For eleven months Anthropic's blocking classifiers for chemical and biological weapons never ran on its human-feedback platforms -- about 50,000 vendor-vetted contractors, some 133 million exchanges, with the logging disabled by the same flag -- and it found out only on an outside tip.
Documents shown to IPO investors put Anthropic's preliminary Q2 revenue above $11.5bn against $787m a year earlier, with positive adjusted operating income -- the first hard profitability figure the company has let out, and it is preliminary.
Nvidia has cut its proposed guarantee for OpenAI's Ohio campus from $250bn to under $120bn after investors questioned its risk exposure -- the second time the same partnership has been trimmed, and the first retreat in the circular financing this page has tracked since July.
A State Department draft reviewed by Reuters warns Pax Silica signatories -- Japan, Australia, South Korea and Kazakhstan among them -- that joining China's rival WAICO framework costs them the US-led AI supply chain.
Models Fail Half of Real Finance Work
The Decoder: no frontier model tops 59.7% on a new vision-perception benchmark
