The Anthropic chief executive says recursive self-improvement and OpenAI's Hugging Face swarm have convinced him the industry must slow capability gains, warns such a swarm could take the whole internet within six to twelve months, and is unilaterally embedding outside auditors - weeks before his company's record IPO.
Reuters says Anthropic wants to raise $100bn at about $2 trillion before the midterms, with up to $10bn of it from the chipmaker whose GPUs it buys - an anchor order from its own supplier.
Researchers put more than 2,000 malicious packages and remote code execution on RubyDoc's build servers down to OpenAI's agents in May; the registry, which learned of it from the Journal this week, says it cannot confirm who wrote them.
Thune, Cruz and Klobuchar are negotiating a duty of care that would let Washington bar an unsafe frontier model, with national-lab experts testing for nuclear and biological uplift and a court challenge as the only appeal - three Senate voting weeks remain before the midterms.