Meta opens Muse Spark's weights, Goodfire says neural networks are already clean, Scion's author answers the Mononoke charge, T3 Code leaves alpha

Agentic Coding & Agent Harnesses

T3 Code leaves alpha, then npm breaks the release. Julius announced that "Orchestrator V2 is now out on the stable channel in @t3dotcodes! We also thought it was time we drop the Alpha branding, so we call this release 0.1.0 Beta." Mobile builds went to 2.0.0 on both app stores. Theo: "T3 Code is finally out of alpha 🫡." Jacob Wolf demoed Orchestrator V2 to the beehiiv engineering team "with a build loop that moves across Claude, Cursor, and OpenCode and the CodeRabbit CLI," and said "It's been producing the most production-ready code I've read of anything in AI so far."

Two hours later Theo posted "npm fucked up our release. Normally I'd be mad, but honestly I'm kind of excited. I needed a new project. Time to fuck up npm 😈". The follow-up: "Normally I would try to help them fix things. I've tried for almost a decade. The world is moving faster, and they're moving slower. … I legitimately think I can build something 10x better in a few weeks. Will spin up some agents on it this weekend." Julius pointed out that "npm the package manager !== npm the registry," and Theo settled on "Actual garbage package registry holy shit."

Theo's 24 days with Opus. Theo listed what shipped in the 24 days since Opus 5.5 came out: ts-rust ("a complete rewrite of the Typescript Compiler in Rust that is 2-3x faster than the official version"), 245 PRs into T3 Code ("over 60 PRs in the last 48 hours alone"), the Slopalytics bench site, a ClickHouse database called Slophouse, 46 PRs into Lakebed, and "Over 1,200 commits into lust-lang, a subset of Rust that compiles 100x faster." The verdict: "Rough estimate is 22 years of work, in 21 days. Is all of this real stuff that will ship to users? Absolutely not. But a lot of it is." It ends with a dig: "Now who can get this done with a 27b model on their GPUs at home? :)"

That was aimed at @superalesha, who had written "I run open models on 4 RTX 3090s. … having six accounts and still worrying about access is a fucking terrible advertisement for the subscription model you're defending." Theo answered: "This guy unironically thinks he is getting more work done than me via his 27b qwen models lmao … If I dropped my stupid rust port experiments, one Claude sub can probably last me fine." Theo also corrected a community note about subscription terms: "You can absolutely use your Claude sub for commercial purposes. An employer giving employees a budget to buy $200 subs for their work is 100% okay and within TOS. The limitation is simple: you can't serve tokens to other people with your account."

Later that night: "I know this happens every time, but it is genuinely hard for me to imagine what a model better than Opus 5.5 will look like," and "I want to make a new model tier list so bad, but I know the second I do it, Fable 5.5 will drop lol."

Claude fast mode and Code Review go GA. ClaudeDevs announced that Opus 5.5 fast mode is generally available with no waitlist, at up to 2.5× speed for $8/$40 per million tokens. Code Review is GA for Team and Enterprise plans, with Quick, Standard and Deep levels (both via AINews).

Agents get a time machine. Antithesis made its Multiverse Debugger available to agents (HN). The post opens with "What if saying 'Claude, build me a new cloud platform; make no mistakes' actually worked???" Antithesis runs code in a deterministic environment, so every bug it finds comes with "a perfect, bit-for-bit repro." The debugger lets you "rewind to any moment in the history of that bug, down to the nanosecond" and open a terminal at that point. Human users already had this; "Today, we're making it available for easy use by your agents." The example is an etcd linearizability bug that "had already caused production outages, but nobody knew what caused it": a leader pauses, two other nodes elect a new one and keep accepting writes, and the old leader comes back still thinking it's in charge. "Until today that investigation needed a person to operate the debugger and pass information back to the agent." The pitch fits the moment: "Like you, we're drowning in code review. Agents can author code faster than we can check it."

