objects as if they were physical ones.
A person will be able to request the necessary pages of books from a large library from home and read them on the screen using an “electric telescope”; first within the city, then over long distances.
Announced
Deadline
No calendar deadline
What we measure
Remote home access to selected pages of a distributed library collection via a screen and communication network
Source
Primary historical text
Editorial analysis
Meaning, limitations and verification criterion
What it means
The technical design was analog, but the user action was recognizable: formulate a request at home, retrieve a specific document from a remote storage and read it on the screen.
Interpretation boundary
The television broadcast of one pre-selected page is not yet equal to the modern Internet: the forecast does not have open publication, independent navigation, full text search and user editing.
Verification protocol
Compare functions, not devices: remote request, page selection, home screen display, distance scale and accessibility for the average person.
By 2001, a home computer console linked to other computers will make it possible to access banking information, book flights, and do work from home that previously required an office presence.
Announced
Deadline
What we measure
Prevalence of connected home computers and share of office tasks that can be completed remotely via the network
Source
Archival video interview
The outcome review has not yet been prepared. A passed deadline alone is not treated as proof or disproof.
Editorial analysis
Meaning, limitations and future assessment protocol
What it means
Clark linked the home computer to practical services and remote work, not just computing. However, by this time, access and the ability to work from home were distributed very unevenly.
Interpretation boundary
The presence of a home computer for a portion of the population does not mean that any work has become remote or that the employer has allowed the employee not to come to the office.
Verification protocol
To estimate by time period, use statistics on home computers and Internet access for 2001, the availability of these services, and the actual share of people who regularly worked from home.
Futurist, author of the “electronic cottage” concept
Recorded wording
By 2006 or earlier, a new production system should emerge in which the home, with its low-cost workstation, computer communications, and teleconferencing, will once again become an important place of work.
Announced
Deadline
What we measure
Proportion of employed people who regularly perform some or all of their paid work from home using computers and telecommunications
Source
Author's publication
The outcome review has not yet been prepared. A passed deadline alone is not treated as proof or disproof.
Editorial analysis
Meaning, limitations and future assessment protocol
What it means
Toffler did not predict single freelancing, but the transfer of a significant part of production and office work to home, along with savings on travel and real estate. Mass participation is more important than the availability of technology.
Interpretation boundary
A few remote jobs, a home computer, or temporary work from home do not mean that the home has become the center of a new production system for a large part of the population.
Verification protocol
Compare regular remote work before and after 2006, separately taking into account entirely home-based, hybrid and temporary work, as well as differences between countries and professions.
By 2014, machines will do almost any routine job better than humans, so people will mostly be left to maintain machines, with the most valuable exception being creative work.
Announced
Deadline
What we measure
The share of routine work tasks where a machine consistently outperforms a human, and the share of employees whose main role is to maintain machines
Source
Author's newspaper publication
The outcome review has not yet been prepared. A passed deadline alone is not treated as proof or disproof.
Editorial analysis
Meaning, limitations and future assessment protocol
What it means
The forecast accurately highlighted the vulnerability of repetitive tasks, but turned the automation of individual tasks into a much stronger statement about the majority of routine work and employment patterns by a specific date.
Interpretation boundary
A robot in a factory, an ATM, or the automation of a single operation does not show that almost all routine work is done by machines better than people and that most workers only serve the equipment.
Verification protocol
For the 2014 assessment, first measure routine at the task level, then measure technical feasibility, actual implementation, and employment structure. Do not bring modern AI capabilities back by the deadline.
Information systems researcher, author of the term “hypertext”
Recorded wording
A computer system for personal documents and creativity should allow a person to build complex connections of his own, freely change the structure, save alternatives and work with interrelated text and graphic material that is inconvenient to represent on paper.
Announced
Deadline
No calendar deadline
What we measure
Availability of a personal, flexible environment where documents are interconnected, support alternative structures, and serve as a tool for creativity
Source
Primary scientific publication
Editorial analysis
Meaning, limitations and verification criterion
What it means
Nelson described hypertext as a more general work environment than a sequence of web pages. The implemented World Wide Web embodied part of the idea, but not all the properties of the original system.
Interpretation boundary
A simple link between two pages does not support the entire concept: it is essential to change personal structures, alternatives, the origin of fragments and complex connections within the material.
Verification protocol
Compare separately links, backlinks, reuse of fragments, saving versions, alternative routes and the ability for the user to rebuild their own document system.
Humanity needs a constantly updated World Encyclopedia: a common, accessible and authoritative body of knowledge, linked to primary sources, for learning, management and collaboration by people around the world.
Announced
Deadline
No calendar deadline
What we measure
Existence of a global, regularly updated and widely accessible body of verifiable knowledge with links to primary materials
Source
Primary historical text
Editorial analysis
Meaning, limitations and verification criterion
What it means
Wells was not just describing an electronic library or a specific future program. His vision centered on renewal, verifiability, education, and the international organization of knowledge.
Interpretation boundary
One large website, search engine or online encyclopedia fulfills only part of the idea. The size of the audience in itself does not say anything about the authority, completeness, stability and independence of knowledge.
Verification protocol
Separate access, coverage, relevance, citations to primary sources, error correction policies, and editorial controls. Do not declare modern service the complete embodiment of Wells' entire social program.
Computers should disappear into the ordinary environment: many specialized devices of different sizes will be connected by a network and will become so familiar that a person will no longer perceive interaction with them as a separate work at the computer.
Announced
Deadline
No calendar deadline
What we measure
The proliferation of networked computing devices in everyday environments and the reduction in the proportion of interactions that require special attention to the computer itself
Source
Primary author's article
Editorial analysis
Meaning, limitations and verification criterion
What it means
Weiser predicted more than just an increase in the number of screens. Its key idea is calm computing, embedded in context, objects and space, rather than the user's constant struggle with the interface.
Interpretation boundary
A large number of smartphones and sensors does not yet prove that the technology has “disappeared”: notifications, customization, incompatibility and constant shifting of attention may mean just the opposite.
Verification protocol
Measure the number and types of embedded devices, compatibility, the share of automatic contextual actions and user attention costs. The prevalence of technology and the calmness of interaction should be assessed separately.
Author of the proposal from which the World Wide Web grew
Recorded wording
CERN should create a universal system of linked data with remote access, support for different computers and without a single center; in ten years, Berners-Lee suggested, there could be many commercial solutions for this problem.
Announced
Deadline
What we measure
Availability by 1999 of many working solutions for network linking of heterogeneous documents and data with access from different systems
Source
Initial technical proposal
The outcome review has not yet been prepared. A passed deadline alone is not treated as proof or disproof.
Editorial analysis
Meaning, limitations and future assessment protocol
What it means
The document simultaneously solved a local problem at CERN and assumed that such a problem would soon become common. Not only links are important, but also portability, decentralization and the ability to connect old databases.
Interpretation boundary
The emergence of one browser or one site does not fulfill the prediction of a multitude of compatible solutions. The mere fact of Internet growth also does not test the architectural properties listed in the proposal.
Verification protocol
Compile a snapshot of 1999: browsers, servers, open standards, number of sites and compatibility between platforms. Separately, check whether decentralization and the ability to connect disparate sources have been preserved.
Apple co-founder speaking at a design conference in Aspen
Recorded wording
The personal computer will become the new primary means of communication, entering the work, school and home environments, and portable computers with radio communications will allow you to receive messages and data on the go; the first phase of mass distribution will be completed in about fifteen years.
