Predictions

“In Part I, we explain why we think that transformative economic and societal impacts will be slow (on the timescale of decades), making a critical distinction between AI methods, AI applications, and AI adoption, arguing that the three happen at different timescales.” source

How we track it. Adoption and adaptation indicators (BTOS firm use, BBD hours assisted, BLS productivity and TFP, labour trackers) stay inside their normal bands; a breach of the fast bands on two adoption indicators and one adaptation indicator for four consecutive readings would count against it.

Indicators: US firms using AI (Census BTOS)Share of work hours assisted by generative AIUS labour productivity, year on yearUS total factor productivity, private nonfarm business

“Thus, we predict that slow diffusion will continue to be the norm in high-consequence tasks.” source

How we track it. FDA AI device clearances, Waymo paid rides, and safety-brake events (lab pauses, GPU reallocation) as evidence that high-consequence deployment is gated; a rapid, ungated deployment in a regulated domain would count against it.

“All this is not just true of today’s AI, but even in the face of hypothetical developments such as self-improvement in AI capabilities. Many of the limits to the power of AI systems are (and should be) external to those systems, so that they cannot be overcome simply by having AI go off and improve its own technical design.” source

How we track it. Independently verified agent-workdays per human-workday above 1 with an intervention rate on 4–8h tasks below 50% (OpenAI's self-reported 3.1 and >50% do not yet qualify) would test it on the invention stage; the diffusion indicators test it downstream.

Indicators: METR 50% time horizon50%/80% horizon ratio

“Every time we solve a benchmark (reach what we thought was the peak), we discover limitations of the benchmark (realize that we’re on a ‘false summit’) and construct a new benchmark (set our sights on what we now think is the summit). This leads to accusations of ‘moving the goalposts’, but this is what we should expect given the intrinsic challenges of benchmarking.” source

How we track it. The 50%/80% horizon ratio staying wide, the developer RCT uplift staying small while benchmark horizons race, and pilot-to-production rates staying low are the product-side evidence; a collapse of the ratio toward 2x with uplift above 40% would count against it.

Indicators: 50%/80% horizon ratioDeveloper productivity uplift (METR RCTs)Enterprise pilots with measurable P&L impact

“There is no particular capability milestone that will unlock all of this economic potential. The economic potential is already there. It really depends on all these downstream actions that we take. It is not gated by capability.” source

How we track it. Not testable as a point prediction; the tracker reads it through the concordance of adoption and adaptation indicators moving at organisational speed regardless of capability releases.

Indicators: Cross-tracker concordance (labour)Share of work hours assisted by generative AI

“According to our measurements, we have now reached the goal, announced last fall, of having an automated research intern by September of this year. By “research intern,” we mean a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days.” source

How we track it. Resolves only on an independent replication of OpenAI's agent-workday ratio and intervention rate (METR's RSI programme or equivalent), not on OpenAI's own measurements. The lab's self-grade is tier 7 and caps the status at emerging.

Indicators: METR 80% time horizonDeveloper productivity uplift (METR RCTs)

“We are making strong progress toward creating an automated AI researcher by March of 2028.” source

How we track it. Resolves when an external evaluation shows autonomous delivery of a multi-week research project without human intervention; proxied until then by the 80% METR horizon reaching weeks.

Indicators: METR 80% time horizon

“I think we might be six to 12 months away from when the model is doing most, maybe all of what software engineers do end-to-end, and then it’s a question of how fast does that loop close?” source

How we track it. 'Most' = more than half of software-engineering task-hours completed without human edits. Proxied by the 80% METR horizon on software tasks reaching working days, and by the developer RCTs turning strongly positive.

Indicators: METR 80% time horizonDeveloper productivity uplift (METR RCTs)50%/80% horizon ratio

“If these scaling laws continue for only a year or two longer, we are likely to get what I’ve called Powerful AI, or “a country of geniuses in a datacenter”.” source

How we track it. Resolves against the 'Machines of Loving Grace' definition (Nobel-level across fields, autonomous multi-day and multi-week tasks, millions of instances); proxied by the 80% METR horizon exceeding weeks and a shipped continual-learning system.

Indicators: METR 80% time horizonMETR 50% time horizon

“It is hard for me to see that there won’t be trillions of dollars in revenue before 2030.” source

How we track it. Sum of frontier-lab run-rates (Epoch's revenue reports plus press-confirmed figures) reaching $1T annualised before 1 Jan 2030; interim path per Amodei's own words on the same page, 'low hundreds of billions by 2028'.

Indicators: Circular financing scaleCloud backlog (remaining performance obligations)

“there’s this bizarre 10x per year growth in revenue that we’ve seen. So in 2023, it was zero to $100 million. In 2024, it was $100 million to $1 billion. In 2025, it was $1 billion to $ 9-10 billion.” source

How we track it. Anthropic's run-rate at end-2026 of $90–100B against $9–10B at end-2025 keeps the 10x-a-year pattern; the same page states the intent to grow '20 or 30x a year instead of 10x'.

