Parallel agents in ChatGPT, quantum circuits Claude redesigned, and $15 million for mathematicians

In the three days to October 17, 2026, eight items from the AI and tech accounts this channel follows were worth carrying. OpenAI gave the top tier of ChatGPT a way to split one task across several agents at once. Google published a set of scientific results it attributes to Gemini 4 Argon, and a research fellow used Claude to redesign eight quantum circuits past the best published designs. Anthropic handed a team from METR employee-like access to its internal systems, and the Simons Foundation, Anthropic and OpenAI put $15 million into helping mathematicians absorb what AI is now producing. The window runs three days because the editions of October 15 and October 16 were not published.

Source scope: the fixed set of public AI and tech accounts this channel monitors, while the personal X connection is not linked. One of those accounts, Benedict Evans, can no longer be read on X, so its posts are absent from this edition.

OpenAI's top ChatGPT tier now splits one task across several agents OpenAI's post on X · 9,365 likes and 1.37 million views on the post at capture · October 16

OpenAI upgraded GPT-6 Astra Pro in ChatGPT on October 16. The company says responses are faster and that the model now uses plugins and tools better 1.

The structural change is the parallel split. Multiple agents can now take different parts of one task at once, and the user can follow each agent's progress and open a completed task to review its work 1. The post presents that as new, alongside the speed and tool-use improvements.

OpenAI's release notes describe the same update under a narrower name, as a change to the GPT-6 Pro model in the chat tab. The notes put availability at eligible Pro users, with managed workspaces depending on their settings 2.

For a working user, the reviewing step now lands on them: several agents finish parts of a job and someone has to read the parts. The release note describes the interface for that review and gives no figure for how much faster a parallel run finishes.

Four Gemini 4 Argon agents beat Minecraft in 27 hours, with no human writing their code The arXiv paper by Faldor, Singh and Raileanu, and Google DeepMind's post on X · 2,037 likes and 131,066 views on the post at capture · October 15–16

Maxence Faldor, Satinder Singh and Roberta Raileanu put four Gemini 4 Argon agents into a sandboxed Minecraft survival world with one goal: defeat the Ender Dragon 3.

Each agent controlled a player through Mineflayer, an open-source library for writing Minecraft bots, and coordinated with its teammates through in-game chat and shared memory. Nobody assigned roles or wrote game logic. The agents planned, wrote and debugged every program themselves.

From a randomly generated world on Easy difficulty, the team wrote 2,783 programs and beat the dragon in 27 hours and 14 minutes, dying 605 times on the way 3. The authors call it the first end-to-end autonomous completion of the game, using only the standard library, with no pre-scouted seeds and no game-specific tools written by humans.

Two limits travel with the result. The difficulty was Easy, and the authors designed and wrote up their own evaluation. Researchers watched on a live dashboard and could pause or stop the run, though the paper says they had no cause to intervene. Ethan Mollick recommended the paper as an accessible account of agents coordinating on a complicated goal without being told how 4.

Gemini 4 Argon's scientific results: a 60-year-old plasma conjecture, five enzymes, a faster quantum decoder Google DeepMind's thread on X · 2,452 likes and 160,589 views on the opening post at capture · October 14

Google DeepMind published a thread of results it attributes to Gemini 4 Argon across four fields, most of the work done by its own researchers. In quantum computing, Ben Jaderberg used Argon to translate an internal version of AlphaQubit, Google's quantum error-correction decoder, into a single optimized hardware megakernel and nearly doubled inference throughput 5.

In plasma physics, a paper with Jonathan Citrin, Brendan Tracey and Matt Landreman reports two new families of counterexamples to Grad's conjecture, an assumption that has stood for nearly 60 years, worked out with Argon 6. The paper generalizes two families Landreman published earlier and removes their stellarator symmetry 7.

In protein design, the company says Argon compressed enzyme engineering into hours and produced five serine esterases confirmed in the lab, in a zero-shot workflow 8. Beside the thread, Google opened its highest Antigravity usage tier to 10,000 scientists free for six months 9.

Every result in the thread comes from Google's own teams and was announced by Google. The plasma paper is on arXiv and can be read on its own terms; the throughput and enzyme figures rest on the company's account.

Claude Mythos 5 redesigned eight quantum circuits past the best published designs Anthropic's research post, and its post on X · 4,262 likes and 561,939 views on Anthropic's post at capture · October 16

Anthropic gave Claude plain-language specifications for eight quantum circuits drawn from chemistry, optimization, physics simulation and number theory, plus an automated evaluator that checks each design for correctness and scores it on qubits multiplied by gates 10.

Claude Mythos 5 then ran the loop on its own: propose a circuit, receive a score, revise, with no human input between rounds. Across the eight tasks it produced designs that used both fewer qubits and fewer gates than the best published versions, sometimes by an order of magnitude or more.

Modular inversion carries most of the cost of adding points on an elliptic curve, and it is where the method shows. A straight quantum version of the standard divsteps routine has to store a long record of its own branching decisions, which costs an enormous number of qubits. Claude proved new algebraic identities that let the circuit recover that record from the numbers themselves, giving the first reversible divsteps circuit with no stored history, then formalized the argument in Lean 10.