Announced
Deadline
What we measure
The massive presence of personal computers in work, education and at home, their role as a communication medium and access to data from portable devices via wireless communication
Source
Archival recording and transcription of the speech
The outcome review has not yet been prepared. A passed deadline alone is not treated as proof or disproof.
Editorial analysis
Meaning, limitations and future assessment protocol
What it means
The forecast combined the device, the network, and a new way of communicating. It is interesting because it was made before the mass Web and long before the smartphone, but some of the technology already existed in laboratories and email.
Interpretation boundary
High sales of computers do not prove that they have become the dominant means of communication. A wireless laptop for one class of users also does not mean mass access on the go.
Verification protocol
For 1998, separately check the penetration of computers at home, work and education, time of use, share of digital communication and the prevalence of portable wireless access.
Mathematician, one of the founders of the theory of computing
Recorded wording
Machines will eventually rival humans in all purely intellectual fields; According to Turing, it was necessary to start with an abstract game and teach the machine to understand and use language.
Announced
Deadline
No calendar deadline
What we measure
Sustained machine competitiveness across a wide range of intelligent domains, rather than just one test or one game
Source
Primary article in Mind magazine
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is Turing's broader bet than his famous numerical prediction about the imitation game. It concerns a variety of intellectual activities and involves learning, language and interaction, not just pre-written rules.
Interpretation boundary
Winning at chess, passing an exam, or scoring high on one benchmark does not prove competitiveness in all intellectual domains. The word "compete" also does not mean a complete replacement of people.
Verification protocol
Define in advance a diverse set of open intelligence areas, compare quality, autonomy, sustainability, and cost with humans, and separately capture areas where humans retain the advantage.
An electronic computer with a program for searching and evaluating positions will be able to play fairly strong chess at a speed comparable to the speed of a person, although a complete search of all games is practically impossible.
Announced
Deadline
No calendar deadline
What we measure
The strength of a chess program against humans with comparable time per move and using selective search instead of brute force
Source
Primary scientific article, archived copy
Editorial analysis
Meaning, limitations and verification criterion
What it means
Shannon described a practical way to machine play: not by solving entire chess, but by limiting the tree of options and evaluating positions. This is an important early prediction of how a computing system will perform in a complex intelligent task.
Interpretation boundary
The mere fact of running a chess program does not confirm that the game is “strong enough.” The modern superiority of a chess engine also does not prove general intelligence outside of strictly formalized play.
Verification protocol
Compare ratings, time control, hardware resources and results against people. Separately mark the date of achievement of the amateur, master and champion levels, and do not transfer chess success to other tasks.
Mathematician; the statement is preserved in the memoirs of Stanislav Ulam
Recorded wording
Accelerating technological progress and changing lifestyles create the impression of approaching a significant historical singularity, after which human affairs can no longer continue in the same way.
Announced
Deadline
No calendar deadline
What we measure
The presence of a qualitative historical break, after which previous institutions, work and everyday life cease to be reproduced in a recognizable form due to the acceleration of technology
Source
Recollection of a direct interlocutor
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is one of the earliest recorded uses of the word "singularity" in technology. It was about the general acceleration of technology and lifestyle, and not about a specific date for the appearance of AGI or a single computing system.
Interpretation boundary
The rapid growth of computing, a single crisis, or the popularity of AI does not in itself confirm an irreversible break in the entire story. The formulation came through Ulam and is not a verbatim publication by von Neumann himself.
Verification protocol
When evaluating, maintain the indirect source type and do not assign a missing date to it. Look for simultaneous long-term gaps in technology, institutions, work and lifestyles, rather than a single quick indicator.
Computing could one day be organized as a utility service, much like the telephone network: subscribers connect over communication lines, pay only for the power they use, and gain access to the capabilities of a larger system.
Announced
Deadline
No calendar deadline
What we measure
Access to computing and applications as a remote, pay-per-use service without the user owning the entire infrastructure
Source
Archival biographical recording of the performance
Editorial analysis
Meaning, limitations and verification criterion
What it means
The forecast anticipated the economic and technical model of cloud computing: remote sharing, network connectivity, and pay-as-you-go. This is a separate idea from McCarthy's 1970 home information terminal.
Interpretation boundary
A remote server or a fixed subscription does not fulfill all the features of a utility model. The modern cloud is also not necessarily a single public good or regulated utility infrastructure.
Verification protocol
Check separately remote access, infrastructure sharing between subscribers, elasticity, consumption pricing and access to software services. The social status of a service should be assessed separately from technical similarity.
Mathematician, cryptanalyst and artificial intelligence researcher
Recorded wording
If an ultra-intelligent machine emerges, it will be able to design even better machines and launch an “intelligence explosion”; the first such machine may be the last invention man needs, if it remains sufficiently controllable.
Announced
Deadline
No calendar deadline
What we measure
The emergence of a system superior to humans in all intellectual activities and capable of creating stronger intellectual systems
Source
Primary scientific work, university archival copy
Editorial analysis
Meaning, limitations and verification criterion
What it means
Good described not just a very smart machine, but a recursive mechanism: the system designs an improved successor, after which the rate of intellectual progress increases dramatically. The controllability condition was part of the original formulation.
Interpretation boundary
Self-improvement of a single algorithm, automatic selection of hyperparameters, or a model that outperforms humans in one area does not yet constitute the explosion of intelligence described.
Verification protocol
First, define a broad set of intellectual areas and a criterion for superiority over the best people. Then check to see if the system created a stronger successor with minimal human design and repeat the cycle of improvement.
Exhibit No. 129The deadline is formulated approximately
J. C. R. Licklider
Psychoacoustician and interactive computing pioneer
Recorded wording
The expected human-computer collaboration would closely connect humans and machines: the human sets goals, hypotheses and criteria, and the computer does the routine preparatory work for decisions and discoveries.
Announced
Deadline
"In not too distant years"
What we measure
Sustainable collaboration, where the computer helps formulate problems and make decisions without rigid dependence on a pre-recorded program
Source
IEEE Primary Research Paper
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is a separate forecast from the same article as the museum card about network “thinking centers.” Here, the object of review is the division of intellectual work between humans and computers, not the access infrastructure.
Interpretation boundary
A calculator, search or chat with a model does not in itself confirm symbiosis. Two-way interactivity, participation in setting a not yet formulated task, and maintaining different roles of partners are important.
Verification protocol
Observe the full work cycle: who sets the goal, forms hypotheses, prepares data, proposes a solution and evaluates the result. Check the increase in quality and speed relative to a person and a computer separately.
Changing work processes could open up a task volume equivalent to 300 million full-time jobs worldwide to automation.
Announced
Deadline
Not named
What we measure
FTE equivalent of tasks subject to automation
Source
Official Study Review
Editorial analysis
Meaning, limitations and verification criterion
What it means
The authors aggregated the shares of tasks within professions. The result was the amount of working time in terms of full employment, and not a list of 300 million people who will lose their jobs.
Interpretation boundary
The report itself says that most professions are only partially affected and will likely be supplemented by AI. Equivalent working hours cannot be translated into layoffs on a one-to-one basis.
Verification protocol
Compare actually automated tasks, employment and creation of new tasks for the same professions. Separately count reduced hours and reduced people.
A model for the future of work in Europe and the US
Recorded wording
In average scenarios, about 27% of current working hours in Europe and 30% in the US could be automated by 2030.