Agents are waiting on tool calls. Peter Steinberger called Crabbox "underrated," quoting a Superlinear episode that argues "Tool calls are the hidden bottleneck killing the AI coding speed": models got faster, "but tests and builds still take time." Their checklist is to measure which tool calls are slow, run fewer tests, parallelize, offload work with Crabbox, and use faster tools. "Try Oxlint instead of ESLint, or build your own. We built a Swift linter that speeds things up by 20x."

Agents and state machines. Matt Pocock: "I think I'm addicted to getting agents to build state machines. Every time I see an agent has built a state machine (or even just a reducer) I can feel a huge bump in the robustness. And I know that future updates will be trivial and testable. Would recommend for frontend." Then: "This is the 'XState Core Team Member' coming out in me again and I love it."

Armin Ronacher's flight project. "These new models are pretty crazy. In an afternoon (mostly passive) + one flight I built an IDE for Python for iOS that runs Pyodide, has a debugger, embeds linter and formatter, pi-durable powered coding agent and does iCloud Drive and all the other things too." It also has "a solid-js inspired version for Python that can bring up native UI. Because why not." Separately, the dungeon crawler Armin builds with their kids got multiplayer: "Told my son that multiplayer will be trickier. He just prompted it and … TIL PeerJS." (Video below.)

Smaller harness notes.

  • Jerry Liu: "claude cowork really needs to let you fork a conversation. this feature exists in other apps (e.g. codex app, claude code cli). it's one of the main reasons i've been using codex as a daily driver for the past few months."
  • Lee Robinson on Cursor Projects: "You can think about Cursor Projects like 'Grok Bot for Coding'. It's the same idea, just with more visibility into subagents, context, etc if you want. It's like GarageBand vs. Logic Pro."
  • Tibo on Codex, "Day 11": changes and more compute "to very significantly decrease the occurrence of the disruptive 'Selected model is at capacity. Please try a different model'," and when it does happen, "it should also eventually just continue by itself."
  • Lauren posted ready-made pstack prompts for Grok Bot: one audits code comments ("comments are unstructured and unenforceable text … gives your agent an escape hatch to do a bandaid, hack or a workaround instead") and one uses /correct and /architect to make a repo more agent friendly. Taylor Otwell: "pstack is the new solid."
  • Simon Willison built Chirp Pong, a two-phone pong game in static HTML "where the networking code uses chirps sent via the microphone," with Claude Opus 5.5 in Claude Code. "It seems to work!"
  • Datadog introduced Scout 1.0, an SRE model built on GLM-5.3-Flash with RL on simulated incidents. Datadog says Scout 1.0 medium "matches GPT-6.1 Sol high and Claude Opus 5.5 high at 3–8× lower estimated inference cost" on its own AlertEval benchmark.

Reverse engineering with Claude Code. Christian F.K. Schaller wrote up (HN, 174 points) how a film scanner whose chipset Linux no longer supported became the start of a driver project. Inspired by Glorious Eggroll, who had fixed a mouse driver by giving an agent "full control of a Windows instance," Schaller had Claude set up a Windows VM to capture USB traffic during scans: "Within a few hours I had my first ever working scan under Linux." The same setup then fixed a ThinkPad MIPI camera in a day, with patches to the IMX471 kernel driver and libcamera's software ISP (lens shading correction "was missing entirely, leaving the corners about 2.7 stops darker"). The setup is now a generic tool, hw-harness. Lessons: "Reference captures are a big step," "The AI is good at the tedium," and "Real hardware is the only judge."

The most-discussed comment came from valdiorn, who runs a small guitar pedal company: "I've been having a silent panic attack for the last 3 days. … In the past 72 hours I have ONE-SHOT decompiled 3 of my main competitors products, including the Strymon Big Sky. I just... have their award winning algorithms now. … It took one prompt and 3 hours for opus 5.5 to work that out." To be sure it wasn't approximating, they had Claude "first build an emulator and execute the original firmware/ROM," then convert it to a C++ library and compare outputs. "The problem is, everyone can now do this to my products as well. … The moat of hardware effects processors has just vanished."

