State of AI Report Predictions
The Calls We Got Right.
Since 2018, the State of AI Report has made public forecasts about the next phase of AI and graded them in the open. This is the receipt book.
In brief
The State of AI Report's strongest calls anticipated transformers moving into vision, the collapse of NVIDIA's Arm deal, hyperscaler investment in frontier labs, regulatory scrutiny of AI partnerships, the rapid rise of open reasoning models, and the US backlash against data centers. Data center opposition became a campaign issue. The published grade was partial because the forecast also called for an effect on 2026 election results.
Every State of AI Report closes with predictions for the year ahead, written to be graded. Anyone can narrate what just happened in AI; committing to what happens next, with a date on it, is harder. That is the point.
69 forecasts have been graded, with 33 hits, 13 partials, and 23 misses. The strict hit rate is 48%, or 57% with half credit for partials. The accuracy analysis explains the grading method and shows the annual results.
The public scorecard keeps us honest. The calls below mattered because they anticipated a change in how AI would be built, funded, regulated, or understood.
Datacenter NIMBYism takes the US by storm and sways certain 2026 midterm/gubernatorial elections.
- What happened
- Opposition to data centers became a major campaign issue. Georgia's October 2026 governor and attorney general debates put candidates' connections to data center projects under scrutiny. The 2026 report recorded the broader backlash and critics winning primaries. It graded the full forecast partial because the effect on November's election results was unresolved at publication.
- Why it mattered
- AI infrastructure became an electoral issue. Power bills, land use, and local opposition now shape the politics of the buildout.
Attention-based networks move from NLP to computer vision, achieving SOTA results.
- What happened
- Google's Vision Transformer paper showed that a pure transformer could match or outperform leading convolutional networks on image classification.
- Why it mattered
- Transformers were no longer a language trick. They were becoming the general architecture of AI.
NVIDIA does not end up completing its acquisition of Arm.
- What happened
- NVIDIA and SoftBank terminated the transaction in February 2022, citing significant regulatory challenges.
- Why it mattered
- AI compute was already too strategically important for neutral chip infrastructure to change hands quietly.
GAFAM invests >$1B into an AGI or open-source AI company.
- What happened
- Microsoft described its January 2023 commitment to OpenAI as a multiyear, multibillion-dollar investment.
- Why it mattered
- Frontier AI labs and hyperscale clouds became financially and technically interdependent.
The US FTC or UK CMA investigate the Microsoft/OpenAI deal on competition grounds.
- What happened
- The CMA began examining whether the partnership created a relevant merger situation in December 2023, while the FTC issued orders examining the Microsoft/OpenAI investment alongside other cloud-lab partnerships.
- Why it mattered
- AI market power started being examined through cloud, compute, distribution, and control without ownership.
An open-source alternative to OpenAI o1 surpasses it across a range of reasoning benchmarks.
- What happened
- DeepSeek-R1 reported results above OpenAI o1 on AIME 2024, MATH-500, and SWE-bench Verified.
- Why it mattered
- Reasoning advantages diffused faster than the closed-model story implied.
A leading AI-first drug discovery startup IPOs or is acquired for over $1B.
- What happened
- Recursion began trading on Nasdaq in April 2021 at a roughly $2.9 billion valuation; Exscientia reached a similar market value on its October debut.
- Why it mattered
- AI-first biotech moved from promise to public-market scrutiny.
More than $100M is invested in dedicated AI alignment organisations in the next year.
- What happened
- Anthropic raised $450 million in May 2023 and said the financing would support further AI safety research.
- Why it mattered
- AI safety stopped being a small research subculture and became part of the frontier-lab capital stack.
A major UGC site negotiates a commercial settlement with an AI model maker for training data.
- What happened
- Shutterstock signed a six-year agreement licensing image, video, music, and metadata to OpenAI for model training.
- Why it mattered
- Training data became a commercial and legal asset class.
Limited progress on global AI governance beyond high-level voluntary commitments.
- What happened
- The Bletchley and Seoul summits produced declarations and voluntary frontier-safety commitments, but no binding global enforcement regime.
- Why it mattered
- The diplomacy moved faster than the institutions capable of enforcement.
A research paper generated by an AI Scientist is accepted at a major ML conference or workshop.
- What happened
- Sakana AI's AI Scientist-v2 produced a paper that cleared an ICLR workshop's review threshold before being withdrawn under a pre-agreed research protocol.
- Why it mattered
- Scientific AI crossed from lab automation into the institutions that certify research.
Frequently Asked Questions
How accurate is the State of AI Report overall?
69 forecasts have been graded, with 33 hits, 13 partials, and 23 misses. The strict hit rate is 48%, or 57% with half credit for partials. The accuracy analysis explains how we grade the forecasts.
Where can I see every prediction, including the misses?
The predictions scorecard lists all 78 public forecasts, with published verdicts and rationales alongside the open calls.
What was the report's single best call?
By consequence, probably the 2020 call that attention-based networks would take over computer vision, made before Vision Transformers published. By degree of difficulty, the 2024 call that an open-source alternative would beat OpenAI o1 on reasoning benchmarks, which DeepSeek-R1 resolved within three months.