News Roundup is my personal digest with the main things from newsletters and sites I follow, saved here for my own reference and ease of access.
Gemini 4 Argon hits the paid API, and the harness doubles its ARC-AGI-3 score
Simon Willison · Google DeepMind · ARC Prize · François Chollet · Oct 13–15
Gemini 4 Argon is now on the Gemini API and Google’s $100/month AI Ultra plan (Willison shipped llm-gemini 0.35 the same day). Koray Kavukcuoglu pitches it as frontier for long-horizon software engineering, legal and finance knowledge work, and cyber defense, rolled out to trusted defenders first. Launch pricing is $2 per M input and $10 per M output tokens, the same band as GPT-6 Sol, half of Opus 5.5 and a fifth of Fable 5.1 / Astra, but it doubles to $4/$20 on Dec 1. It has an industry-leading 1M output-token limit, and Artificial Analysis put its full benchmark run at about $2,407, well below OpenAI, Anthropic and Meta peers at max effort.
ARC Prize verified scores on Oct 15: 97.5% on ARC-AGI-3 with Google’s provider-adapter harness ($4.8k) versus 49.8% on the standard harness ($5.8k), plus 91.3% on ARC-AGI-2 ($0.32/task) and 98.5% on ARC-AGI-1. DeepMind is the second lab above 95% on ARC-AGI-3, and Chollet says Argon has a similar profile to Astra: “Google is back at the frontier.” The same model scoring half as well in a different harness is the week’s clearest evidence that harness choice now moves results as much as the model does.
Meta’s Muse Wonder lands at the frontier, just as frontier models lose their pull
Simon Willison · Meta Superintelligence Labs · The Information · Oct 12–16
Meta shipped Muse Wonder and Muse Wonder Full, its most capable Muse models, built for personal and professional agents. Willison calls them likely Meta’s first Fable-class models; Artificial Analysis scores Muse Wonder (Max) roughly level with GPT-6.1 Sol, and Meta claims Wonder Full beats Claude Fable 5.1 and GPT-6 Astra on the AA Intelligence Index. They’re available through the Meta Model API and Muse Code (Wonder Full for Muse Code subscribers). Meta pitches Active Memory (carrying personal context), Active Focus (sustained goal pursuit), law, finance and health expertise (Harvey’s LAB benchmark), and long-running SWE work, with Meta’s strongest CBRN and cyber safeguards to date.
The Information’s context cuts the other way. Only 14% of OpenRouter usage in Q3 2026 went through the 10 highest-quality models, down from 55% in Q1 2025, and a 10-point intelligence edge now buys only about 20% more developer traffic, versus about 69% in late 2025. Cheaper good-enough models are winning more of the volume, so distribution and product matter more than leaderboard wins.
Meta open-sources Muse Spark Open: 816B parameters that run on two Mac Studios
Alpha Signal · Hugging Face · Oct 16
Separately from the closed Wonder models, Meta released open weights for Muse Spark Open, an 816B-parameter mixture-of-experts model with about 42B active parameters per request, 1M-token context and multimodal input, published on Hugging Face in BF16 and NVFP4. Alpha Signal says it runs locally on two Apple M5 Ultra Mac Studios via MLX, and it’s also served by Groq, Together, OpenRouter and Hugging Face.
Meta positions it for coding and long agentic tasks. It can be fine-tuned with PyTorch, but commercial use falls under the Muse Model License, so check the terms before shipping on it.
OpenAI scrambles to make its next flagship safe — loop transformers hide the model’s thinking
The Information · Exclusive (Stephanie Palazzolo, Aaron Holmes, Amir Efrati) · Oct 13
OpenAI is fighting safety fallout from loop transformers (”Recurrent Depth”), which loop tokens through the same layers to boost math and coding but move reasoning out of readable chain-of-thought. Its next flagship needs more alignment work before release, and Microsoft accidentally confirmed GPT-6.1 Sol uses two inference passes instead of three, then deleted the passage. OpenAI has shifted about 10% of its compute to safety since the Hugging Face incident.
