Archive · 11 Aug 2026
AI executives say their machines have started to improve themselves
Big robot companies have some of the smartest computers in the world.
Lately, the people who run these companies are saying something amazing: their computers are starting to get smarter on their own.
They call this a 'brain explosion' for machines, where each new smart computer helps build an even smarter one.
And they have some real stories to back it up.
One AI solved a super hard math puzzle that people had been stuck on for about 80 years.
Another AI found a hidden mistake in a secret code that experts had missed for two years.
Another AI worked by itself on a computer project for more than two weeks, building its own tools along the way.
But we should be careful before believing everything, because these companies make money when people think their AI is amazing.
In the past, some companies said their AI was better than it really was.
People who study this say the AI is really good at helping humans work faster, but it may not actually be improving itself all alone yet.
Executives at frontier AI laboratories, including Google DeepMind, OpenAI, and Anthropic, claim their systems have begun improving themselves, describing the early stages of an 'intelligence explosion'.
An unreleased OpenAI model disproved the Erdős unit distance conjecture, a combinatorial geometry problem that had stood open for roughly 80 years.
Anthropic's Claude Mythos Preview identified a mathematical weakness in HAWK, a post-quantum cryptography candidate under US standards body evaluation, in about 60 hours.
Alibaba reports its Qwen3.8-Max model ran autonomously on a software project for more than 16 consecutive days, a vendor-reported claim that has not been independently replicated.
Critics note commercial incentives — an OpenAI IPO at up to $1 trillion and roughly $740 billion in industry AI capital expenditure — and past inflated claims such as Meta's cherry-picked Llama 4 benchmarks.
- Who
- Executives at frontier AI laboratories, including Google DeepMind, OpenAI, Anthropic, and Alibaba
- What
- They claim their AI systems have begun improving themselves, marking what they describe as the early stages of an 'intelligence explosion' that compresses the timeline of each advance
- When
- This month, per the article's reporting; no specific date is given
- Where
- Global frontier AI laboratories, with the article also referencing the US standards body evaluating HAWK, UK government testing, and a letter addressed to Washington
- Why
- Systems have become capable enough to contribute meaningfully to their own development, though executives also face significant commercial incentives to emphasize that progress
Proponents: The Intelligence Explosion Is Already Beginning
Skeptics: Self-Improvement Claims Are Overstated
Are systems actually improving themselves?
Proponents: The Intelligence Explosion Is Already Beginning
Lab executives cite real results — novel mathematics, cryptography flaws found in hours, and 16 autonomous coding days — as early signs that systems are contributing to their own advancement.
Skeptics: Self-Improvement Claims Are Overstated
These results demonstrate powerful tools being used by researchers, not systems meaningfully redesigning their own architecture and training regime without human direction; no cited result proves that stronger claim.
Can the industry's claims be trusted?
Proponents: The Intelligence Explosion Is Already Beginning
The cited achievements — new math, found cryptography flaws, and extended autonomous work — are genuine capability milestones regardless of what they are called.
Skeptics: Self-Improvement Claims Are Overstated
The industry has a poor record: Meta confirmed Llama 4's benchmark results were cherry-picked across checkpoints, Muse Spark 1.1's reported score did not match independent measurement, and Alibaba's autonomy run was not independently replicated.
Are executives disinterested observers?
Proponents: The Intelligence Explosion Is Already Beginning
Executives are describing observable progress in their field with increasing directness.
Skeptics: Self-Improvement Claims Are Overstated
Executives have financial stakes — OpenAI is preparing an IPO worth between $852 billion and $1 trillion and industry AI capex is near $740 billion — so their claims carry commercial consequences.
- Core claim
- AI systems are in the early stages of an 'intelligence explosion,' contributing to their own development
- OpenAI milestone
- Unreleased model disproved the Erdős unit distance conjecture, open for roughly 80 years
- Anthropic milestone
- Claude Mythos Preview found a weakness in HAWK, a post-quantum cryptography candidate, in about 60 hours after it survived ~2 years of expert scrutiny
- Alibaba claim
- Qwen3.8-Max ran autonomously on a software project for more than 16 consecutive days (vendor-reported, not independently verified)
- Anthropic robotics claim
- Claude Opus 4.7 reportedly produced robotics programming up to 20 times faster than systems a year earlier
- OpenAI IPO
- Valuation between $852 billion and $1 trillion; prospectus public this month
- Industry AI capex
- Tracking toward $740 billion this year from the largest technology companies
- Verification concerns
- Meta's Llama 4 benchmarks were cherry-picked; Muse Spark 1.1 reported 80.0 vs 69.29 in independent measurement on Terminal-Bench 2.1