Announced
Deadline
What we measure
Proportion of working hours and transitions between professions
Source
Official MGI report
Editorial analysis
Meaning, limitations and verification criterion
What it means
The model considers automated activities and implementation speed. At the same time, she expects an increase in demand for STEM and other high-skilled professions and up to 12 million career transitions in each region.
Interpretation boundary
Thirty percent of hours does not mean thirty percent of positions gone. The freed up time can change the composition of the work, the price of the service and the volume of demand.
Verification protocol
By 2030, compare the actual share of automated hours, the number of job transitions, and changes in employment. Save exactly the midpoint scenario and geography of the report.
In a sample of OECD countries, occupations with the highest risk of automation accounted for 27% of employment, but there has not yet been a widespread slowdown in labor demand due to AI.
Announced
Deadline
Not named
What we measure
Employment in the top automation risk group
Source
Official OECD report
Editorial analysis
Meaning, limitations and verification criterion
What it means
The indicator ranks professions based on the reproducibility of skills and abilities by automation technologies. It helps look for areas of pressure, but does not predict the literal disappearance of every position in the group.
Interpretation boundary
Twenty-seven percent is the share of employment in occupations in the top risk group, not the probability of layoffs of 27% of workers. High exposure to AI can also mean labor addition.
Verification protocol
Monitor employment, hiring, hours and wages in at-risk groups relative to comparable occupations, taking into account the creation of new tasks and delays in implementation.
Staff Discussion Note on GenAI and the Future of Work
Recorded wording
AI affects almost 40% of global employment and about 60% of employment in developed economies; about half of the exposure there can be complementary to humans.
Announced
Deadline
Not named
What we measure
Exposure of professions and potential complementarity
Source
Official IMF Policy Brief
Editorial analysis
Meaning, limitations and verification criterion
What it means
The same technical exposure is divided into substitution risk and productivity growth opportunity. Therefore, a high percentage in developed countries simultaneously means more risk and more potential for supplementation.
Interpretation boundary
Forty percent of employment affected does not equal forty percent of jobs lost. This includes professions where AI changes some of the tasks, but the person remains central.
Verification protocol
Compare employment, wages and productivity of high-exposure occupations with high and low complementarity separately, maintaining the breakdown by country group.
For GenAI's deep integration, the model produced a range from zero job losses with full addition to 7.9 million job losses with full displacement.
Announced
Deadline
What we measure
Scenario job losses and UK GDP growth
Source
IPPR Official Report and Press Release
Editorial analysis
Meaning, limitations and verification criterion
What it means
Extreme results were obtained for one technology under different decisions of employers and the state. The main prediction here is not the point, but the dependence of the result on the method of implementation.
Interpretation boundary
The 7.9 million number is a worst-case scenario, not a central estimate. The central scenario of the model gave 4.4 million, the best - zero losses and GDP growth.
Verification protocol
By the five-year horizon, fix the share of addition and displacement of tasks, actual employment and GDP. Compare the result with the entire range, and not just with the scary maximum.
AI Jobs Transition Framework for 921 US Occupations
Recorded wording
About 18% of jobs are at high risk of automation, 24% are at risk of disruption, 12% are likely to grow with AI, and 46% are less likely to change in the near future.
Announced
Deadline
Not named
What we measure
Four trajectories of occupational change
Source
Official OpenAI methodology and report
Editorial analysis
Meaning, limitations and verification criterion
What it means
Software developers are relegated not to extinction, but to reorganization: code is speeding up, but professional judgment, responsibility, exceptions and relationships remain with the people.
Interpretation boundary
The authors explicitly caution that the categories are not a predictor of the share of jobs lost. Technical ability still requires integration, economic benefit and employer decision.
Verification protocol
Observe task composition, entry role trajectories, human control requirements, and occupancy for each category. Don't mix risk, reorganization and growth into one percent automation.
Analytical forecast of the impact of AI on employment in the United States
Recorded wording
Between 2025 and 2030, automation and AI could lead to the loss of approximately 10.4 million roles in the US; at the same time, AI should complement human work in 20% of jobs.
Announced
Deadline
What we measure
Automated roles, waste share and augmented jobs
Source
Official press release of the study
Editorial analysis
Meaning, limitations and verification criterion
What it means
The card deliberately preserves the problem of the original source: the headline calls the 6% share of US jobs, and the body text calls the share of all job losses, equating it to 10.4 million roles. These denominators are not identical.
Interpretation boundary
6% automation and 20% augmentation cannot be added together. Addition of tasks does not mean the disappearance of a position, and the loss of a role does not always mean long-term unemployment for a particular person.
Verification protocol
First, consolidate the definition of 6% in the full report. Then, by the end of 2030, separately examine the number of automated roles, the proportion of associated losses, and the proportion of jobs where AI has changed tasks.
Authors of a study on the susceptibility of professions to computerization
Recorded wording
About 47% of U.S. employment was in occupations with a high estimated probability of full computerization—potentially within an uncertain horizon, “perhaps a decade or two.”
Announced
Deadline
What we measure
Share of employment in 2010 in occupations with a probability of full computerization above 0.7
Source
Primary scientific work
Editorial analysis
Meaning, limitations and verification criterion
What it means
The authors assessed the technological susceptibility of 702 occupations, not the number of future layoffs. They did not directly try to predict how many jobs would actually be automated.
Interpretation boundary
Automating some tasks, using AI, or reducing employment for any reason does not support the 47% number. The risk of the profession cannot be translated one to one into unemployment of people.
Verification protocol
Maintain the 2010 occupational and employment database, define a test for full computerization, and evaluate technical feasibility separately from implementation. Do not turn the author's soft guideline into the promised deadline.
Corporate foresight study with the participation of more than 20 experts
Recorded wording
The study estimates that 85% of the jobs of 2030 had not yet been invented in 2017.
Announced
Deadline
What we measure
Share of 2030 jobs classified as jobs that did not exist in 2017
Source
Official press release of the study
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is a well-known foresight thesis, but the public material does not define the “invented work,” the denominator, or the procedure for obtaining 85%. Its museum value is to show the limit of verifiability of the viral percentage.
Interpretation boundary
New job titles, new skills within an old profession, and new tasks are not the same thing. They cannot automatically be considered new jobs and added up to 85%.
Verification protocol
Until 2030, fix the classification: for example, the share of people employed in occupational codes that were not in the comparable classification of 2017. Without such operationalization, the exact percentage will remain unverifiable.
By 2030, 70% of the skills used in most jobs will have changed; artificial intelligence will be one of the main catalysts.
Announced
Deadline
What we measure
Share of Skills Changed in LinkedIn Taxonomy for Most Jobs
Source
Official summary of platform research
Editorial analysis
Meaning, limitations and verification criterion
What it means
The forecast describes the renewal of skill sets within jobs, rather than the disappearance of a corresponding proportion of jobs. Its denominator depends on LinkedIn's skill taxonomy, user profiles, and job openings.
Interpretation boundary
70% of skills changed cannot be read as 70% of people being laid off or jobs being automated. Users on a single platform do not represent the entire global workforce without adjustment.
Verification protocol
Save the taxonomy version and the 2025 sample, determine the threshold for “change” in a skill and repeat the calculation in 2030 for comparable professions; separately show the contribution of AI and other factors.