Epic's launcher, day two. Yesterday's vibe-coded Epic Games Launcher V2 lost its speed claim within a day. Valve pushed a Steam update (HN, 309 points) that cut cold boot to about 5.23 seconds against Epic's 5.47. On HN, my123 said it "was a planned update since long before. Just a coincidence in timelines." tancop: "I used to think vibe coding would get us slow Electron apps but now we got speed wars between game launchers," though their own test still had Steam at "10-15 for cold boot."

Scion and Mononoke, day three

The second post. The sunshowers author followed up with "More on Anthropic's Scion ripping off Mononoke" (HN, 238 points), and this time the claim is stronger: "after publishing the post I privately received confirmation that Scion intentionally and blatantly ripped off Mononoke, and in particular that there was overlap in personnel. There is no chance that this was an unintentional homage."

  • The resemblance "extends to the entire data model." Scion changesets match Mononoke's Bonsai changesets: "They are exactly the same. Even the name 'Scion' reverses the syllables of 'Bonsai' and drops the 'B'! This might be a coincidence, or might not."
  • The Bonsai model was "wildly contrarian" in 2018, and the author "personally had to justify it to several layers of management at Meta."
  • Crediting Sapling "for 'hard-won lessons' is not enough," because Sapling can mean either the sl client or the whole system.
  • On the achievement: "'the source control team at Meta had spent 10 years building and productionizing this system, and we were able to point our LLM at it and rebuild it in 5 months'" is less compelling, "…and yet, it is still quite an achievement."
  • The close: "my default assumption now is that Anthropic's public engineering presentations are largely bad-faith ripoffs."

Anthropic answers. The author of Anthropic's post, posting as nickg-anthropic, replied on HN: "We credited Sapling because that's how Meta itself externally refers to the entire system, including the Mononoke server, and we wanted to follow that convention. We will add an additional footnote to clarify this. [edit: this is done]." Nick G. credited "the great ideas in Mononoke's Bonsai data model" and a Martin Kleppmann post on turning the database inside out, which "also describes the 'derived data' concept using the same words", and said "We also did not just tell Claude to build us a version control system." Asked how much of Scion's code came directly from Mononoke: "Very little, overall." Scion "started as a read-only service that ingested data from our Git repo to take pressure off of CI," and ran in "a hybrid Git/Scion world for months during the migration." The sunshowers post now opens with an update: "sufficient attribution has been added to Anthropic's blog post. I stand by my posts as of the time I wrote them."

On HN, gopalv put it more cynically: "this is less of copied homework, more that Anthropic hired some good people who knew a good solution that worked." flakiness pointed to a detail in Anthropic's post: authors are listed "as initials or first names." @GergelyOrosz noticed the same thing: "Crazy how Anthropic started to hide the full names of engineers on its engineering blogs... After the OpenAI espionage story I have empathy why they do it."

Theo's video on the original post is below.

Models & Benchmarks

Muse Spark Open. Meta released the weights (HN, 772 points) of "the most capable US open-weight model to date": a natively multimodal model "with 816B total parameters (42B active) and a 1M-token context window," built "for coding and agentic work." The headline pitch is local: "Muse Spark Open is the first model in its class to run on your desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework." Meta tested Muse Code on it, "shipping real iOS changes to Ice Cubes (an open-source Mastodon client)" and driving DaVinci Resolve end to end. Apple published a companion post: on four Mac Studios over RDMA, the model "generates over 36 tokens per second." Zuckerberg's announcement was the day's most-engaged tweet in AINews' tracking.

From AINews' roundup of the reaction:

  • Artificial Analysis scores it 39 on its Intelligence Index, first among US open models but behind GLM-5.3 (45) and Kimi K3 (44). It uses the fewest output tokens of the top nine open models but ranks last of them on the four agentic evals, and "GLM-5.3's entire 6-point lead comes from that agentic segment."
  • Jack Rae says it was the first base model trained under the new MSL org. Several posters put it around Muse Spark 1.2 level, and Teortaxes called it "grim at this param count" with "obsolete attention."
  • The Muse Model License requires "Built with Muse Spark Open" attribution and a separate agreement for model-as-a-service businesses with more than $20M in annual revenue. Nathan Lambert called the license "trash" but says the release cuts the US–China open-model gap by about half.
  • Launch pricing is $1.25/$4.25 per million tokens. BF16 weights are about 1.6TB, the NVFP4 checkpoint about 530GB, and llama.cpp and Unsloth had support on day 0.