One design in the resulting family uses 12 per cent more qubits than the smallest previous circuit and 17 times fewer gates. Point addition took 53 hours of largely unsupervised work. Andi Gu led the project inside Anthropic's STEM Fellows Program, with mentoring from Alek Kemeny and collaboration from Harvard's Susanne Yelin and Mikhail Lukin; the full paper is on arXiv.

Mollick: the chain of thought has stopped being a window into what an agent is doing Ethan Mollick's thread on X · 321 likes and 28,874 views on the opening post at capture · October 15

Ethan Mollick, a Wharton professor who studies how people use AI, argues that the value of a model's chain of thought has collapsed in the agentic era, the summarized version included 11.

Volume did it. When agents call other agents and a chain of thought runs to thousands of pages for a simple task, reading it no longer tells you where the work is heading 12.

His substitute is procedural: ask for interim outputs and tests, and use AI to troubleshoot the problems you find. That puts more weight on the orchestrator model, which becomes the only thing tracking every task 13.

The claim comes from use and carries no measurement. Its practical value is in the check it leaves behind: when the reasoning runs too long to read, inspect the intermediate results instead.

Mollick: a cheaper model is not a drop-in replacement, because small accuracy gains compound Ethan Mollick's posts on X, citing a paper by Sinha and colleagues · 1,181 likes and 61,757 views on the paper post at capture · October 15–16

Mollick used a paper by Akshit Sinha, Arvindh Arun, Shashwat Goel, Steffen Staab and Jonas Geiping to argue against swapping frontier models for cheaper ones. The paper finds that small gains in single-step accuracy compound into exponential gains in the length of task a model can finish 1415.

To separate execution from reasoning, the authors hand the model the plan and the knowledge a task needs. Larger models still complete many more turns, even where small models score near-perfectly on a single turn. Per-step accuracy falls as steps accumulate, and a model grows likelier to err when its own earlier mistakes sit in its context, an effect that thinking reduces.

Mollick's conclusion is that the saving arrives up front and the worse work arrives at the end. Small models do plenty of valuable work, he writes, as long as larger ones orchestrate them and check their errors 16.

The day before, he made a larger claim: that harnesses plus strong models are retiring the last famous hard benchmarks. He named three: METR's long-task evaluations, which time how long a piece of work a model can finish unaided; GDPval, OpenAI's set of economically valuable tasks; and ARC-AGI-3, the newest version of the abstract-reasoning test 17. That is his forecast. The paper explains the mechanism behind it; the benchmark retirements themselves are his reading of the state of the field.

Anthropic gives a METR team employee-like access, and keeps the right to redact Anthropic's announcement, and its post on X · 3,728 likes and 518,185 views on Anthropic's post at capture · October 15

Anthropic has onboarded six employees of METR, a nonprofit that evaluates frontier models, with company hardware, office space and access to internal tools roughly matching what its own risk-assessment teams get 18.

METR will publish its analysis of Anthropic's risk levels, incidents and practices. Anthropic cannot veto publication, but it can require redaction of predefined categories such as customer data, privileged information and model architectures, and Anthropic decides what falls into them 18.

Access runs in three tiers. A first-tier evaluator gets the access Anthropic chooses and can publish only with its approval. A second-tier evaluator publishes without approval and sees more. METR sits at the third tier, with employee-like access and publication rights close to the second 18.

SecureBio is reviewing the chemical and biological weapons chapter of Anthropic's August risk report, and Accenture with Faculty has a wider evaluation and red-teaming partnership; Anthropic says it is in early talks with Hugging Face. The post compares the setup to bank supervision, where regulators seat examiners inside the largest banks. One asymmetry stays in place: the host decides which information counts as sensitive.

$15 million to help mathematicians absorb what AI is now producing The Simons Foundation's announcement, with posts from Anthropic and OpenAI · 6,159 likes and 912,302 views on OpenAI's post at capture · October 15

The Simons Foundation announced a joint program with Anthropic and OpenAI worth $15 million in combined funding and resources. The foundation manages and awards the money, including through an open call for proposals, for workshops and initiatives that help researchers evaluate recent AI-assisted results, learn the tools and set standards for their use 19.

Anthropic and OpenAI each put in $5 million. Anthropic describes its share as funding, compute and Claude access; OpenAI describes its as a contribution to the initiative plus access to its models and computing resources for foundation-supported researchers 2021. The foundation says that access extends past mathematics into basic science.

The money answers a concrete problem. AI systems now produce mathematical results that other mathematicians have to check, understand and build on, and the foundation was an early backer of Lean, the proof assistant that made machine-checkable AI proofs possible 19.

Simons Foundation president David Spergel put the program in the community's hands: the mathematics community should be driving how these tools are evaluated, adopted and used, "not any single company or institution" 19. Both companies are among those producing the AI-generated results the program exists to examine.