World Bank Study on Growth, Technology and Employment
Recorded wording
4.5% of existing jobs in low- and middle-income countries are amenable to generative automation versus 14.2% in rich ones; AI addition - 16.2% versus 18.7%.
Announced
Deadline
Not named
What we measure
Share of existing jobs whose tasks can be automated or augmented by generative AI
Source
Official release of the world report with reproducibility package
Editorial analysis
Meaning, limitations and verification criterion
What it means
The comparison shows lower technical exposure in developing economies where there is more manual labor. The potential for AI augmentation is closer to that of rich countries, especially in digital sectors.
Interpretation boundary
“Amenable to automation” does not mean that the job will be automated, disappear or be reduced. The assessment does not specify the pace of implementation, cost, infrastructure or final employment balance.
Verification protocol
When updating the report, reproduce the task classification on the published dataset, check for differences by income group, and compare exposure to actual implementation and employment—without substituting one for the other.
Application Development and DevOps Market Forecast
Recorded wording
By 2028, natural language will become the most widely used programming language and will be involved in the creation of 55% of new applications.
Announced
Deadline
What we measure
Share of net-new applications in which natural language is used as a programming interface
Source
Official Business Forecast Summary
Editorial analysis
Meaning, limitations and verification criterion
What it means
The statement brings programming to the level of intent: a person describes the result in natural language, and tools create or change software components. The number refers to applications, not lines of code.
Interpretation boundary
Natural language does not become a formal execution language in the usual sense. Participating in the creation of an application does not mean generating most of its code and does not prove the disappearance of the developer.
Verification protocol
By 2028, define a “new application” unit, a minimum threshold for natural language participation, and deduplication rules. Then reproduce the share on a comparable sample and separately measure the volume of generated code.
In the absence of wars and disasters and with favorable political and economic development, the probability of human-level machine intelligence will be 10% by 2018, 50% by 2028 and 90% by 2050.
Announced
Deadline
What we measure
Conditional probability distribution: 10% / 50% / 90% for three dates
Source
Public Q&A with direct response from the author
Editorial analysis
Meaning, limitations and verification criterion
What it means
Legg did not promise a 2028 event. He recorded the distribution of uncertainty and directly made the forecast dependent on external conditions.
Interpretation boundary
The fact that 2018 passed without generally accepted human-level AI does not refute the distribution with a probability of 10%. And vice versa, 50% by 2028 cannot be assessed with just one “fulfilled” mark.
Verification protocol
By 2028, determine the operational criterion of human-level machine intelligence and the rules for taking into account conditions. Calibration can only be assessed using a series of comparable probabilistic forecasts.
By the end of 2030, training launches on the order of 2×10²⁹ FLOPs are likely to be technically feasible—about ten thousand times larger than the cutting-edge launches of mid-2024.
Announced
Deadline
What we measure
Maximum technically feasible amount of computation for one training run, taking into account energy, chips, data and delays
Source
Primary Research
Editorial analysis
Meaning, limitations and verification criterion
What it means
The forecast answers the question of the technical feasibility of the scale, and not whether the company will decide to pay for such a launch and what capabilities the model will receive. The authors separately consider four physical limitations.
Interpretation boundary
The total calculations of many models, the capacity of the data center built, or the announced budget do not equal one completed 2×10²⁹ FLOP run. Feasibility does not guarantee actual learning.
Verification protocol
For the largest launches in 2030, collect verifiable estimates of the number of accelerators, their performance, load, and duration. Separately evaluate the technical limit and the largest actually completed launch.
Working Study; the conclusions are not the official position of the IMF
Recorded wording
Ten years from now, global GDP in the model could be 1.3% above the baseline path if productivity growth is low, or almost 4% above the baseline if productivity growth is high; after five years the range is 0.8–2.4%.
Announced
Deadline
What we measure
Deviation of the level of world GDP from the baseline trajectory in the low and high scenarios of total factor productivity growth from the IMF model
Source
IMF Working Study
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is the conditional range of model results given assumptions about productivity, access to AI, and economic response. It shows the possible macroeconomic impact, rather than a guaranteed increase on top of any future GDP.
Interpretation boundary
The usual growth of the world economy by a few percent does not confirm the scenario: growth relative to a comparable base trajectory is needed. The upper limit cannot be passed off as an official unconditional forecast by the IMF.
Verification protocol
In 2030 and 2035, restore the baseline path and decompose the actual variance into productivity, capital, demand, and other shocks. Assess both scenarios and differences between country groups.
If widely adopted, generative AI could increase global GDP levels by about 7%, or nearly $7 trillion, and add about 1.5 percentage points to productivity growth over a ten-year period.
Announced
Deadline
What we measure
Deviation of global GDP and average annual productivity growth from a comparable baseline trajectory with widespread AI adoption
Source
Official review of economic research
Editorial analysis
Meaning, limitations and verification criterion
What it means
From the same study, the museum already has a card about 300 million FTE-equivalent tasks susceptible to automation. The new card stores another result of the model - a macroeconomic growth scenario.
Interpretation boundary
The usual 7% global GDP growth does not support the forecast: the increment is measured relative to a world without the AI implementation in question. The potential for widespread adoption of the technology is not a guaranteed outcome.
Verification protocol
In April 2033, construct a comparable counterfactual path, estimate actual adoption, and decompose growth into AI, capital, labor, and other shocks. Check GDP and productivity separately.
Taking into account difficult-to-learn tasks, total factor productivity due to AI is estimated by the author to grow by less than 0.53% over the next ten years.
Announced
Deadline
What we measure
Cumulative addition to US total factor productivity over ten years caused by achievable automation and AI task augmentation
Source
MIT Primary Economics Work
Editorial analysis
Meaning, limitations and verification criterion
What it means
This is a conservative quantitative rate, useful for comparison with higher scenarios from Goldman Sachs, OECD and IMF. The numerator is TFP, not GDP, wages or the share of automated occupations.
Interpretation boundary
0.53% per decade cannot be read as 0.53 percentage points annually. Overall productivity growth does not show the contribution of AI without a causal assessment, and the local effects of individual tools cannot be directly transferred to the economy.
Verification protocol
By May 2034, reproduce the author's decomposition of the share of affected tasks and cost savings, update the implementation data and compare the accumulated contribution of AI to TFP with the 0.53% limit.
G7 countries with high exposure and rapid adoption of AI, primarily the US and UK, could gain an additional 0.4–1.3 percentage points in average annual productivity growth over the decade; in some other countries the effect is up to 50% less.
Announced
Deadline
What we measure
Average annual increment to aggregate labor productivity growth due to AI by country and deployment scenario
Source
Official OECD Working Paper
Editorial analysis
Meaning, limitations and verification criterion
What it means
The range of 0.4–1.3 does not apply to the entire G7 equally, but to the most AI-prone economies across the three scenarios. The difference between countries depends on the structure of industries and the expected speed of adoption.
Interpretation boundary
This annual pace cannot be directly compared with Acemoglu's accumulated TFP or the level of global GDP in IMF scenarios. The fact that productivity is growing without assessing the underlying trajectory is also insufficient.
Verification protocol
For each G7 country, reconstruct the original implementation scenario, measure the average annual increase in labor productivity over comparable ten years, and test the high and low bounds separately.
Mathematician, author of the article Computing Machinery and Intelligence
Recorded wording
In about fifty years, a machine will be able to complete a five-minute text-based “imitation game” so that the average judge will identify it correctly no more than 70% of the time.
