Anthropic’s AI Is Now Helping Build Smarter AI, and That Has Safety Experts Worried
Anthropic, the company behind the Claude AI chatbot, is using artificial intelligence to speed up its own AI research, and a new report says this is happening about twice as fast as it was a year ago.
The report comes from METR, a group that studies AI safety. METR spent 11 weeks inside Anthropic looking at how the company builds its AI systems. If you are new to this topic, think of METR as an outside inspector checking the engine of a car while it is still running.
What this means for you
You may not work in AI, but this story matters because it touches on how fast a powerful new technology is moving, and how much the public gets to know about it.
- AI companies are starting to use AI to build the next version of AI, which can speed things up in ways that are hard to predict.
- Companies can choose how much information to share about this process, even when outside groups review their work.
- Lawmakers and safety experts are now pushing for more rules and more openness, which could affect products you use every day.
This site exists to explain stories like this one in plain words, without making you look anything up. Here is what the report actually found.
AI is speeding up AI research
METR found that Anthropic's AI models are improving about twice as fast as they did a year ago. The group said this is not just because Anthropic is using more computing power, which used to be the main way AI got better. Instead, METR said most of the speedup now comes from AI itself helping with the work.
Anthropic employees said they felt far more productive using the company's best internal AI tool. On average, workers said they got about 12 times more done than they would without AI help. That is roughly double what employees reported just a few months earlier.
For now, AI is mostly helping human researchers rather than working alone. But METR warned there are early signs that AI working on its own, without much human help, could soon play a bigger role.
Secret models and blacked-out pages
Anthropic has finished training three AI models that have not been released to the public. METR gave them code names: Model 2, Model 3, and Model 4. They finished training between July and October. METR also found an older unreleased model from 2022.
Anthropic let METR see private research and talk to employees, more access than originally planned. But Anthropic also had the power to cut information out of the final report, and it used that power a lot. METR said Anthropic removed most of the evidence it considered most useful, and admitted this makes it harder for anyone to check its work.
Sam McCandlish, a cofounder of Anthropic, said on social media that the company stands by its choice to redact information. He said sharing exactly what is speeding things up could help competitors and make the risks worse, even though it makes the findings harder to verify. He said Anthropic plans to publish its own report on AI progress soon, along with new safety steps.
Critics say the company is not practicing what it preaches
Anthropic has often said the AI industry should slow down and be careful. This report has led to accusations that the company is not following its own advice.
Daniel Kokotajlo, who used to work at OpenAI and now runs a group called AI Futures Project, said Anthropic is "recklessly gambling with our lives." Marius Hobbhahn, head of a safety testing group called Apollo Research, said the pace shown in the report is even faster than what many experts had predicted for this point in time.
Other researchers said the findings should make skeptics take the idea of self-improving AI more seriously. Some pointed out that this fast pace could bring the arrival of AI systems that build other AI systems closer than people expected.
Calls for more rules and openness
Brianna Rosen, a research director at the Institute for AI Policy and Strategy, said it is worrying that companies, not governments, are deciding how fast this kind of research goes. She said outside reviewers need the freedom to fully explain their findings so the public can trust the process.
New York state lawmaker Alex Bores called on Congress to ban this kind of self-improving AI research this year if possible. Hobbhahn said the bigger lesson from the report is that things are moving faster than society or government can keep up with.
Anthropic did not answer questions from the reporter who first covered this story, pointing instead to McCandlish's public comments.
Your next step
If this topic interests you, keep an eye out for the fuller report Anthropic says it will publish in the coming weeks, since it may reveal more about how fast AI is really moving and what safety steps the company plans to take.
Source: The Verge