Indicators: Circular financing scale

“According to a recent METR report, the length of coding tasks AIs can handle, their “time horizon”, doubled every 7 months from 2019 - 2024 and every 4 months from 2024-onward.” source

How we track it. The 50% METR horizon doubling time fitted over the 2024-onward window stays at or under four months (122 days); the tracker's own fit is the band input on the horizon indicators.

Indicators: METR 50% time horizonMETR 80% time horizon

“If the trend continues to speed up, by March 2027 AIs could succeed with 80% reliability on software tasks that would take a skilled human years to complete.” source

How we track it. The 80% METR horizon reaching a working year (about 2,000 hours) by March 2027; the tracker's 80% horizon indicator is the input.

Indicators: METR 80% time horizon50%/80% horizon ratio

“All model-based forecasts have 2027 as one of the most likely years that SC being developed, which is when an SC arrives in the AI 2027 scenario.” source

How we track it. SC = an AI that does any coding task of the best lab engineer, 30x faster and cheaper, on 5% of frontier compute; proxied by the 80% METR horizon and a lab disclosure of fully autonomous coding of frontier training runs.

Indicators: METR 80% time horizonMETR 50% time horizon

“AI’s $200B question is now AI’s $600B question.” source

How we track it. Capex-to-revenue gap for the stack: Nvidia data-centre run-rate doubled for the rest of the data centre and doubled again for a 50% margin, minus AI ecosystem revenue. Tracked as a gap, not pass/fail, once the capex_to_revenue_stack metric lands.

Indicators: Circular financing scaleSemiconductors' share of stack operating incomeCloud backlog (remaining performance obligations)

“A lot of people have reached out to me for the updated math behind AI’s $600B question. It is now AI’s $1.5T question.” source

How we track it. Same gap measure: about $3T of cumulative lifetime revenue required against realised AI revenue.

Indicators: Circular financing scaleSemiconductors' share of stack operating income

“Two trillion dollars in annual revenue is what’s needed to fund computing power needed to meet anticipated AI demand by 2030. However, even with AI-related savings, the world is still $800 billion short to keep pace with demand, new research by Bain & Company finds.” source

How we track it. Global AI revenue (labs, applications, cloud AI) trajectory against $2T a year by 2030; interim reading from the capex-to-revenue gap.

Indicators: Circular financing scaleCloud backlog (remaining performance obligations)

“AI technology is exceptionally expensive, and to justify those costs, the technology must be able to solve complex problems, which it isn’t designed to do,” source

How we track it. Direction claim: application-layer margins net of inference and reported enterprise P&L impact fail to cover the capex; contradicted if the lab recoupment ratio exceeds one or the developer RCTs turn strongly positive.

Indicators: Developer productivity uplift (METR RCTs)METR 50% time horizonSemiconductors' share of stack operating income

“Hyperscalers are spending $12 on AI infrastructure for every $1 of AI revenue, betting on 5x revenue growth within five years to justify $575B in capex.” source

How we track it. Hyperscaler AI capex over hyperscaler AI revenue near 12; resolves in Tunguz's favour if the ratio has not fallen toward the low single digits within five years.

Indicators: Cloud segment operating marginsCloud backlog (remaining performance obligations)Circular financing scale

“Under modest change, AI adds less than half a point to GDP growth by 2030 and raises unemployment by a tenth of a point.” source

How we track it. US GDP growth and TFP against the no-AI trend; unemployment-rate delta. The authors state the scenarios only separate after 2027, so no macro reading before 2028 discriminates them.

Indicators: US total factor productivity, private nonfarm businessUS labour productivity, year on yearUS labour share of income, year on year

“the median respondent’s answers are consistent with our substantial change scenario in which, by 2030, GDP rises by 8 percent and cognitive employment declines by 4 percent.” source

How we track it. US GDP 8% above the no-AI path by 2030; employment in cognitive occupations down 4%, read off the labour trackers' exposed-occupation series.

Indicators: US total factor productivity, private nonfarm businessCross-tracker concordance (labour)Entry-level employment shortfall in AI-exposed occupations

“GDP growth then rises to 15 percent per year, the labor share of income falls from 60 to 45 percent, and nearly one in five cognitive workers is unemployed.” source

How we track it. BLS labour share of nonfarm business income falling toward 45%; GDP growth toward 15% a year; cognitive-worker unemployment near 20%.

Indicators: US labour share of income, year on yearUS labour productivity, year on yearRecent-graduate unemployment rate (NY Fed)

“Second, almost all of the divergence comes after 2027, a year and a half from the mid-2026 anchor, because the scenarios share today’s readings and separate only as the reach and use of AI diverge.” source

How we track it. A meta-claim about the tracker's own power: no macro indicator (GDP, TFP, labour share, unemployment) should separate the three scenarios before 2028. Falsified if a macro series breaches its fast band before then.

Indicators: US total factor productivity, private nonfarm businessUS labour share of income, year on yearCross-tracker concordance (labour)