Announced
Deadline
What we measure
Proportion of correct decisions by judges after five minutes of text dialogue
Source
Original article in Mind magazine
Result source published
Basis for the verdict
Independent research; the protocol is different
Editorial analysis
Meaning, outcome, caveats and verification criterion
What it means
Turing named a specific behavioral threshold and deadline. Therefore, the prediction can be compared with experiments if they actually reproduce a short text game with a person and a machine.
Interpretation boundary
Successful imitation of a human does not prove consciousness, general intelligence, or the ability to replace a human in all types of work.
What happened
In a preregistered study published in May 2026, GPT-4.5 with a persona cue was selected as a person on 73% of five-minute trials. Without such a person, the result of the same model dropped to 36%.
Verdict: Not confirmed by deadline; the later result is comparable
The numerical threshold was passed approximately 26 years later than the named horizon. The modern tripartite protocol and the strong dependence on hints do not allow us to call this an exact repetition of the Turing experiment.
Verification protocol
Compare conversation length, number of judges, selection rules, models and system prompts. A result without a description of the protocol should not be considered a direct verification of the prediction.
Head of research at Fairchild Semiconductor, future co-founder of Intel
Recorded wording
By 1975, a cost-effective integrated circuit could contain about 65,000 components—about a thousand times more than in 1965.
Announced
Deadline
What we measure
Number of components on one integrated circuit at minimum component cost
Source
Original Electronics article, Intel archive
Result report published
Basis for the verdict
Author's analysis and technical data
Editorial analysis
Meaning, outcome, caveats and verification criterion
What it means
This is a rare ten-year technical forecast with a pre-named value. It referred to the density and economics of chip production, not directly to the speed of programs or the development of AI.
Interpretation boundary
The coincidence of one curve does not mean that exponential growth must continue forever or is equally transferred to the quality of models, power consumption and cost of calculations.
What happened
In a 1975 report, Moore added factual data and wrote that complexity was growing at a roughly doubling annual rate. The resulting order of magnitude coincided with the line drawn ten years earlier.
Verdict: The trend was confirmed; the exact metric is limitedly comparable
Counting components, transistors, and memory bits are not completely interchangeable. In addition, the well-known forecast has gradually become a reference point for the semiconductor industry itself, so the result cannot be considered an independent natural law.
Verification protocol
Compare the original definition of "component" and minimum cost criterion to the 1975 devices without replacing them with the later formula of doubling the number of transistors every two years.
During 1996, the Internet would experience a catastrophic collapse.
Announced
Deadline
What we measure
The initial criterion is “catastrophic collapse”; later clarification by the author - more than a billion lost user hours
Source
Scan of the original InfoWorld column
Result period
Outcome source dated
Basis for the verdict
Later author's analysis and retrospective criterion
Editorial analysis
Meaning, outcome, caveats and verification criterion
What it means
In the original column, Metcalfe predicted a catastrophic collapse without a numerical threshold. He later described gigalapse - a billion lost user hours; the museum preserves this as a subsequent clarification and not as part of the 1995 text.
Interpretation boundary
An erroneous prediction of a crash does not prove that the Internet was reliable at all points or that the technical risks listed by the author were fictitious.
What happened
In an interview with the Computer History Museum, Metcalfe pegged the largest outage he knew of, in 1996, at about 118 million lost user hours—about one-eighth of the promised threshold. Afterwards, he publicly ate the printed column.
Verdict: Didn't come true; numerical estimate based on late refinement
The numerical criterion was explained in detail by the author later, and some of the original columns were written with sarcasm. Even with this caveat, the claimed billion hours were not achieved.
Verification protocol
Check two layers separately: the qualitative formulation of the 1995 column and the author’s later numerical criterion. For the second, estimate the number of affected users and the duration of the major outages of 1996.
In 2021, Ford planned to bring SAE Level 4 to commercial fleets for ridesharing, ride-hailing or delivery—without a steering wheel or pedals in a designated geofence.
Announced
Deadline
What we measure
Commercial fleet of SAE Level 4 vehicles rather than research trips
Source
Ford official report 2016/17
Result source published
Basis for the verdict
Corporate report after the deadline; indirect comparison
Editorial analysis
Meaning, outcome, caveats and verification criterion
What it means
The application contains the level of autonomy, commercial format and calendar year. Therefore, test cars with a safety driver cannot be counted as fulfilling the plan.
Interpretation boundary
Missing the deadline doesn't mean that autonomous driving research has stopped or that driver assistance systems haven't been improved.
What happened
In its third-quarter 2022 report, Ford said profitable mass-scale fully autonomous vehicles were further behind expectations and decided to stop investing in Argo AI, reallocating resources toward L2+/L3 systems.
Verdict: Not completed on time
This is a post-deadline corporate source and indirect comparison, not a separate audit of the 2021 promise. It does not show the launch of the announced commercial L4 fleet; the original document was a company plan, not a probabilistic forecast.
Verification protocol
Seek confirmation of the paid availability of a commercial service, the declared SAE level and operation without a driver in the permitted area until the end of 2021.
By mid-2020, there should have been more than a million Teslas on the road with full self-driving hardware that could operate as robotaxis.
Announced
Deadline
What we measure
More than a million active robotaxis, not sold cars with a set of sensors
Source
Official Tesla video
Result period
Outcome source dated
Basis for the verdict
Regulatory corporate reporting
Editorial analysis
Meaning, outcome, caveats and verification criterion
What it means
The key criterion is the ability of the machines to independently transport passengers in the network. Having driver assistance equipment or beta features is not the same as a functioning robotaxi.
Interpretation boundary
Missing the deadline does not answer the question of whether Tesla or another company will be able to deploy such a network at a later date.
What happened
In its 2020 annual report, Tesla continued to describe features as requiring active driver control: the driver remained in charge of control. The robotaxi network was mentioned as a future possibility rather than a million cars running.
Verdict: Not completed on time
Selling cars with suitable sensors does not fulfill the stated criterion. Regulatory differences also do not turn the driver assistance system into a mass commercial unmanned service.
Verification protocol
Check the actual number of vehicles approved for fully unmanned commercial passenger transportation without permanent driver responsibility.
Corporate evaluation of the effect of AI support assistant
Recorded wording
The AI assistant was expected to improve Klarna's earnings by approximately $40 million over the course of 2024.
Announced
Deadline
What we measure
Monetary effect of AI assistant in 2024
Source
Official press release from Klarna
Result period
Outcome source dated
Basis for the verdict
Self-assessment of the company in the regulatory document
Editorial analysis
Meaning, outcome, caveats and verification criterion
What it means
The forecast has a short duration and monetary value, and the company later disclosed a similar number in a regulatory filing. This is a convenient example of checking a corporate promise with your own reporting data.
Interpretation boundary
The number does not prove that the entire effect is caused by the model alone, that the quality is the same for all calls, or that the equivalent of agents' work is equal to the number of actual layoffs.
What happened
In the registration document, Klarna estimated cost savings from the assistant for 2024 at approximately $39 million - a value close to the reported $40 million.
Verdict: Close number by another metric; company self-assessment
“Profit improvement” and “cost savings” are related, but not identical. Both numbers were calculated by the company itself, so the correct verdict is a close match in scale, rather than independent confirmation of an accurate forecast.
Verification protocol
Verify the calculation methodology, period, currency and the difference between profit improvement and cost savings; Whenever possible, seek independent audits of effect attribution.
