Anthropic has launched its first update to the system that powers its Claude chatbot – shortly after telling the world to slow down AI development.
The new model, Opus 5.5, is the first model released since Anthropic’s chief executive Dario Amodei backed calls to “pace the frontier”, or slow down AI development to give the world time to ensure systems are safe.
Anthropic acknowledged that concern but said that testing had showed that the system was better behaved than any other models it had tested. It includes a range of safeguards it had developed for its most powerful models, it said.
The new model is touted as being better in performance, so that it is now Anthropic’s leading model. But it is also more efficient than the Opus 5 model it replaces, the company said, with a 40 per cent reduction in costs.
After the launch was announced ,Sholto Douglas, a member of Anthropic’s technical team, tweeted “also important news we fixed the writing”. In recent months, Claude has been repeatedly criticised for leaning on a very particular and frustrating style, such that the quality of its writing decreased and it was especially identifiable as being produced by the system.
“Early testers found its writing clearer and easier to follow, which addresses some of the common feedback we heard about Opus 5,” Anthropic said in its announcement. “It puts the most important information up front, and its style makes it a better work partner over long sessions.”
But Anthropic was particularly keen to focus on the safety improvements in the new model. It said that it had tested it across thousands of simulated scenarios and found that it performed better than any other model, and that it was less likely to take “hard-to-reverse actions or act outside the boundaries it’s been given”.
The new system’s performance is similar to that of Mythos, a model that has been deemed so powerful at dangerous tasks such as biological research and cyber security that it has not been released to the public, Anthropic said. As such, it will come with safeguards that mean it may refuse to act on dangerous requests, and researchers in cyber security and biology will be required to apply for special permission to use it in their work.
Those safety concerns come amid renewed worry about the danger posed by advanced models. Mr Amodei’s commitment to “pacing the frontier” came after an Anthropic engineer quit the company and warned that the people building AI technology “earnestly believe that it could kill us all by the end of the decade”.
“Do not underestimate the power of this technology,” warned Jacob Coxon, in a series of highly publicised tweets. “These will soon be superhuman systems that can hack anything, revolutionise any field overnight, and acquire real power and resources.”
Those posts were echoed by researchers still working at Anthropic, and led to a new focus on what might be done to protect the world from the potentially disastrous consequences of increasingly developed artificial intelligence systems.
Read MorePrince Harry forced to halt AI speech after technical glitch
Meta’s AI agent Muse just overtook ChatGPT
Why is everyone talking about Jev? The AI tool ‘200 times’ faster than ChatGPT
