Anthropic warns of risks of self-improving AI
US-based AI company Anthropic has warned that the development of artificial intelligence is accelerating so quickly that, in the near future, AI agents could independently create, train and improve new systems without human involvement. Against this backdrop, the company has called for a slowdown in the development of neural networks.
In a blog post, Marina Favaro, head of the Anthropic Institute, and Anthropic co-founder Jack Clark said AI agents are already capable of running code on their own, delegating hours of work to other agents and gradually approaching the point where they could take over the entire development cycle.
“For most of AI’s history, humans controlled every stage of its development. But at Anthropic, we are delegating an increasing share of AI development to AI systems themselves, which is accelerating our work,” they said.
According to Favaro and Clark, if this trend continues and systems receive enough computing power, AI could eventually emerge that is capable of fully autonomously designing and developing its own successor.
Focus on safety
The growth of such capabilities is raising concerns about a scenario in which AI could become smarter on its own. In December, OpenAI said it was studying safe ways to develop and deploy increasingly advanced AI systems, including models capable of recursive self-improvement.
OpenAI said it wants such systems to consistently follow human intent in complex real-world scenarios and adversarial conditions, avoid catastrophic behavior, remain controllable and auditable, and stay aligned with human values.
They added that once the quality of human-written and AI-written code reaches parity, people may stop writing code altogether and shift only to reviewing it. However, if they cannot review code as quickly as Claude generates it, human review will become the main bottleneck in AI development.
Why slow down AI development
Favaro and Clark also said the ideal scenario would be to slow development so society has more time to understand the “immense” consequences of this technology.
In April, Anthropic refused to release its AI model Claude Mythos to the public due to concerns related to global cybersecurity threats. The model could easily create software exploits, so the company decided not to make it public for now.
On the same day, a group of technology leaders, including representatives from Anthropic and OpenAI, published an open letter urging lawmakers to introduce stricter limits on AI. The authors warned that the technology could help malicious actors overcome the “knowledge barriers” that had previously prevented them from creating biological weapons.
At the same time, they stressed that simply slowing down could create new risks if it allowed less cautious players to catch up technologically with industry leaders. Without a global coordination mechanism, companies and governments will be forced to make difficult safety decisions under competitive and geopolitical pressure.
What makes Claude valuable
Anthropic’s flagship product is the AI assistant Claude. It is a family of language models designed to work with text, code, data and complex tasks where not only response speed matters, but also safety, controllability and reliability. Claude is used as a universal assistant: it can write and edit texts, analyze documents, explain complex topics, help with programming, find errors in code, and prepare summaries, emails, reports and other work materials.
For businesses, Claude is valuable because it can take over part of routine intellectual work. It is used in customer support, analytics, software development, legal and financial processes, marketing and internal corporate tools. Companies use Claude as a chat assistant, an API for integration into their products and a tool for automating tasks that require quickly processing large amounts of information and receiving clear answers in natural language.
As a reminder, Claude helped recover the password to a wallet containing 5 Bitcoin.
Latest Anthropic News
- Forex
- Crypto