Artificial Intelligence Pioneers; authors of a work on problem solving by humans and machines
Recorded wording
Within ten years, the digital machine will become the world chess champion, unless the rules of the competition prohibit it from participating.
Announced
Deadline
What we measure
Computer victory over the current world champion in a standard match
Source
Primary scientific publication
Result date
Basis for the verdict
Chronology of the Computer History Museum
Editorial analysis
Meaning, outcome, caveats and verification criterion
What it means
This is a rare early forecast with a specific deadline and public criterion. For a fair test, the museum uses the closest observable equivalent: a computer system victory over the reigning world champion in a standard rules match.
Interpretation boundary
Victories over individual grandmasters, success in computer tournaments and one won game are not equal to victory over the current world champion in a full match. Even such a victory does not prove the general intelligence of the machine.
What happened
In May 1997, Deep Blue became the first computer system to defeat reigning world champion Garry Kasparov in a match under standard tournament time control.
Verdict: Deadline missed; the closest equivalent appeared approximately 29.5 years later
Deep Blue misses out on human world title: IBM records victory over reigning champion. This is a strong operational equivalent of a forecast, but it occurred almost three decades after the deadline and does not make the original deadline true.
Verification protocol
Compare the deadline of November 14, 1967 with the first officially documented match in which a computer defeated the reigning world champion under standard time control.
Architect Deep Blue; specialized computing systems researcher
Recorded wording
A world-champion Go machine could be built within ten years using intensive analysis—“essentially brute force”—modeled on Deep Blue.
Announced
Deadline
What we measure
World Go champion level system within ten years
Source
Primary author's article
Result date
Basis for the verdict
Official history of the system developer
Editorial analysis
Meaning, outcome, caveats and verification criterion
What it means
One formulation combines two different predictions: achieving a champion level and a specific path through a Deep-Blue-like intensive search. The museum checks them separately.
Interpretation boundary
Beating an amateur or mid-level professional does not equal championship level. But getting to the level of the game does not in itself confirm the computational method proposed by the author.
What happened
In March 2016, AlphaGo defeated Lee Sedol, winner of 18 world titles, 4:1 - before the end of the stated ten-year window.
Verdict: The level was reached approximately 19 months earlier; the proposed method was not confirmed
AlphaGo combined deep neural networks, advanced search, and reinforcement learning rather than replicating Deep Blue through brute force alone. The source of the result belongs to the developer, but the match itself and the score are publicly recorded.
Verification protocol
Find a documented match of the system with a world champion-level player before the end of October 2017, then separately compare the system architecture with the method described in the article.
Exhibit No. 062The main deadline has passed October 2021
Geoffrey Hinton
Machine learning researcher; Turing Award and Nobel Prize winner
Recorded wording
In 2016, Hinton called for an end to radiology training: within five years, he said, deep learning would make better radiologists; he then admitted that it could take ten years.
Announced
Deadline
What we measure
Interpretation of medical images and practical implications of discontinuing specialist training
Source
Video recording of a public speech
Result period
Outcome source dated
Basis for the verdict
Census of 100% of UK radiology managers
Editorial analysis
Meaning, outcome, caveats and verification criterion
What it means
The statement makes a narrow technical thesis about image reading and a broad personnel thesis about the profession. A strong model result in one task does not make these theses equivalent.
Interpretation boundary
The superiority of an algorithm on a selected set of images does not mean that it does all the work of a clinical radiologist. Regulatory approval of the tool also does not prove the disappearance of the need for specialists.
What happened
According to 2025 data, the UK was short of 32% of clinical radiologists - more than 2,300 specialists; the college demanded that training be expanded rather than stopped.
Verdict: The personnel conclusion was not confirmed; technical metric has not been defined
The census reliably refutes the practical conclusion that training in the selected country is unnecessary, but does not compare the accuracy of the algorithms and doctors. Hinton's additional ten-year clause technically remained open until October 27, 2026.
Verification protocol
The technical part requires a predetermined broad set of clinical tasks and independent comparison. For the personnel part - data on training, employment, vacancies and the need for radiologists after the deadline.
Roboticist, co-founder of iRobot and former director of MIT CSAIL
Recorded wording
The next generally accepted big idea in AI after deep learning, based on already published work, will appear no earlier than 2023 and no later than 2027.
Announced
Deadline
What we measure
The emergence of a generally accepted next trend in AI from work published no later than 2018
Source
Primary author's table of dated forecasts
Result period
Outcome source dated
Basis for the verdict
Author's retrospective review with named original work
Editorial analysis
Meaning, outcome, caveats and verification criterion
What it means
Brooks set not only a deadline, but also a lower bound, and also required that the basic idea already exist in the scientific literature. This makes the time part of the forecast noticeably more accurate than the usual “soon”.
Interpretation boundary
Product popularity does not equal scientific consensus about a new paradigm. In addition, LLMs are built on deep learning, so their status as an idea “after deep learning” depends on interpretation.
What happened
In his review, Brooks placed the next big idea in 2023: LLMs had gained mainstream acceptance, and their key architectural source—the 2017 Transformer paper—predated the prediction.
Verdict: The time window has been met; the definition of a new paradigm remains controversial
The result is assessed by the author himself, and the “generally accepted” criterion was not quantitatively specified. LLMs are deep learning systems, so the museum counts a temporary hit, but not full confirmation of the literal thesis “after deep learning.”
Verification protocol
Compare the date of mass recognition of the direction with the window of 2023–2027, find the original work no later than 2018, and separately check whether experts really consider the direction to be the next stage, and not the scaling of deep learning.
Co-founder of the MIT Artificial Intelligence Laboratory, Turing Award winner
Recorded wording
In 3-8 years, there will be a machine with the general intelligence of the average person: it will be able to read Shakespeare, oil a car, engage in office politics, joke and fight.
Announced
Deadline
What we measure
One machine that combines the listed intellectual, everyday and social abilities
Source
Scan of the original issue of LIFE
Result source published
Basis for the verdict
The author's own later assessment
Editorial analysis
Meaning, outcome, caveats and verification criterion
What it means
This is not a prediction of a single test or a single model. Minsky described a general machine that would connect text comprehension, physical actions and behavior among people.
Interpretation boundary
The success of a system in chess, dialogue or one everyday action does not confirm the entire set. The current AGI name also cannot be automatically substituted in place of the original description.
What happened
In a 2005 draft, Minsky wrote that no machine can yet read a book, clean a house, or look after a child. Even this later benchmark did not live up to the overall machine's promised set.
Verdict: At the upper limit of the deadline did not come true
The source of the result was published 27 years after the deadline and is not a formal comparative test. Therefore, it reliably shows a missed deadline, but does not measure the exact distance to the “intelligence of the average person.”
Verification protocol
Check the upper limit of the window - November 1978 - according to the state of the unified system. Require demonstration of the entire set without switching between unrelated specialized programs.
Predictions, scenarios, targets and corporate plans are shown together. The date marks the public statement or editorial snapshot, not necessarily the first appearance of the idea.
If the relationships of musical sounds can be expressed by formal rules, the Analytical Engine will be able to compose complex scientifically constructed musical works of any complexity and duration.
In a hundred years, the standard of living in progressive countries will be 4–8 times higher; the remaining work can be distributed in three-hour shifts, or 15-hour weeks.
A person will be able to request the necessary pages of books from a large library from home and read them on the screen using an “electric telescope”; first within the city, then over long distances.
