Good morning.
The promise of AI has always been tangible progress in science, physics, medicine and mathematics.
Well, it's happening.
On Tuesday, OpenAI released a collection of AI-generated mathematical results. Its public repository groups the manuscripts into 372 families. There are proof files, revisions and caveats.
Some of these manuscripts are very early stage concepts, others are far along.
Obviously every such release starts the predictable circus again. People start proclaiming the death of mathematics and tell us mathematics is now solved. Others tell us that this entire exercise is useless but don’t tell us why. They just hate the technology I guess. But nobody bothers to look at this from up close.
Because that requires someone to actually read these papers. An outrageous demand in the age of instant opinions.
OpenAI’s own documentation makes an essential distinction: these results are at different stages of verification. Some have formal proofs ; others still need scrutiny. Even a formalised result does not automatically certify every claim in the accompanying paper.
So no, a mountain of PDFs is not a mountain of settled truth. But dismissing the whole thing because it arrived by machine would be equally stupid.
The reaction reported by WIRED includes concerns about credit and the sheer volume of work being released. Those are legitimate questions. Who checks the results? Who explains them?
We have spent years worrying about machines doing our work. We should also prepare for machines producing more work than we know how to judge. Proposals. Designs. Legal arguments. Software patches. Each one looking plausible, polished and perfect.
And more difficult to imagine : we might soon go into the territory where us humans cannot check the AI’s work anymore because it has surpassed us in intellect. Where the AI is doing things we cannot comprehend anymore.
It’s like Einstein would attempt to explain his theory of relativity to a 5-year old. They simply don’t have the brainspace for it.
The scarce skill becomes deciding what deserves to survive contact with reality. And even that skill is being put into question by the limits to our intellect.
We’re truly in uncharted waters.
Welcome to the Blacklynx Brief
Leave Granola and get up to 12 months free of Wispr Flow Notetaker + Dictation
If you have paid time left on an individual Granola plan, we'll match it with a Wispr Flow subscription that includes Notetaker and dictation, and add bonus time, up to 12 months total. Sign in or create a Wispr account and submit proof of your plan to check eligibility.
AI News
OpenAI releases a mathematical avalanche. OpenAI published hundreds of manuscripts from an internal frontier model, alongside research details and some Lean proof formalizations. The results have differing verification status; the release offers material for researchers to examine, rather than blanket confirmation that every claimed result is correct. (OpenAI)
Mistral previews its largest model. Mistral Large 4 is available through a public preview API, with downloadable weights promised by the end of October. The European lab is testing expanded cybersecurity capabilities with vetted partners before that release; its performance claims remain company-reported. (Mistral)
OpenAI introduces text watermarking. OpenAI opened optional textGrain watermarking for selected API models and plans to mark eligible ChatGPT and Codex output in the EU over coming weeks. Detector access initially targets approved experts, and OpenAI warns that editing and short passages undermine reliability. (OpenAI)
OpenAI safety-report lead resigns. David Robinson, who led the writing of OpenAI’s launch safety reports, resigned and publicly criticised the company’s culture of perpetual sprints. His essay argues for stronger organisational safeguards while noting that OpenAI stands by its safety practices. (The Atlantic)
Anthropic expands verified cybersecurity access. Anthropic introduced three Cyber Verification Program tiers, extending advanced models and reduced cyber blocking to qualifying security professionals. Access depends on the work involved and verification requirements, with the least restricted tier reserved for closely vetted organisations handling especially sensitive systems. (Anthropic)
AI Quick News
Trump named intelligence director Jay Clayton to lead a new Super Intelligence Force coordinating the federal government’s AI efforts.
Sesame’s personal assistants can connect to services including Gmail, Calendar and Drive, with app permissions managed separately for each assistant.
The Wall Street Journal reported three OpenAI researcher dismissals over alleged misconduct including sharing confidential information with an outside AI-safety group.
OpenAI’s Meetings plugin is in beta for Pro and Business users on Mac, turning calls into notes and action items with participants’ consent.
HeyGen launched a general-purpose video model built on MiniMax H3, with introductory pricing from $0.01 per second through October.
Elon Musk said SpaceXAI would become SpaceXSI, adopting Trump’s preferred “super intelligence” terminology.
Claude Code Mods let developers customise agent behaviour and interfaces using JavaScript or TypeScript.
Anthropic put Claude for Google Workspace into public beta for paid plans, adding assistance inside Docs, Sheets and Slides plus connectors for editing those files.
Vals AI reported two antiferromagnetic semiconductor candidates found with Claude agents, with predicted properties still requiring experimental validation.
Sam Altman told Politico that society should accept some harmful outcomes in exchange for AI’s benefits.
Meta published six research papers developed through collaborations between mathematicians and Muse Spark.
OpenAI says it disrupted a reasoning-extraction campaign, attributing a core cluster to people associated with Moonshot AI rather than claiming all activity came from one actor.
DoorDash introduced a corporate-ordering connector that lets company AI agents assemble, place and track office food orders.
The Pentagon’s Project Meridian gives Elon Musk, Palmer Luckey and Newt Gingrich 120 days to recommend technologies for future warfare.
Norway’s government proposed a temporary smart-glasses ban in selected public settings over privacy concerns; parliamentary approval remains ahead.
Carter Church published a solution to an 1809 military cipher, reporting that GPT-6 Astra completed the decoding workflow in roughly six hours.
California signed new AI worker protections, including restrictions on employers relying solely on automated systems for disciplinary decisions.
Nolla launched a Utah pilot for AI-issued acne prescriptions, beginning with physician approval of every prescription before any later supervised autonomy.
Microsoft launched MAI-Transcribe-2-Streaming and new MAI-Voice models, with the transcription system supporting 60 languages.
CBS opened a 60 Minutes report with an AI version of correspondent Jon Wertheim before revealing the real presenter.
Muse released open-source gadget code for developers who want to build its assistant into their own hardware projects.
Closing Thoughts
That’s it for us this week. Please like and subscribe 🙂