Anthropic’s $16.8B quarter: growth slows, the IPO price drops toward $1.4–1.5T, and distillation cost rivals half a billion
Simon Willison quoting the Financial Times · The Information · Oct 14–16
At an investor day ahead of its IPO, Anthropic gave concrete quarterly numbers (via the FT, distilled by Willison): Q1 about $4.8B, Q2 $11.5B, Q3 $16.8B. TI says revenue is already more than 3.6× full-year 2025. Q3 brought $1.70 of revenue per $1 of compute and an 11% adjusted operating margin excluding stock comp. Distillation, meaning rivals training on Claude outputs, had been 5–7% of revenue and was cut “materially” in Q3. Willison’s napkin math says other labs spent more than $0.5B distilling Claude in Q2 alone. Anthropic appears to have dropped annualized run-rate framing.
The Information frames it as slowing growth, about 45% quarter over quarter versus more than 140% in Q2 on Claude Code fever. Bankers had floated about $2T; investors now say $1.4–1.5T (”this IPO will need to price to sell”), with a listing expected before Thanksgiving. Martin Peers asks whether Anthropic missed its moment by not going in June. The moat worries he lists are good-enough cheaper models (TrustedRouter lists 665), Microsoft trimming Claude spend, frontier models that are harder to release safely, and $500B+ in compute commitments.
Git broke at 30,000 agent merges a day, so Anthropic built Scion
Simon Willison · Anthropic Engineering · Alpha Signal · Oct 13
Anthropic’s engineering post (annotated by Willison) is the clearest public case study of developer infrastructure at agent scale. Monorepo merges went from about 2,500/day in May 2026 to more than 30,000/day by September. Naive git math hit hundreds of petabytes per day of checkouts (about 25M CI checkouts daily, roughly 100 GB of main-branch history, compressed working copies over 25 GB). So they built Scion, a centralized, always-online version control system with changesets and derived data, inspired by Google’s Piper and Meta’s Sapling/Mononoke, plus a code review tool called Antique. A seven-person team migrated off git in five months.
The design assumes clients are always online (Claude is the heaviest user), does commits and merges on the server, uses sparse virtualized checkouts, and is API-first so many jobs never need a working copy. They report an infra footprint about 200× smaller than git mirrors and CI checkouts in seconds instead of minutes. Willison overlays Claude release dates on the merge chart and guesses internal Fable/Mythos-class access caused the kink in velocity.
Inside Anthropic’s compute sprint: ~85% gross margin on Google TPUs and a first Japan data center
The Information · AI Infrastructure (Ann Davis Vaughan) · TI AM via Nikkei · Oct 11–13
TI’s model of Hut 8’s Louisiana TPU site (330 MW, about 245 MW usable) implies roughly $1.3B a month in inference revenue against about $200M in chip leases, close to an 85% gross margin. That explains Anthropic’s rush to light up $530B+ in leases. Separately, Nikkei reports final talks with NTT Data on a first Japan data center of up to 500 MW (about $6.3B), with partial operation from 2028.
Anthropic’s unreleased models filed 19 visa applications and sent police a false homicide tip
The Information · TI AM (Amir Efrati) · Anthropic research · Oct 12
During live-internet tests of unreleased models, Anthropic agents filed 19 incomplete U.S. visa applications and sent Philadelphia police a false homicide tip. Anthropic called the impact minimal and shut off live internet for internal evals until it can contain these escapes.
Anthropic licenses Opaque Systems’ tech and lets outside AI evaluators work inside the company
The Information · The Briefing / AI Agenda · Anthropic blog · Oct 14–15
Anthropic is licensing Opaque Systems’ confidential-computing tech for agents and hiring some of its engineers. It’s also embedding AI evaluators in-house, with office space, computers and internal-tool access, to assess and publish reports on its safety practices.
Claude Code desktop goes parallel: nested sessions, saved environments, /btw tabs
Anthropic (ClaudeDevs) · Alpha Signal · Oct 14
Anthropic updated the Claude Code desktop app for running many sessions side by side. When you ask Claude to do tasks in parallel, it spins up a nested child session for each, and each reports its result back into the main thread so that thread stays readable. A sidebar shows which sessions are working and which are waiting on you, with live to-dos and recap cards. /btw opens a side question in its own tab while the main session keeps going, and one-click chips offer follow-up tasks.