Humanity needs a constantly updated World Encyclopedia: a common, accessible and authoritative body of knowledge, linked to primary sources, for learning, management and collaboration by people around the world.
Announced: August 1937 · deadline: No calendar deadline
The future personal Memex will allow a person to store books, records and messages, quickly find them, link materials into associative routes and transfer such routes to other people.
Announced: July 1945 · deadline: No calendar deadline
A flexible system of computers, sensors and control mechanisms should lead to factories without workers; a critical situation with industrial employment, according to Wiener, could arise in ten to twenty years.
An electronic computer with a program for searching and evaluating positions will be able to play fairly strong chess at a speed comparable to the speed of a person, although a complete search of all games is practically impossible.
Announced: March 1950 · deadline: No calendar deadline
In about fifty years, a machine will be able to complete a five-minute text-based “imitation game” so that the average judge will identify it correctly no more than 70% of the time.
Announced: October 1950 · deadline: Around 2000
Result source published: May 19, 2026 · Not confirmed by deadline; the later result is comparable
Machines will eventually rival humans in all purely intellectual fields; According to Turing, it was necessary to start with an abstract game and teach the machine to understand and use language.
Announced: October 1950 · deadline: No calendar deadline
The authors suggested that any aspect of learning and intelligence could in principle be described by a machine, and ten researchers could make significant progress in one summer.
Announced: August 31, 1955 · deadline: Summer 1956
The authors estimated the reduction in coding time against assembly to be about 4 to 20 times and showed that 47 lines of FORTRAN replaced about 1,000 machine instructions.
Within ten years, the digital machine will become the world chess champion, unless the rules of the competition prohibit it from participating.
Announced: November 14, 1957 · deadline: November 14, 1967
Result date: May 11, 1997 · Deadline missed; the closest equivalent appeared approximately 29.5 years later
Accelerating technological progress and changing lifestyles create the impression of approaching a significant historical singularity, after which human affairs can no longer continue in the same way.
Announced: published May 1958 · deadline: No calendar deadline
In 10–15 years, a “thinking center” should appear, combining the functions of a library, storing and retrieving information, computing and a network of similar centers for remote users.
The expected human-computer collaboration would closely connect humans and machines: the human sets goals, hypotheses and criteria, and the computer does the routine preparatory work for decisions and discoveries.
Announced: March 1960 · deadline: "In not too distant years"
In 1960, Simon wrote that within twenty years machines would technically be able to do any human job—though economically, humans would retain a comparative advantage.
Computing could one day be organized as a utility service, much like the telephone network: subscribers connect over communication lines, pay only for the power they use, and gain access to the capabilities of a larger system.
Computer tools should not just automate individual operations, but enhance human intelligence: speed up and deepen understanding of complex situations, help find faster and better solutions.
Announced: October 1962 · deadline: No calendar deadline
By 2014, machines will do almost any routine job better than humans, so people will mostly be left to maintain machines, with the most valuable exception being creative work.
If an ultra-intelligent machine emerges, it will be able to design even better machines and launch an “intelligence explosion”; the first such machine may be the last invention man needs, if it remains sufficiently controllable.
The ultimate development of a computer display should be an environment in which the machine controls not only the image, but also the perceived properties of objects: a virtual chair should feel like a chair, and the world created by the program should feel like a space that can be entered.
Announced: May 1965 · deadline: No calendar deadline
A computer system for personal documents and creativity should allow a person to build complex connections of his own, freely change the structure, save alternatives and work with interrelated text and graphic material that is inconvenient to represent on paper.
Announced: August 1965 · deadline: No calendar deadline
A home information terminal with a keyboard and screen will access a shared computer through the telephone network and provide instant access to books, newspapers, catalogs, schedules and personal files.
In 3-8 years, there will be a machine with the general intelligence of the average person: it will be able to read Shakespeare, oil a car, engage in office politics, joke and fight.
Announced: November 20, 1970 · deadline: No later than November 20, 1978
Result source published: July 28, 2005 · At the upper limit of the deadline did not come true
A personal computer should belong to the user, be portable, cost no more than a television, and serve simultaneously as a medium of expression, a set of modifiable tools, a book, a notebook, and a communication channel.
Announced: August 1972 · deadline: No calendar deadline
By 2001, a home computer console linked to other computers will make it possible to access banking information, book flights, and do work from home that previously required an office presence.
By 2006 or earlier, a new production system should emerge in which the home, with its low-cost workstation, computer communications, and teleconferencing, will once again become an important place of work.
If productivity doesn't improve, Martin wrote, within ten years the industry will need 93.1 times more programmers—about 28 million—with a 41-fold increase in the number of applications.
The personal computer will become the new primary means of communication, entering the work, school and home environments, and portable computers with radio communications will allow you to receive messages and data on the go; the first phase of mass distribution will be completed in about fifteen years.
In the next ten years, there will not be one technology or management technique that, on its own, promises a tenfold improvement in the productivity, reliability and ease of software development.
Announced: September 1986 · deadline: September 1996
CERN should create a universal system of linked data with remote access, support for different computers and without a single center; in ten years, Berners-Lee suggested, there could be many commercial solutions for this problem.
Computers should disappear into the ordinary environment: many specialized devices of different sizes will be connected by a network and will become so familiar that a person will no longer perceive interaction with them as a separate work at the computer.
Announced: September 1991 · deadline: No calendar deadline
Within thirty years the technological means to create superhuman intelligence will be available; soon after this, according to the author's forecast, the human era will end.
Announced: March 31, 1993 · deadline: March 31, 2023
During 1996, the Internet would experience a catastrophic collapse.
Announced: December 4, 1995 · deadline: Until the end of 1996
Result period: Results for 1996 · source: 2006 · Didn't come true; numerical estimate based on late refinement
The singularity will occur in 2045; The non-biological intelligence created by this moment will be a billion times more powerful than the total human intelligence of the beginning of the 21st century.
A world-champion Go machine could be built within ten years using intensive analysis—“essentially brute force”—modeled on Deep Blue.
Announced: October 2007 · deadline: October 2017
Result date: March 15, 2016 · The level was reached approximately 19 months earlier; the proposed method was not confirmed
In the absence of wars and disasters and with favorable political and economic development, the probability of human-level machine intelligence will be 10% by 2018, 50% by 2028 and 90% by 2050.
Announced: June 17, 2011 · deadline: Median point - 2028
About 47% of U.S. employment was in occupations with a high estimated probability of full computerization—potentially within an uncertain horizon, “perhaps a decade or two.”
Announced: September 17, 2013 · deadline: Target 2023–2033, no hard deadline
In 2021, Ford planned to bring SAE Level 4 to commercial fleets for ridesharing, ride-hailing or delivery—without a steering wheel or pedals in a designated geofence.
Announced: August 16, 2016 · deadline: 2021
Result source published: October 26, 2022 · Not completed on time
In 2016, Hinton called for an end to radiology training: within five years, he said, deep learning would make better radiologists; he then admitted that it could take ten years.
Announced: October 27, 2016 · deadline: October 27, 2021 · Disclaimer: Possibly 2026
Result period: 2025 · source: June 18, 2026 · The personnel conclusion was not confirmed; technical metric has not been defined
Neural network weights are becoming Software 2.0: instead of explicitly writing all the instructions, people specify the data, the goal, and the architecture, and optimization finds a program that works.
