16 highlights from the State of AI Report 2025
Math progress, paid AI adoption trends, impact on elections, and more
The annual State of AI report is out, and I’ve spent about six hours going through the 313 slides and taking extensive notes along the way.
I spend a lot of time following AI progress on a day-to-day basis, but there was still a good deal of new and interesting information for me.
Here I have collected 16 key pieces of information that I think could be useful to many people. I’ll include the corresponding slides for each if you want more context.
1) OpenAI models remain at the frontier of intelligence, but the gap has narrowed
“A fast-moving open-weights pack from China (DeepSeek, Qwen, Kimi) and closed-source group in the US (Gemini, Claude, Grok) sits within a few points on reasoning/coding.”
2) Researchers from OpenAI, Google, Anthropic, and more sounded alarm over possible loss of ability to understand AI
The matter is the so-called Chain-of-Thought (CoT), which shows how models think, and therefore for instance can surface explicit malicious intents.
But, they are incomplete “and can drift away from faithful reasoning while advanced, situationally-aware models might hide or compress their thoughts.”
3) Substantial progress in math
OpenAI, Google DeepMind and the startup Harmonic hit gold-medal performance in The International Mathematical Olympiad (IMO).
The improvements in the field “point to the very real possibility of a non-trivial research-level result in mathematics being proven and formalized by an AI system (with some human supervision involved) within the next year.”
4) Chess grandmasters learn new concepts from AI
“Researchers extracted novel chess concepts from AlphaZero (an AI system that mastered chess via self-play without human supervision) and successfully taught them to 4 world champion grandmasters, demonstrating that superhuman AI systems can advance human knowledge at the highest expert levels.”
5) China leads the pack in computer-use agents
OpenAI and Anthropic have both released tools for letting the AI actually do tasks on a computer, but the new UI-TARS-2 model from China’s ByteDance has blown the competition out of the water with new high-scores across benchmarks.
However, long-horizon problems like browsing for hard-to-find information ”remain brittle”.
6) Intense AI research activity - China dominates
The 2026 conference for the Association for the Advancement of Artificial Intelligence (AAAI) has received an unprecedented 29,000 research submissions, almost double that of 2025, and more than two-thirds (20,000) come from China.
The venue for the event is in Singapore, though, which could have led to extra proposals from the region (as pointed out by a Reddit user).
7) Pattern: from model release to fundraising
There is a predictable cadence between releases of new frontier models from the leading labs and their fundraising events.
For instance, on average, Anthropic announces a fundraising round 44 days after they’ve released a new model. For OpenAI and xAI, it’s 50 and 77 days, respectively.
8) Card data shows paid AI adoption surges more than government estimates
Based on card/bill-pay data from 45,000+ US businesses using the fintech Ramp’s platform, paid AI adoption reached 43.8% in September 2025, up from 5.2% in January 2023 (OpenAI introduced a paid tier for ChatGPT in February that year).
In contrast, the US Government based on its Business Trends and Outlook Survey (BTOS) estimates the number to be 9.2%.
There could be a higher share of early adopters among Ramp’s customers, although that does not explain the recent surge since the beginning of 2025.
9) AI search engines surge - Google is losing terrain
ChatGPT just reached 800 million weekly users, and it could seem that some are using it as a substitute for the classic Google Search.
”SEM Rush data shows Google’s global search traffic fell ~7.9% year-over-year, the first significant dip in decades, even as it retained ~90% global share.“
10) AI search is high-intent
Retail visits referred by ChatGPT now converts better than every other major marketing channel with 11.4% conversion, up from ~6% a year ago. For two comparable numbers, Paid Search and Email are at 9.3% and 4.6%, respectively.
11) US dominates computing power, leading to a self-reinforcing cycle
The US controls ~75% of global supercomputer capacity with 850,000 equivalents of the H100 GPU from NVIDIA, compared to China’s 110,000 (~14%). EU is at <5%.
The American lead “creates a self-reinforcing cycle where compute advantage drives breakthroughs that attract more investment.”
12) European brain drain for top AI researchers
“22% of the world’s leading AI researchers studied in Europe, but only 14% continue to work in the EU.”
13) Elections - little to no negative impact from generative AI
“Despite rampant worries of AI-generated election dis/misinformation during the “largest election year in global history,” there was almost little to no negative impact from GenAI in any of the 2024 elections.”
14) AI helped parties reach voters across a breadth of languages
”In India, political parties spent $50M on legal AIGen, using it for voter outreach via AI voice clone calls, personalized videos, and translating speeches into one of 22 official and 780 unofficial languages.”
(Same slide as above)
15) AI labs spend more in a day than 11 safety organisations do in a year
11 of the most prominent American AI safety-science organisations is estimated to spend $133.4 million in 2025. That compares to the $92 billion that the AI labs will spend, which is the equivalent of ~$252 million per day.
16) Lowered barrier for sophisticated cybercrime
The improved technical capabilities of the leading models come with a downside.
”North Korean operators with minimal technical skills leveraged Claude to pass technical interviews at Fortune 500 tech companies and maintain engineering positions. These salaries directly fund North Korea’s government and military programs.”
Obviously there are way more insights in the full report, which you can read for free here.
And if you found this useful, consider sending it to a friend or colleague.