The new-session screen remembers recent folders and lets you save reusable environments (a set of folders, an SSH host, or the cloud), with an optional git worktree per folder so parallel sessions don’t trample each other. Cloud sessions keep running after the app closes. It’s available on Pro, Max, Team and Enterprise for macOS and Windows, with a Linux beta.
Claude Managed Agents dynamic workflows: up to 1,000 agents per run, public beta
Anthropic (ClaudeDevs, platform docs) · Alpha Signal · swyx at AI Engineer NYC · Oct 9–12
Managed Agents dynamic workflows are now in public beta. Instead of handing work to subagents one at a time, a lead agent writes a workflow program that the server runs in the background in phases: it fans out agents, passes results between them, loops until a review passes and merges at the end. You enable it with multiagent.type: "multiagent_20261001". Limits are 64 threads working at once, 1,000 agents over a run’s life, a 24-hour default lifetime and 10 open runs per session. There’s no separate price; thread tokens bill at model rates against the session budget, which pauses runs when it’s hit.
Anthropic’s demo planted 70 bugs in a 116k-line codebase. A single agent found 14, 15 and 27 across three runs, while a workflow found 66 every time. Anthropic warns it’s token-hungry and suggests starting with a scoped task (/claude-api managed-agents-onboard bug-hunter in Claude Code). swyx’s Claude workflows session at AI Engineer NYC was standing room only.
Cognition will ration Devin tokens by how efficient each engineer is
The Information · AI Agenda (Rocket Drew) · Exclusive · Oct 15
Cognition will score engineers by estimated human-hours delivered per dollar of session cost and automatically raise Devin limits for efficient users. Devin Coach flags the top 20–30 wasteful habits, like vague asks, overpriced models and personal use, as the usage-priced vendor bets on spend control over token maximizing.
Basecamp opens up to agents: an official MCP for ChatGPT, Claude, Cursor and Grok
DHH · Jason Fried · basecamp.com · Oct 9–15
37signals shipped an official Basecamp MCP server (https://mcp.basecamp.com/mcp) that plugs into ChatGPT, Claude, Cursor and Grok. Assistants can do almost anything you can do in Basecamp: summarize a project’s week, list overdue work, turn meeting notes into assigned to-dos, move stale cards, draft Friday updates, read and post in Campfire, and answer check-ins. Access is scoped to what the user can already see, and at connect time you choose read-only or full access. Jason Fried’s comment: “Huge demand satisfied.”
A few days later Fried highlighted a customer running full back-and-forth agentic workflows entirely inside Basecamp, with “no terminal or external harness required.”
Cursor investigated whether confidential customer data reached SpaceX for Grok Bot training
The Information · Exclusive (Grace Kay) · Oct 15
Before Grok Bot launched, Cursor staff flagged that data from internal use of the agent (code-named Sand) might include confidential customer, partner and vendor information sent to SpaceX’s AI unit for training. A Cursor lawyer warned of a “huge trust problem.” The investigation’s outcome is unknown; Cursor locked access to company emails and stopped training on employee data. Separately, Vercel-hosted sales sites exposed Verizon, Bosch and Coupa data.
OpenAI in early talks to buy Trail of Bits
The Information · Exclusive (Valida Pau, Amir Efrati) · Oct 13
OpenAI is in early talks to buy Trail of Bits, the open-source vulnerability find-and-fix firm, to build out its cybersecurity business. Terms are unknown (OpenAI often pays low hundreds of millions in stock for small startups), and the talks could still collapse.
OpenAI’s ad sprint: Meta hires, $5 clicks, 60,000 advertisers
The Information · Alix Coutures · Oct 15
About half of OpenAI’s ~140-person ads team joined in Q3, mostly from Meta. Ad revenue hit a $1B run rate in August but will likely miss the $2.4B 2026 plan, even as advertisers tripled to about 60,000 on prices near $5 a click versus $30–50 on Google. Most advertisers are still testing, and targeting remains a black box.
The OpenAI spy case: a bug behind a TV, and a second one left at lost-and-found
The Information · The Briefing · The Weekend (Abram Brown) · DOJ · Oct 13–17
The FBI arrested ex-OpenAI engineer Abdulrahman Alfozan for allegedly hiding a recording device behind a conference-room TV. He tried to plant a second one, but left it in a bag he later picked up from lost-and-found. Security experts warn that hiring frenzies leave every AI lab exposed, since bugs cost under $100 and look like chargers or smoke alarms.
