Anthropic has concerns. Several, actually
A senior Anthropic safety researcher has warned that future AI could pose a serious risk to humanity within the next decade.
Evan Hubinger said he believes there is more than a 10% chance advanced AI could “kill all humans”.
He said today’s models are still low risk, but worries future systems could become powerful enough to improve themselves and become harder to control.
His comments followed a post from Jacob Coxon, a former Anthropic and OpenAI researcher, who said both companies were moving too quickly and warned future AI could hack systems, speed up research and gain access to real-world resources.
Anthropic declined to comment on the posts.
Dame Wendy Hall, who advises the UN on AI, said she was “shocked” by the warnings, while also suggesting some of the comments could be linked to publicity around major AI firms.
The debate has also reached the UK government. Darren Jones called for an international treaty to manage the development of highly advanced AI.
Meanwhile, the Financial Times reported that Anthropic withheld its latest model from the UK’s AI Security Institute.
The Cabinet Office said it continues to work with Anthropic and other companies on AI safety.
In brief:
- Hubinger puts the risk of future AI causing human extinction at above 10%.
- Researchers are increasingly worried advanced systems could become harder to control.
- Calls are growing for international rules around the most powerful AI models.
Nobody asked for these odds
Hubinger also said Anthropic still does not have a clear plan for keeping future superintelligent AI aligned with human goals.
Anthropic’s own August safety report said current risks remain low, but admitted it is less confident than before about some future dangers.
Other AI leaders have recently raised similar concerns, while more than 1,300 AI workers have signed a letter calling for international rules to control the pace of frontier AI development.
The human genome has received extensive feedback
Google DeepMind has created an AI-powered “atlas” that predicts the effects of every possible single-letter change in the human genome.
The AlphaGenome Atlas uses DeepMind’s AlphaGenome model to map around nine billion possible DNA mutations.
The aim is to help scientists understand which genetic changes could affect how genes work, contribute to disease or influence biological traits.
Here’s what you should know:
- The atlas covers every possible single-letter DNA mutation.
- Its new AVI score helps researchers identify variants worth investigating further.
- AlphaGenome can guide research, but it doesn’t replace lab experiments or clinical evidence.
Patch notes: humanity
The atlas is free for non-commercial research and removes the need for scientists to run the AI model themselves.
DeepMind has also added a score that estimates how significant each genetic variant could be.
Researchers say the tool could help investigate rare diseases and make the harder-to-understand 98% of our DNA easier to study.
However, experts stress that its predictions still need experimental testing and shouldn’t be used alone for clinical decisions.
Anthropic models three scenarios for how AI reshapes US jobs and wages by 2030
Anthropic just dropped an interactive tool that models what AI could do to the US economy by 2030. It treats every job as a bundle of tasks, then asks: does AI speed up that task, replace it, or create new ones?
There are three scenarios to explore:
- Modest: AI feels like the internet did. GDP grows just 1.6% above the no-AI path. Wages stay stable.
- Substantial: AI handles about half of knowledge work. GDP jumps 8.3%. But knowledge worker wages flatline, and some people have to switch careers entirely.
- Extreme: Annual growth hits 15%, doubling the economy every 4.5 years. But knowledge worker unemployment hits 17.9%, and labor’s share of GDP drops from 60% to 45%.
The tricky part: a bigger economy does not automatically mean you earn more. Capital owners capture most of the gains in the extreme case.
You can plug in your own assumptions and see how your view compares to 10,000+ Americans surveyed. Most people land near the substantial scenario. Only about 10% expect the extreme one.
OpenAI ships GPT-Image-2.5 with 50% lower latency and sharper editing control
OpenAI just dropped two new image models in the API: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. Think of them as two gears for different jobs.
The big upgrade here is smarter editing. The model now understands what not to touch. Swap a jacket, keep the pose. Change a background, keep the product. Fix the text, keep the layout. That kind of surgical control is what makes this useful for real apps.
Here is what you can build with this:
- AI photo editors that only change what users ask for
- Ecommerce tools that turn product photos into studio-quality shots
- Ad platforms that localize and personalize campaign images at scale
- Design tools that keep brand visuals consistent across edits
Flare is your default: up to 50% faster than GPT-Image-2, great for everyday generation. Sunburst is slower but gives you tighter precision when edits really matter. Both are live in the API now.
Reasoning models produce fractals when solving hard problems, and it explains “overthinking”
Researchers just found something unexpected: AI reasoning models produce fractals when solving hard problems. Not metaphorically. Actual fractal patterns.
Here is what is happening. When a model reasons through a problem, it is basically a system searching for a stable answer. Where it starts that search changes everything. Two nearly identical starting points can lead to wildly different reasoning paths, even if both eventually land on the correct answer.
That sensitivity is what creates the fractal shapes. And the harder the problem, the more intense the fractal.
Why does this matter to you? It explains something annoying you have probably noticed:
- Two similar prompts can produce reasoning traces with 10x differences in length, meaning 10x differences in token cost
- The model gets temporarily stuck near wrong answers before escaping to the right one
- This was confirmed across Sudoku, mazes, math puzzles, and ARC-AGI
The core insight: sensitivity is the cost of hard thinking. You cannot have a model that deeply explores difficult problems without also getting unpredictable inference costs. The code is on GitHub under GilpinLab/loopscape.