Good morning. The math story we’ve been tracking took a turn overnight: three of OpenAI’s claimed proofs have been retracted over a sign error, prompting a wider conversation — including from Terence Tao and Scott Aaronson — about what it even means to “do math” when the proofs are unreadable. Meanwhile, OpenAI’s finances got a $20B haircut, Anthropic is giving away vulnerability scans to open source, and someone finally asked out loud why nobody’s panicking about DeepSeek.
OpenAI retracts three math results. Dan Roberts posted on X that OpenAI withdrew three of its ~400 claimed proofs after a sign error invalidated the arguments, while adding six new Lean formalizations and 19 modifications. About 42% of the top-line results are now formally verified. HN was mixed — some called 3/400 a respectable error rate, others pointed out the retracted proofs were the natural-language ones and asked why those were mixed in with Lean-verified work at all. One commenter noted the whole repo now reads like a software changelog: “release 1.3.42: retracted papers 139 and 140, fixed a sign error in paper 47.”
Scott Aaronson and Terence Tao weigh in. Aaronson’s post, The Mathocalypse, recounts his wife Dana Moshkovitz — who spent her career on the Unique Games Conjecture — describing OpenAI’s apparent proof as “alien” and “psychedelic,” so poorly written she needed AI help just to parse it. Tao’s take, circulated on Mastodon, is that “Math 2.0” needs to value understanding and communication, not just outputs, and that dumping unreviewed proofs on the community shifts a lot of thankless verification work onto humans. Buried in Aaronson’s post: the model was given ~8,000 open problems and solved about 5% of them, which is either stunning or modest depending on your priors.
OpenAI’s revenue is $20B lower than reported. OpenAI’s annualized revenue is “approaching $50 billion,” not the $70B figure that circulated last month, per TechCrunch and CNBC. The higher number came from investors trying to match Anthropic’s accounting method, which includes cloud partner sales. The correction dragged Nvidia down 3%, Oracle 6%, and CoreWeave 8%, and sharpens questions about the $852B valuation ahead of a 2027 IPO. Ed Zitron readers were, predictably, not surprised.
Fired OpenAI safety researchers push back. Three safety researchers dismissed last week for allegedly mishandling sensitive information published an open letter denying wrongdoing, saying their external collaborations were standard and within their mandate. They warn the firings have created a chilling effect on internal safety culture. OpenAI shared an internal memo denying the dismissals were retaliatory but hasn’t formally responded to the letter.
Claude Haiku 5.5, one day later. We covered the launch yesterday, but the pricing debate has crystallized: input is $0.10/MTok under 100K tokens and $0.50 above — a 5x cliff that several developers are calling unworkable for anything long-context. On the upside, Plotly’s benchmark showed 9x cheaper runs with two letter grades better accuracy, and the monthly API credits for Max subscribers ($100–$500) are getting consistent praise as a genuine perk. Details here.
Why isn’t anyone freaking out about DeepSeek 4.1 Flash? A blog post from dgt.is argues DeepSeek’s latest Flash model matches Claude Opus on real coding and research tasks for a fraction of the price, thanks to a ~437x KV cache reduction. The HN response was skeptical but informative: most heavy users are on flat-rate subscriptions where API pricing is beside the point, Flash tends to burn more tokens so raw cost comparisons mislead, and GPT 5.6 Sol and Opus 5.5 still beat it on quality for complex work. Still, several commenters said Flash has quietly become their default for agentic workflows.
Anthropic offers free security scans for open source. OSS Scanner is a new opt-in service that runs periodic vulnerability scans on open source projects using Claude Mythos. There’s no human review, so false positives are expected. It arrives at an awkward moment: the Linux kernel and other projects are already drowning in AI-generated bug reports, and it’s not obvious whether more automated scanning helps or compounds the problem.
That’s it for today. If you’ve been following the math saga, Aaronson’s post is worth reading in full — it’s the most honest account yet of what it actually feels like to watch your life’s work get solved by something you can’t read.