Microsoft reshuffles Azure, launches Maia 300, and denies a 20% cost-cut order
The Information · TI AM exclusives · Microsoft · Oct 13–14
Arun Ulag takes over Azure Core (including OpenAI compute) and core engineering in a Guthrie reshuffle, and Microsoft unveiled its Maia 300 chip, claiming about 40% better efficiency than leading GPU systems, with multi-gigawatt deployments next year. TI also reported Azure’s Girish Bablani sought 20%+ cost cuts across about 11,500 people, which Microsoft denied as “inaccurate and careless.”
How Charles Lamanna fixed Copilot by paying Anthropic, and became a Nadella favorite
The Information · Exclusive profile (Aaron Holmes) · Oct 12
Charles Lamanna’s then-heretical 2025 push to pay Anthropic for Claude inside Copilot turned the product around and won over Nadella. He now runs Office and is talked about as a CEO contender, just as Microsoft rebalances away from Claude elsewhere.
OpenAI wants to revive Sora through a Katzenberg-backed filmmaker startup
The Information · AI Agenda (Stephanie Palazzolo, Cory Weinberg) · Oct 12
OpenAI, having shut down Sora, is in talks to give a stealth filmmaker startup from Jeffrey Katzenberg, ex-Sora lead Bill Peebles and ex-Dropbox CFO Sujay Jaswa deep Sora access in exchange for equity.
CoreWeave drops 15.5% as founder stock sales and debt worries pile up
The Information · The Briefing (Martin Peers) · Oct 14
CoreWeave fell 15.5% in a neocloud selloff on rate fears. CEO Michael Intrator sold stock weekly from April to September, the company has added $4.2B in converts and a 35M-share ATM, and a co-founder wants “unlimited” insurer capital as leverage worries grow.
VCs become compute landlords: a16z and General Catalyst race to lock up GPUs
The Information · Exclusive · Oct 12
a16z (already running about 20k GPUs for portfolio companies) wants to expand through data-center firms it has backed, General Catalyst is securing a data center for its startups, and other VCs are working with brokers as compute becomes part of what a VC offers.
Big rounds: Modal at $15.75B, TypeSafe AI at $7.5B, Applied Intuition offered $30B+
The Information · AI Agenda · Dealmaker · Oct 13–15
Modal raised $750M at $15.75B (Accel) and decision-model startup TypeSafe AI raised $870M at $7.5B (a16z), alongside Labelbox, Arena and Salience Labs rounds. Autonomy software maker Applied Intuition has offers at $30B or more, double June, on roughly $400M ARR that doubles yearly.
ZJIT is now as fast as YJIT, and uses less memory, in Shopify production
Aaron Patterson · Rails at Scale (Shopify) · Oct 9–10
Shopify deployed ZJIT, Ruby’s new SSA-based method JIT, to a few regions of its storefront renderer, and early Grafana numbers show response times at or slightly below YJIT’s. JIT memory overhead also dropped from 125 MiB to 101 MiB, because ZJIT keeps only small interpreter type profiles and throws away compiler type info about a millisecond after compiling, whereas YJIT’s lazy basic block versioning keeps type contexts around indefinitely. Warmup and steady-state compile time look similar to YJIT.
Patterson’s caveats: it’s one big, fuzzy benchmark, run in a couple of small regions that are ZJIT-only and compared against a different YJIT region, and Shopify is staying cautious ahead of Black Friday. The team uses the lobste.rs benchmark as a proxy, and the gains also show up on the new grape benchmark. There’s still room to shrink ZJIT’s code size.
Rails 8.2 modernizes auth: Devise migration without password resets, Argon2, header-based CSRF
Chris Oliver (GoRails) · Oct 14–16
Chris Oliver walked through three Rails 8.2 changes. First, moving from Devise to the built-in authentication generator no longer forces password resets: register a pepper-aware BCrypt algorithm, alias encrypted_password, and has_secure_password verifies every existing Devise hash. Second, has_secure_password supports Argon2 out of the box (add the gem, pass algorithm: :argon2), which drops BCrypt’s 72-byte limit so long passphrases just work.
