AI wrote 75% of the code, but it didn’t make itself move faster. Google just poured cold water on AI self-improvement 🦖
🕐 最新解读群里更新
The investigators: Google DeepMind, one of the world’s largest AI labs. What they investigated: the hypothesis of recursive self-improvement—the idea that AI creates a smarter version of itself, which then creates an even smarter next generation, so quickly that humans can’t test, regulate, or simply shut it down in time. The conclusion, in the latest flagship model safety report published on October 14: no evidence of acceleration.
First, three numbers. One: more than 75% of production code at Google DeepMind is now written by AI. The share of code submissions handled entirely by AI rose from 7.5% in April to 11.9% in September. AI is clearly becoming more deeply embedded. Two: around 88% of code changes still need multiple rounds of human guidance—or are written by people themselves. Three: the company’s internal composite capability index, which combines 123 models and 92 benchmarks into a single score, has risen by about 0.1 points per month. There was no statistically significant acceleration after April 1, 2026, either.
The shape of the curve says even more. The flagship generation at the end of 2024 scored around 0.5 points. The previous generation, from this September, scored just over 2.9. Almost every version lies along a straight line. The latest model, Argon, scored around 3.3—the only point that sticks out. But its margin of error is wide enough to overlap with the trend line. Researcher Haydn Belfield, who took part in the evaluation, explained that most of this outperformance comes from increased computing power and organizational investment, not AI accelerating itself.
The data on actual work is even more direct. Google examined 1,389 work logs from Argon acting as an internal agent. The conclusion: the bigger the task, the less likely it is to complete it independently, and the more often a human needs to step in. It managed only about half of internal research engineering tasks, while DeepMind considers 90% the threshold for full automation. Another survey of 12 senior leaders and researchers found that automation is highest for clearly defined, verifiable tasks, and lowest for open-ended scientific creativity, designing new experiments, diagnosing unstable training, and long-term research judgment.
Some context: a month ago, Anthropic CEO Dario Amodei publicly said recursive self-improvement was already accelerating. Sam Altman and Elon Musk echoed him. Those remarks sent chip stocks down nearly 6% that day, while Bitcoin was rising ⚠️
My take: this audit isn’t declaring that AI has hit a wall. It is pulling the runaway-acceleration hypothesis down from market expectations. For crypto markets, some assets with an AI narrative premium are priced partly on the assumption that AI will keep getting faster on its own. Now the company best placed to build AI says its internal data shows otherwise. In the short term, that’s clearly a sentiment headwind. But the 75% and 11.9% figures also show that AI is already deeply embedded in production workflows. The underlying demand for payments and settlement in the agent economy hasn’t been overturned 📉
A quick look at the markets: Bitcoin is fluctuating around $82,000, and Ethereum is hovering below $2,500. The market-implied odds of a Fed rate hike in October have risen from 18% to 29% over the past week. In other words, as the AI narrative cools, macroeconomic tensions are tightening.
Do you think AI will lose control of itself first, or will data like this gradually disprove the idea? If fears of another AI acceleration surge return, will Bitcoin once again become a safe haven for capital? Share your thoughts in the comments.
Tap the profile picture to watch the livestream.
Every day, I bring you the latest in AI and crypto—not just what’s happening, but the logic and opportunities behind it 👀🚀
🕐 最新解读群里更新
The investigators: Google DeepMind, one of the world’s largest AI labs. What they investigated: the hypothesis of recursive self-improvement—the idea that AI creates a smarter version of itself, which then creates an even smarter next generation, so quickly that humans can’t test, regulate, or simply shut it down in time. The conclusion, in the latest flagship model safety report published on October 14: no evidence of acceleration.
First, three numbers. One: more than 75% of production code at Google DeepMind is now written by AI. The share of code submissions handled entirely by AI rose from 7.5% in April to 11.9% in September. AI is clearly becoming more deeply embedded. Two: around 88% of code changes still need multiple rounds of human guidance—or are written by people themselves. Three: the company’s internal composite capability index, which combines 123 models and 92 benchmarks into a single score, has risen by about 0.1 points per month. There was no statistically significant acceleration after April 1, 2026, either.
The shape of the curve says even more. The flagship generation at the end of 2024 scored around 0.5 points. The previous generation, from this September, scored just over 2.9. Almost every version lies along a straight line. The latest model, Argon, scored around 3.3—the only point that sticks out. But its margin of error is wide enough to overlap with the trend line. Researcher Haydn Belfield, who took part in the evaluation, explained that most of this outperformance comes from increased computing power and organizational investment, not AI accelerating itself.
The data on actual work is even more direct. Google examined 1,389 work logs from Argon acting as an internal agent. The conclusion: the bigger the task, the less likely it is to complete it independently, and the more often a human needs to step in. It managed only about half of internal research engineering tasks, while DeepMind considers 90% the threshold for full automation. Another survey of 12 senior leaders and researchers found that automation is highest for clearly defined, verifiable tasks, and lowest for open-ended scientific creativity, designing new experiments, diagnosing unstable training, and long-term research judgment.
Some context: a month ago, Anthropic CEO Dario Amodei publicly said recursive self-improvement was already accelerating. Sam Altman and Elon Musk echoed him. Those remarks sent chip stocks down nearly 6% that day, while Bitcoin was rising ⚠️
My take: this audit isn’t declaring that AI has hit a wall. It is pulling the runaway-acceleration hypothesis down from market expectations. For crypto markets, some assets with an AI narrative premium are priced partly on the assumption that AI will keep getting faster on its own. Now the company best placed to build AI says its internal data shows otherwise. In the short term, that’s clearly a sentiment headwind. But the 75% and 11.9% figures also show that AI is already deeply embedded in production workflows. The underlying demand for payments and settlement in the agent economy hasn’t been overturned 📉
A quick look at the markets: Bitcoin is fluctuating around $82,000, and Ethereum is hovering below $2,500. The market-implied odds of a Fed rate hike in October have risen from 18% to 29% over the past week. In other words, as the AI narrative cools, macroeconomic tensions are tightening.
Do you think AI will lose control of itself first, or will data like this gradually disprove the idea? If fears of another AI acceleration surge return, will Bitcoin once again become a safe haven for capital? Share your thoughts in the comments.
Tap the profile picture to watch the livestream.
Every day, I bring you the latest in AI and crypto—not just what’s happening, but the logic and opportunities behind it 👀🚀