Announced: November 11, 2017 · deadline: Not named
The next generally accepted big idea in AI after deep learning, based on already published work, will appear no earlier than 2023 and no later than 2027.
Announced: January 1, 2018 · deadline: Window 2023–2027
Result period: 2023 · source: January 1, 2026 · The time window has been met; the definition of a new paradigm remains controversial
Despite the alarming headline, the author did not predict the disappearance of developers by 2030, but a renewal of the role: less routine, more planning, creativity and leadership.
If widely adopted, generative AI could increase global GDP levels by about 7%, or nearly $7 trillion, and add about 1.5 percentage points to productivity growth over a ten-year period.
Announced: April 5, 2023 · deadline: Ten years · April 2033
Large language models can make the source code a secondary artifact: the natural language becomes the main interface, and the model becomes a new virtual machine.
In a sample of OECD countries, occupations with the highest risk of automation accounted for 27% of employment, but there has not yet been a widespread slowdown in labor demand due to AI.
Within five years, people will no longer have to choose a separate application for each task: the user will explain the goal in ordinary language to a personal AI agent, and agents will become the next computing platform.
Announced: November 9, 2023 · deadline: November 9, 2028
AI affects almost 40% of global employment and about 60% of employment in developed economies; about half of the exposure there can be complementary to humans.
Computing technologies must reach a state where they do not need to be programmed in the traditional way: human language becomes the programming language, and everyone becomes a programmer.
Announced: February 12, 2024 · deadline: No calendar deadline
The AI assistant was expected to improve Klarna's earnings by approximately $40 million over the course of 2024.
Announced: February 27, 2024 · deadline: 2024
Result period: 2024 · source: September 8, 2025 · Close number by another metric; company self-assessment
The total addressable market for humanoid robots could reach $38 billion by 2035; the base case separately assumes more than 250,000 deliveries in 2030, almost entirely to industry.
Announced: February 27, 2024 · deadline: 2030 and 2035
Taking into account difficult-to-learn tasks, total factor productivity due to AI is estimated by the author to grow by less than 0.53% over the next ten years.
Announced: May 12, 2024 · deadline: Ten years May 2034
By the end of 2030, training launches on the order of 2×10²⁹ FLOPs are likely to be technically feasible—about ten thousand times larger than the cutting-edge launches of mid-2024.
Powerful AI could speed up discoveries by at least tenfold and compress the next 50-100 years of progress in biology and medicine into 5-10 years after its introduction.
Announced: October 2024 · deadline: 5-10 years after powerful AI; in the early scenario - approximately 2031–2036
Employers expect 39% of workers' core skills to change by 2030; out of a nominal 100 workers, 59 will need training, and for 11 it may remain unavailable.
41% of surveyed employers expect to reduce the number of employees by 2030 where AI can automate tasks; at the same time, 77% plan to upgrade employee skills, and 47% plan to move people out of affected roles.
By 2030, the total number of employees in the Russian IT industry should grow from almost one million to 1.4 million people; at least 250 thousand students must undergo training with the participation of business.
By 2029, agent-based AI will independently, without human intervention, resolve 80% of typical customer support requests, which will lead to a reduction in operating costs by 30%.
Powerful AI systems could emerge in late 2026 or early 2027: matching or surpassing Nobel laureates in most disciplines, operating through digital interfaces, and autonomously solving complex problems over hours, days, or weeks.
Announced: March 6, 2025 · deadline: End 2026 – beginning 2027
By 2033, the global AI market could grow from $189 billion in 2023 to approximately $4.8 trillion, becoming the largest emerging technology and capturing about 30% of the total market for such technologies.
Global electricity consumption by data centers will increase from approximately 415 TWh in 2024 to 945 TWh in 2030; AI will be the main but not the only source of this growth, and the demand for AI-focused data centers will more than quadruple.
Ten years from now, global GDP in the model could be 1.3% above the baseline path if productivity growth is low, or almost 4% above the baseline if productivity growth is high; after five years the range is 0.8–2.4%.
Announced: April 11, 2025 · deadline: Five and ten years: 2030 and 2035
Within 2–5 years, every organization will begin the path to a “Frontier Firm”—a company where work is built around hybrid teams of humans and AI agents.
Announced: April 23, 2025 · deadline: Window: April 2027 - April 2030
By 2050, more than a billion humanoid robots could be used in the world, about 90% in industry and commerce; the market, including supply chains and services, could exceed $5 trillion.
There is about a 50% chance that DeepMind's definition of AGI will appear in the next 5–10 years. Such an AGI must have the full range of human cognitive abilities.
Announced: June 4, 2025 · deadline: Window: June 2030 - June 2035
In 2026, systems capable of obtaining new scientific insights are likely to appear, and in 2027, robots that perform tasks in the physical world may appear.
Announced: June 10, 2025 · deadline: Two points: 2026 and 2027
G7 countries with high exposure and rapid adoption of AI, primarily the US and UK, could gain an additional 0.4–1.3 percentage points in average annual productivity growth over the decade; in some other countries the effect is up to 50% less.
Announced: June 30, 2025 · deadline: Ten-year horizon June 2035
The global number of new installations of industrial robots should exceed 700,000 units per year by 2028; for 2025, 575,000 installations were expected.
By 2030, AI will impact all IT work: CIOs expect 75% of work will be done by AI-enhanced people, 25% by AI alone, and the proportion of work done by people without AI will drop to zero.
By 2028, with the number of agent deployments growing tenfold, 60% of G2000 companies will use a separate agent development lifecycle to scale and control agents.
Between 2025 and 2030, automation and AI could lead to the loss of approximately 10.4 million roles in the US; at the same time, AI should complement human work in 20% of jobs.
By 2030, 65% of enterprises will integrate AI agents into DevOps and DevSecOps pipelines so that agents execute development and security workflows as an ongoing part of software delivery.
The distinct role of a software engineer may eventually become dissolved among generalists on products, problems, and systems; The creation of programs will not disappear.
If agents make an engineer 2 to 10 times more productive, previously unprofitable software projects will become viable and the demand for engineers will spread across all functions of the business.
About 18% of jobs are at high risk of automation, 24% are at risk of disruption, 12% are likely to grow with AI, and 46% are less likely to change in the near future.
By 2027, more than 65% of engineering teams using agent-based programming will consider the traditional integrated development environment optional and will move management, control and verification to automated platforms.
In agent-based development, human control remains necessary, and newcomers need a preceptor model: real work and safe mistakes instead of disappearing entry into the profession.
The autonomy of the agent should depend on the risk: 60–70% of low-risk pull requests can be automatically checked and merged, but even two lines in the authentication module can be left to a person.
By 2029, 60% of organizations will widely use smaller development teams, up from 15% in 2026. Such teams typically consist of four to five people, with two to three-person teams expected to become more common as the AI's skills and capabilities increase.
Between 2023 and 2035, employment in the EU27 will grow by more than 4%, and employment of ICT specialists and technicians by an average of 2.5% per year.
4.5% of existing jobs in low- and middle-income countries are amenable to generative automation versus 14.2% in rich ones; AI addition - 16.2% versus 18.7%.
.jpg)
.jpg)
.jpg)
_(cropped).jpg)
.jpg)
.jpg)
.jpg)
_(portrait_crop).JPG)
.jpg)
.jpg)
_(cropped).jpg)
.jpg)
.jpg)
_(crop).jpg)
.jpg)
_(cropped_2)_(b).jpg)
.jpg)
