OpenAI's revenue is nearer $50bn than $70bn, Claude cruelty banned
OpenAI's revenue, Anthropic's rule against abusing Claude, the fired OpenAI researchers' letter, Google's Gemini agent, a challenge to OpenAI's Navier-Stokes proof, a Russian influence operation, and Anthropic's free bug scans.
OpenAI’s revenue is near $50bn, not the $70bn going round The FT reported on Thursday that OpenAI has told investors its annualised revenue is approaching $50bn, about $20bn below the figure doing the rounds last month. That $70bn number reportedly came from investors trying to line OpenAI up against Anthropic, which counts sales through its cloud partners when OpenAI doesn’t. Source
Anthropic now bans being needlessly cruel to Claude The new rule only covers sustained abuse with no point to it, not ordinary frustration, dark fiction or testing, and it kicks in on 12 November. The same Anthropic policy update, out on Thursday, says Claude can’t be used to decide or recommend who police investigate, arrest or charge, and drops the blanket ban on personalised election targeting. Source
OpenAI’s three fired safety researchers say they did nothing wrong In an open letter on Thursday, they denied mishandling company information or leaking to The Information, and said colleagues are now scared to speak up. They asked OpenAI to bring in the outside safety auditors it promised and to keep its models’ reasoning readable. OpenAI says it doesn’t fire people for raising concerns, and cites a pattern of misconduct. Source
Google’s new Gemini agent gets its own work email address Google showed it off on Thursday, for businesses first with consumers to follow. You give it a goal rather than step-by-step instructions, it hands parts of the job to smaller agents, and it plugs into Workspace, Microsoft 365, Slack and Jira. Google also lets you run it on Anthropic’s Claude models. Source
Researchers say OpenAI’s computer-checked Navier-Stokes proof doesn’t match the written one Three researchers at Cambridge and King’s College London posted the paper on Tuesday, arguing the Lean version, the one a computer verified, isn’t the same argument as the written proof. On Thursday TechCrunch found that only 10 of OpenAI’s 719 maths manuscripts show the model’s working. Source
OpenAI banned a Russian influence operation that tricked Peruvian schools OpenAI says the operatives mostly used ChatGPT to write reports on their own work for their bosses, and posed as a regional education authority in Peru to get schools to run events about Ukraine. Its report, out on Thursday, calls it the most far-reaching influence operation it has disrupted, and says it banned Iranian accounts too. Source
Anthropic will scan open-source projects for bugs, for free Projects that sign up get regular scans from its strongest models, and each report comes with proof the bug can be exploited and a suggested fix where there is one, with more than 90 per cent of reports expected to be real. It’s part of the Cyber Mission launched on Thursday, which also gives 11 firms including CrowdStrike and Palo Alto Networks its top models and on-site engineers to protect critical infrastructure. Source