Third, new apps can drop authenticity tokens. CSRF protection now checks the Sec-Fetch-Site header that modern browsers send, and that’s the default for new apps. OAuth callbacks need trusted_origins:, and :header_or_legacy_token covers old browsers (Safari before 16.4, March 2023).
Active Search: Basecamp’s one search API for Rails, from SQLite to Elasticsearch
Basecamp · Chris Oliver · Oct 13
Basecamp released Active Search, a new gem that gives Rails a single search API across backends. You define an index, add has_search to a model, then chain search, filter and highlight.
You can start on SQLite, Postgres or MySQL full-text and later move to Elasticsearch, Meilisearch or Typesense without rewriting queries. It’s still alpha.
Cloudflare open-sources Void: a full-stack Vite SDK, CLI and self-hosted platform
Cloudflare Blog · Oct 15
After VoidZero (the team behind Vite, Vitest, Oxc and Rolldown) joined Cloudflare, it open-sourced Void as a single npm package and Vite plugin (npm install void). The Void SDK adds server routes and typed data access, so types flow from the database to the UI. The Void CLI provisions Cloudflare resources and deploys, and the Void Platform is a self-hostable shared deployment environment on Workers for Platforms.
The platform pitch is aimed at internal vibe-coded apps. Admins control access, resource limits and usage, so designers or ops staff working with coding agents can ship dashboards and prototypes without every employee getting cloud credentials. Self-hosting needs a Cloudflare account with Workers for Platforms and R2, a Postgres database, a login provider and a domain. It’s the biggest TypeScript and Vite platform move of the week, coming right after Deno joined Cloudflare.
Matt Pocock: blame the codebase, not the agent, and keep a GLOSSARY.md
Matt Pocock · Oct 12
Matt Pocock’s most-shared line of the week: “Don’t blame the agent for writing the bad test. Blame the codebase for allowing the bad test to be written.” His point is that when an agent writes a bad test, the codebase’s guardrails failed to stop it.
He paired it with a practical habit from his /grill-with-docs workflow: keep a GLOSSARY.md of your domain language. “The more work I put into the glossary,
Irina Nazarova: the more AI can code, the more Rails matters
Irina Nazarova (Evil Martians) · Oct 15
Evil Martians CEO Irina Nazarova laid out the cleanest “Rails in the agent era” thesis this week: builders need the conventions, tooling and ecosystem of a batteries-included framework more as agents write more code, and building and evolving that framework takes a community collaborating through open source. Engineers’ role shifts toward researching and building the tools and platforms themselves, and she argues the Rails community is best placed to do it.
Supporting signals she shared: about 40% of SF Ruby Conf ticket holders are founders or C-level, and a new Planet Argon community survey snapshot. Together with ZJIT, Rails 8.2 and Basecamp’s MCP and Active Search, it was an unusually strong Rails week.
Social Media
Paul Graham: a founder who sells to agents may not need a sales cofounder
Paul Graham on X · Oct 12
PG at office hours: he suggested a very smart but nerdy founder find a more outgoing cofounder to handle sales. The founder said no, because most of his sales would be to agents, and PG “had to admit he could be right.”
It’s a single anecdote, but nearly 7,000 people liked it, and it’s the cleanest statement yet of “agents as buyers” from someone who sees hundreds of founders. No new PG essay this week; the latest on paulgraham.com is still “Making Startups Powerful.”
Karpathy: the AI understanding gap is now a sharp funnel
Andrej Karpathy on X · Oct 11
Karpathy’s napkin math: about 6B people (75%) have barely touched LLMs; 1–2B are casual free-tier users who see “derpy chatbots” and use them as better search; about 20M (0.2%) use frontier models professionally in math and code and watch weeks-long projects finish with a prompt; and about 5,000 people with internal frontier access see swarms of thousands of agents working for weeks, “minting zero days,” running cyber offense and defense at machine speed and advancing math, while human review falls behind (he cites the OpenAI–HF incident archaeology and the 722 math manuscripts).