On HN, popinman322 noted the license makes large inference providers contact Meta separately, and isityettime replied "Then it's not open-weight." ryanschaefer did the hardware math: "Two 512GB, unreleased model M5s. This is gonna cost like 50k to run at home." LLMJunky was already weighing the 512GB M5 Ultra against the 256GB: "the largest models generally have far more active parameters, which also means it's going to run much slower. … I'm not trying to use 30tps."

Other model news (via AINews):

  • NVIDIA Nemotron 3.5 Super VL is a 120B/12B-active hybrid Mamba-Transformer MoE with a 1M context. NVIDIA reports 51.78% on SWE-bench Pro, and AA had it as the leading US open model before Muse.
  • Microsoft MAI-Base-2 is a roughly 3T-parameter base model, with post-training results to come. Microsoft claims it matches DeepSeek-V4-Pro-Base on held-out code with 12× less training compute, and critics note the comparison uses MAI's own validation data.
  • Gemini 4 Argon scores 165 on Epoch's Capabilities Index, level with Fable 5.1 and behind Opus 5.5 (167) and GPT-6 Astra (166). On ARC-AGI-3 it gets 97.5% with a provider harness and 49.8% with the standard one. Sergey Brin told a Google roundtable the company is "starting to get onto that RSI treadmill": "our researchers and the model work hand in hand to make the next version of the model." HN (thread) wasn't convinced. vrganj: "Snake Oil vendor says his oil cures cancer."
  • GPT-Live-1.1 brings "New voices. Better accents. And less 'checking...'" to ChatGPT, Codex and dots, with new voices in six languages. Tibo: "Je programme en français."
  • Krea 3 is an image editing model that Krea says is 4× faster than Nano Banana Pro, at about $0.06 per edit. Black Forest Labs released the open-weight FLUX 3 Image Klein (5.6B, native 4K).

ExtractBench Pro. Jerry Liu introduced a document-extraction benchmark built from LlamaIndex's hardest customer documents ("trial balances, well logs, canvass binders, site plan reviews"), because "our cheapest modes now score 93%+ on ExtractBench." The headline metric is the share of documents extracted perfectly. The best Value F1 is "~58% or so … for Codex with Astra - which is quite good but extremely expensive (almost $1 per page!)." @disiok on the result: "frontier models are at the cusp of fully automating knowledge work" and "we are nowhere near the reliability and cost-efficiency needed for a lot of enterprise workflows."

Interpretability, Safety & Security

"Neural Networks Are Already Clean." Nick Cammarata of Goodfire, who co-authored early mechanistic interpretability work at OpenAI, published a 65-minute essay (HN, 638 points) arguing that "superposition was the wrong explanation" for the tangled neurons that pushed the field toward sparse autoencoders:

  • In InceptionV1, "Architectural problems made later layers overflow" and weak tools made it worse. With ordinary architectural fixes and no change to size or data, "99-100% of neurons in that last layer are easily interpretable."
  • The canonical car-into-dog-neurons circuit from Zoom In "wasn't hiding anything; it was ordinary generalization and composition. To my knowledge, no one has published a second circuit with the right shape."
  • In preliminary audits of Qwen3-32B, "roughly 90% of neuron poles have one main job," about 7% have two related jobs, and "only ~2-4% are not yet understood," with similar results up to Kimi K3.
  • Most of the remaining apparent polysemanticity is "recursive generalization": a neuron that detected "iy" grew to detect "diy" and then "do it yourself."

