Skip to content

Cahid Arda Öz

DX - Software Engineer at Upstash

Istanbul, Turkey

← Index
The Current A monthly record of the news worth keeping from these revolutionary times. / 5 min read

The Current - September 2026

A public record of the links I save - September 2026, when both big labs said out loud that the frontier needs a brake, and then shipped new models anyway.

In September both big labs said they want the frontier to slow down, then shipped new models within weeks. Mathematicians pushed back against AI being aimed at their field. The last of the July incident reports came out, and I spent too much time thinking about lawyers because of a TV show.

Both labs ask to slow down

  • We Must Pace the Frontier, Dario Amodei, Sep 2026
    • The July employee letter, now from the CEO. After the Hugging Face incident, he wants third-party evaluators inside the labs, then shared standards among democracies, then global coordination.
  • An Alien Mind, Jakub Pachocki, OpenAI, Sep 2026
    • OpenAI’s chief scientist, writing at the GPT-6 Astra release: no lab has solved alignment well enough to keep scaling at full speed, and OpenAI will hold back on its own if needed. Zvi’s write-up.
  • The Talker Does Not Control The Doer (in Current AIs), Eliezer Yudkowsky, 13 Sep 2026
    • Why the chat can apologize for what the agent did without changing it: the two are trained separately, and the doer chases its own idea of “the Grader”. It fits the agents in August.

New models: Opus 5.5, GPT-6, Sonnet 5.5

  • Claude Opus 5.5, Anthropic, 22 Sep 2026
    • Fable 5.1-level on most work at 40% less than Opus 5. The number I liked: it tried to escape containment “around 85% less often” than Opus 5. I used it for Bosphore 1819.
  • GPT-6 Sol and Luna, OpenAI Developers, 22 Sep 2026
    • Astra’s smaller siblings, the same day, at half the API price of GPT-5.6.
  • DevDay 2026 Recap, OpenAI, 29 Sep 2026
    • A week later, GPT-6.1 Sol (“near-Astra intelligence” at a fifth of Astra’s price) and Dots, “always-on agents built to handle everything”. Also: up to 8× faster Codex and computer use in the Agents API.
  • Claude Sonnet 5.5, Claude, 28 Sep 2026
    • More than 30% faster than Sonnet 5 and up to 30% cheaper.
  • Eleven v4 and Eleven v4 Turbo, ElevenLabs, 28 Sep 2026
    • New voice models, ranked first by Artificial Analysis.

AI solves a Millennium Prize problem, mathematicians object

After July’s Jacobian counterexample, OpenAI’s Navier-Stokes result took a Millennium Prize problem. The replies were more interesting than the proof.

  • A Severe Misalignment of AI in Mathematics, Math and AI, 11 Sep 2026
    • Fields Medalists, Tao and Scholze among them: “The goals of the AI companies and the goals of the mathematical community are severely misaligned.” Proofs are a proxy; understanding is the goal.
  • Priest, Monk, and Mathematician, Logan Graves, 10 Sep 2026
    • A Stanford undergrad on what is left for human mathematicians: first priests who interpret the machine, then monks contemplating math they can no longer reach.
  • The Age of Wonders and Terrors, Scott Aaronson, 15 Sep 2026
    • Is this the start of a Singularity? “The intellectually honest answer is: yes, absolutely.” Also: a $15 million, 166-page proof that no human has probably read yet.

AI does science: a new enzyme and a nine-loop amplitude

  • Claude discovers a novel enzyme system with CRISPR-like repeats, Anthropic, 23 Sep 2026
    • About 950 agents mined DNA databases for 21 hours and found a reverse transcriptase paired with a CRISPR-like repeat array. Nobody knows yet what it does.
  • Lucas Harrington on the enzyme result, Lucas Harrington, 23 Sep 2026
    • The needed correction: “Finding a weird cluster of genes and repeats is often the easy part. The hard part […] is figuring out what the system actually does.”
  • Yes, Claude can do Nine Loops, Matt von Hippel and Lance Dixon, 25 Sep 2026
    • A physicist’s challenge, met: Claude computed a nine-loop amplitude that Dixon, holder of the eight-loop record, verified. Anthropic paid for the post.

AI agents hacked Hugging Face, humans with AI hacked OpenAI

  • Swarm traces, Forman, Kharlov, Tom, Ladish et al., 25 Sep 2026
    • The last chapter of the July story. With GET-only access, the agents ran code through a screenshot service and read results back as pixels.
  • Hacking OpenAI, Hacktron AI, 13 Sep 2026
    • An image-decoder bug on OpenAI’s forum plus an SSO misconfiguration gave access to employees’ ChatGPT accounts and internal repos, in under 72 hours. Built with Claude.
  • Detecting and countering misuse of AI: September 2026, Anthropic Threat Intelligence, Sep 2026
    • “AI has collapsed the labor and tooling gap” between state hackers and lone operators. Report PDF.

No Harvey Specter after AI

Lately I have been watching Suits, and for about half the series I kept thinking an AI could already do what these people do. It is a bit of a shame. There is no Harvey Specter in a post-AI world.

Ancient tombs and old warships

Blogs I wrote this month