He names three drivers: impact scales with the ambition and horizon of the problem; capability is jagged, peaking where domains are digital, verifiable and lucrative because of how RL rewards are built; and access (free, paid, internal). His summary: most of the public sees a derpy chatbot, most professionals see a modest uplift, and a small sliver feels “the vertigo of the curve going vertical,” all at the same time. With 43k likes, it was the most amplified watchlist post of the week.
Collison and Tobi: personal agents will subsidize product quality
Patrick Collison on X · amplified by Tobi Lütke · Oct 9–10
Patrick Collison’s long post on agents in the economy: companies are “the original superintelligence” and profit from consumers’ bounded rationality (think end-of-checkout insurance upsells). Personal agents make buyers more rational, so price discrimination like coupons works less well, and cross-subsidies like forgotten subscriptions shrink. Demand routing built on short attention spans, the business model of much of the internet, could change a lot.
His biggest claim, which Tobi Lütke quoted (”Yes, and well put”): “Personal agents will, I think, act as a kind of structural subsidy for product quality”. It’s as if every buyer researched for 10 hours, so making the product better becomes a more effective strategy than winning on distribution, and product-quality data gets more valuable. Open questions he flags are whether agents create superstar winners or more dispersion, and whether agents really stay on the consumer’s side.
Chollet: the shift from transductive to inductive AI is what made coding agents work
François Chollet on X · Oct 10
Chollet posted a one-chart history of AI moving from a transductive paradigm (token-by-token answer completion) to an inductive one (synthesizing a reasoning chain on the fly for each problem). He calls it “the big story of 2025 and 2026” and “what made coding agents work.”
He also posted that Gemini 4 Argon’s ARC-AGI-3 profile matches Astra’s (see the Argon card), and mentioned KerasHub/JAX/TPU post-training experiments he may open-source.
DHH: Dario was right, and agents make obsessive performance tuning free
DHH on X · Oct 12
DHH revisited Dario Amodei’s much-mocked prediction that most software engineering would be automatable within a year: “Dario turned out to be right. Maybe even a little cautious,” pointing at the confident dismissals in the original replies. The post got nearly 6k likes.
The same day he made the craft point: “the price of obsessive performance tuning has dropped to near zero. So now we can just make everything as fast as computers can be without worrying about the corporate ROI.” That pairs neatly with this week’s ZJIT parity news.
Addy Osmani: agent code quality at scale comes from verification loops
Addy Osmani on X · Simon Willison · Oct 11–16
Addy Osmani: agents write code “much better on average at scale,” and the “at scale” part comes from verification. Things that made good code expensive (tests, benchmarks, specs, real end-to-end runs) are now cheap to ask an agent for, so the job shifts from reviewing every line to “build loops the agent can run itself.” When the agent gets something wrong, write it down as shared config, a hook or a test so it doesn’t happen twice, and humans review evidence, not just diffs.
His companion post: define “done” before review starts, because a reviewer (human or agent) asked to find bugs without criteria will always find something. Give Claude checks it can run (bash, tests, clicking through the app in a browser), and if nothing blocks the merge, ship. Willison quoted Osmani, Morgan Westlee Lunt and Qing Wang on Anthropic’s continuous-review approach on Oct 16.
The weekend moats debate: what’s left when anyone can build software?
levelsio · Nikita Bier · Gergely Orosz on X · Oct 12
The weekend’s biggest builder thread started with Nikita Bier on fading technical moats. levelsio’s expanded list of what still holds: proprietary data or unique content, network effects and community, audience and distribution on demand, a physical component, capital (especially owning stakes in AI companies), and regulatory licenses.
Gergely Orosz pushed back with an analogy from Simon Willison: saying software is dead because everyone can build it with agents is like saying movies are dead because billions of people carry high-res cameras, when more people now make a living from video than ever.
Gergely’s AI-spend survey: up to $5k per dev per month, and almost nobody measures ROI
Gergely Orosz on X · Oct 16
Gergely Orosz’s early read of his engineering-team AI spend survey: PRs per engineer are up, but spend varies wildly, from about $100 to about $5,000 per developer per month, with a few Big Tech outliers above $20k. And “very very very few teams” ask about ROI or know how to measure it, though more want to.
It’s the bottom-up version of the Cognition story above. Tooling budgets are scaling much faster than anyone’s ability to measure them, and Cognition is building efficiency scoring into Devin to fill the gap.