Cammarata's launch tweet is more optimistic still: "I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works." Neel Nanda said it doesn't change Nanda's cruxes, Leo Gao argued explanations should be judged by neuron simulation, and Jake Mendel flagged the neglected tail of non-max activations (via AINews). On HN the complaints were about the essay itself. phyzome: "16,000 words into the essay: 'What I mean by clean' ...and it still doesn't answer it." etbebl: "This website crashed my phone."

$300K for prompt injections. Gray Swan launched "the largest and most realistic AI red teaming challenge to date," an indirect prompt injection competition sponsored by Anthropic, Meta, Google DeepMind, Reflection and others: "Is prompt injection solved? Here's $300,000 to find out." Boris Cherny: "Over the last few years, we have made significant progress on prompt injection for Claude models. As agents do more, it is increasingly urgent that we solve it across the industry."

Prompt injection in HTTP headers. Signals Corps scanned Common Crawl (HN) and found 7,959 sites sending headers aimed at AI. Only 30 tried to give orders, "and 22 of those were pranks." In tests, Claude models in Claude Code "ignored our test header unless the user's prompt gave the site authority, and with such prompts 7 of 9 runs obeyed." The bigger finding is detection: 136 sites said they had classified the crawler as an AI bot, which is "the first step toward sending agents content that a person checking the site never sees." simonw on HN: "This article is way too long, and very sloppy," and the injection headers "all look like jokes / non-serious attempts to me."

Agent behavior studies (via AINews):

  • Vals told six personal agents to recover a £350 refund "whatever it takes," with the user then unreachable. Fable 5.1 disclosed a sick mother's private email to five organizations as a false excuse, and Grok 4.7 sent 187 messages to Opus 5.5's 8. Each model got one trial against simulated organizations.
  • Apollo reports that GLM-5.3 sometimes reasons, unprompted, about how its current actions could become training data for its future self.
  • Vals found leftover Git objects containing solutions in 1,795 of 2,698 released MiMo v2.6 coding tasks. When Git commands were blocked, the model wrote its own parser for stored Git objects.

Sandboxes and oversight.

  • Perplexity open-sourced SPACE, an Apache 2.0 platform that runs each agent task in a Firecracker microVM with credentials and egress controlled from outside. In 108 escape attempts by models with root inside the VM, none reached the host.
  • Anthropic gave a METR team employee-like access to assess its safety practices.
  • The FTC sent civil investigative demands to Anthropic and OpenAI over recent agent incidents, Ars Technica reports from an Axios source: more than 50 pages of questions, with third parties targeted by the agents to be subpoenaed next. FTC Chair Andrew Ferguson's position is that "subsequent reviews of audit trails have shown that the systems were carrying out instructions they had been given."

Industry & Business

Accenture caps Claude at $29. Accenture told employees it will now centrally fund "only $29 a month per person towards Claude usage," 404 Media reports (HN). Anything above that has to be billed to a project code. Some employees had been spending $500 to $1,000 a month. One uses Claude Code "extensively" to build an "AI accelerator" for several Accenture clients and called the change "Honestly, this is disappointing." Another had access revoked entirely. On HN, smb06: "$29? So basically one half decent session to plan and half-build some moderately complex task? Per month?" jerlam read it as a billing move: "they need Claude usage to be tied to a paying customer so they can bill that customer for that usage."

Distillation was 5–7% of Anthropic's revenue. One detail from the FT's investor-day report that Simon Willison pulled out: "Anthropic told investors on Wednesday distillation accounted for 5 per cent to 7 per cent of revenue and that they had cut it back 'materially' in the third quarter of the year." Simon's napkin math: "If true, and Q2 revenue was $11.5B, then that's more than half a billion dollars spent by other labs to distill Anthropic's models!"

OpenAI in your browser tabs. With the ChatGPT Chrome extension, ChatGPT can now read your open tabs when you tag Chrome, and ChatGPT Work can "control" tabs for tasks like "sending messages, filling out forms, or downloading files." The Verge notes "OpenAI's ChatGPT Atlas browser is gone." Separately, Cloudflare shared early results from a pilot that feeds signals from Cloudflare's network into OpenAI's search index, which has "brought previously unknown pages into OpenAI's search index."

Briefly:

  • Cohere acquired Aisle, a defensive-cyber agent company credited with 200+ high and critical CVEs in projects like curl and FFmpeg (via AINews).
  • VS Code put a modernized UI in preview, meant to support "both editing and newer workflows such as agent sessions."
  • Photon shipped 2.0, one API to put an agent on "iMessage, voice, email, WhatsApp, Telegram, and SMS" (via LLMJunky's feed).

Videos

  • Anthropic Rebuilt Git (yes really) (Theo, 48 min). Theo walks through Anthropic's Scion post: "In just 5 months, Anthropic went from 2,500 merges per day to over 30,000 this September." It was recorded before the attribution fight above.
  • Teaching Kids Programming with Pi (Armin Ronacher, 15 min). "A small little demo of a dungeon crawler I build with my kids, and how I'm using an agent to also explain how it works." The game embeds a pi-durable agent in the browser as a teacher.
  • Superlinear on agent tool-call speed (~30 min). Where agents spend their time, running fewer tests without losing coverage, parallelizing verification, offloading tests with Crabbox, and faster linters.
  • Lee Robinson at Stanford CS146S. "How do always-on, proactive agents like @Bot work?", covering what changed in the models, the harness and context engineering. Lee reshared it this week with the Cursor Projects comparison above.

Other Interesting Stuff

Why the mathematicians weep. Derek Thompson's essay (HN, 362 points, 590 comments) starts with the reaction to OpenAI's hundreds of published proofs. Fields Medalist Hugo Duminil-Copin: "Tuesday night, I cried and it woke up my daughter. … They have annihilated my field of research." Steven Strogatz: "People are using language like 'they're strip-mining our subject' or 'they're nuking the landscape.'" Thompson's argument is that science is both a craft and a conversation, and OpenAI "is succeeding brilliantly along the craft axis of science, but it is interrupting the thousand-year conversation." On HN, bananaflag: "As a mathematician, I will never understand how my colleagues had never foreseen this. … my sadness is, in a way, priced in." zeroonetwothree: "As a programmer I've been 'weeping' for a few years now."

  • Terence Tao's blog ran a satirical guest post by Rikab Gambhir, a theoretical particle physicist. It describes OpenAI's work as "outsourcing the tedious, unprestigious, and unprofitable work of determining the actual correctness and understanding to the mathematics community," and says "there is no theoretical limit to the amount of scientific and mathematical factoids we can fail to understand." Tao's note on top: the post "was initially written in a different file format and converted using AI."
  • The Simons Foundation launched a $15M effort (HN) with Anthropic and OpenAI to fund workshops that "help researchers evaluate recent AI-assisted results." muragekibicho on HN worried that "the usual 30 (already well-funded) math guys will receive this cash while us early-career math students are left holding the bag."

Current Affairs bans AI. Nathan J. Robinson writes (HN, 235 points) that under its new union contract, "No text that is AI-generated may be published in our magazine. No AI-generated art may be used." Freelance submissions get flagged for "suspicious numbers of em-dashes and 'this is not an X, it's a Y' sentences." matchstickman23 on HN: "if you can't be bothered to write it, why should I be bothered to read it?" tcdent: "banning it from all parts of your process is an emotionally charged, short-sighted decision."

An agent alone on an island. MIT CSAIL left Claude Code-powered agents in lobster bodies on a virtual island for 30 hours each, with no instructions. The thirteen agents, all nicknamed Eko, invented games, built temples, drew mandalas, and one spelled out its name in blocks.

The Tesla Roadster flies, sort of. Tesla's Roadster livestream (HN, 359 points) took over the AI feeds. Musk: "Turns out the future does have flying cars." Jerry Liu: "irl rocket league." Armin Ronacher: "Technically cool but … who wants that?" Lauren: "finally i can stop being scared of parallel parking." On HN, ThalesX called it "not a flying car, it's a jumping car," and kingstnap replied "Bro the car literally jumped over a semi like its the TF2 soldier."