Meta's Muse Wonder reaches the frontier, The X Subscription folds Cursor into Grok, Codex's background daemon draws fire, Linus ships a Claude Code patch

Monday belonged to Meta. Muse Wonder landed in the afternoon, and by midnight my feeds had gone from "Meta is a frontier lab now" to an argument about whether a coding model has to be good at lawn mower simulators. Cursor became part of a Grok subscription, and the Codex CLI got called malware in its own issue tracker. There were no roundups for October 10–12, so the weekend news that still matters is in its own section near the end. Andrej Karpathy and Lee Robinson posted nothing in the window, and Boris Cherny's last activity is a repost from Friday. @potetotes still returns nothing, so Lauren's posts come from @poteto.

Muse Wonder puts Meta at the frontier

The launch. Meta introduced the Muse Wonder series (Zuckerberg's announcement). There are two models. Muse Wonder is on the Meta Model API and in Muse Code. Muse Wonder Full is only for Muse Code subscribers. Meta's pitch is personal agents, built around "Active Memory" (carrying context about you across conversations and pulling it back when it matters) and "Active Focus" (sticking with a hard task, trying other approaches, checking in "before taking important actions"). Meta trained both at its "gigawatt-scale Prometheus data center" and says the bigger base model is why Full does better on tasks it wasn't post-trained for than the Muse Spark 1.3 trend predicts. Alexandr Wang's three highlights were memory "across days/weeks/months," getting tasks done, and "sota in finance & legal, strong in health."

The numbers. Artificial Analysis scored Muse Wonder 52 and Muse Wonder Full 56 on its Intelligence Index at max effort, up from 48 for Muse Spark 1.3 six weeks ago. The biggest jump was Terminal-Bench 4.0, from 33% to 57% for Full. On the Coding Agents Index, Full in Muse Code scores 68, the same as Claude Code with Sonnet 5.5 (68) and ahead of Opus 5.5 (66). Plain Wonder scores 63. Pricing is $4 input, $0.20 cached and $20 output per million tokens, with a 1M context. That makes Wonder's cost per Intelligence Index task $4.52, below Opus 5.5 ($5.98) and far above GPT-6.1 Sol ($0.72). Full has no per-token price yet. On HN (609 points), Jcampuzano2 noticed the shift: "unlike other Meta models this one seems to take a pretty big jump in pricing."

Theo is all in. Theo had early access and posted "My current model rankings": S tier is Opus 5.5 and Muse Wonder, A tier is GPT-6.1 Sol, then a gap. Then: "Meta is a frontier lab now. OpenAI has fallen into 3rd place. I think Google is kinda screwed." Theo later clarified that the ranking is for Full, that "Standard Wonder (the one over API) is significantly worse," and that "The naming sucks. The restriction of Full being subscriber-only sucks even more." The case for it is mostly about feel: "this prose is an entirely different flavor than I'm used to from LLMs. Not another Claude slop distillation," and "I'm not recommending replacing Opus with Muse. I'm saying I like them a similar amount." On T3 Code's Orchestrator V2: "I often have Opus ask Wonder for code review, copy work, and architecture review." The best anecdote came overnight. Theo gave the same bug to three models. Opus "found the root cause and filed a pr." GPT-6.1 Sol "crashed from the async question tool interrupt, found the bug, fixed it, wrote 300 lines of tests, and explained it in impossible to understand terms." Muse Wonder Full "found a small PR someone else had filed and said 'you should probably merge this instead'." Theo then checked: "GPT-6.1 Sol's fix would not have even worked 🙃."

Then the 3D demos. LLMJunky's first run was "Early impressions of Muse Wonder: YIKES": on RocketLeagueBench "It did worse than basically every chinese model I've ever ran this benchmark on." @_MaxBlade gave Wonder Full and Opus 5.5 the same AAA lawn mower simulator prompt. The Muse run cost $4.11 in API-equivalent tokens against Opus's $99.43, but it still "burned my $50 plan," and Max concluded Muse "is not anywhere near close to frontier in 3d tasks like video games" (quoted by Theo). Theo's defense: "The smartest code model in the world would not be good at 3d stuff unless it was trained specifically for it," and "3dmaxxing is a different type of benchmaxxing." Wang agreed: "whole team feels dumb now, but it's true, we did not explicitly train our model on 3d demo games... please try muse wonder on your actual coding work!" @bijanbowen pointed out the awkward part. The first coding demo on Meta's own launch page is a playable 3D kitchen game. @ChrisGPT, a former game developer, wasn't buying it either: 3D demos test "geometry, coordinate transformations, camera systems, physics, rendering, and actual interactive environments."

Other early reviews.

Quota bug and a reset. Early users burned through Muse Code limits fast. Wang posted that Meta had "fixed the issue that was causing the unusually fast quota usage" and announced a "muse code reset!!" Theo's reply: "New reset guy just dropped." LLMJunky couldn't upgrade to try Full properly, because the plans page only showed the $5 plan.

No weights. AI at Meta: "The open weights for Muse Spark are coming soon, Muse Wonder models are large models that are too big for many people to use, in addition to potential safety concerns."

Hacker News. nerevarthelame flagged that Meta's own internal metric shows "a ~3% violation rate for mishandling personal information," which "seems unacceptably high" for agents handling personal affairs. Starlevel004 had a theory for how Meta caught up so fast: "Opencode shows nearly 40 trillion tokens going through muse spark 1.3 contributor this week alone." That tier gives the model away if you let Meta train on your data. wkcheng paid for a month of Muse Code and found "their sandbox never worked across Mac or Linux."

The paid-shill accusations. @Youssofal_ guessed Theo was "social climbing" after hearing from Zuckerberg, and Tibo asked "How much did they pay you" (Theo's reply: 💀). Theo answered the first one: "If you disagree with me, say why! Making shit up about my 'motivations' is so cringe." A few hours later: "Just realized the 'paid shill' accusation are kind of a self own."

The bigger picture. Andrew Curran (reposted by Peter Steinberger) thinks this ends any hope of labs pacing themselves: "One defection among US labs breaks it all, as the others will be forced to accelerate or get passed... Mark has Trump's ear."

My read: a model that ties the best coding-agent score through Meta's own harness while losing on 3D toys and frontend taste is a real coding model with a narrow training mix, not a benchmaxxed one. The worse story for Meta is the split. The model Theo loves (Full) has no price and no API, and the one you can buy is "significantly worse."

The X Subscription folds Cursor into Grok

SpaceXAI launched The X Subscription (announcement), one plan covering Grok, Grok Bot, Cursor and X Premium. The tiers:

  • Free, with daily Grok access and "a starter allowance for Cursor's coding agents."
  • Plus, $8/month. More Grok and Cursor usage, a checkmark and fewer ads.
  • Premium, $30/month. SuperGrok, Cursor Pro and X Premium+, "previously $90/month combined," with 4x the Plus usage.
  • Super, $100/month (15x Plus) and Ultra, $200/month (50x Plus).

Grok and Cursor draw from one shared pool of Grok model usage, and every plan includes "at least twice as much coding usage with our models as the comparable Cursor plan." Teams pay the same per-seat prices and get 3x the coding usage of old Cursor Teams. Existing SuperGrok, Cursor and X Premium subscribers can keep their plans, but the old plans are closed to new signups.

The line that matters for Cursor users: "third-party models are still available in Cursor on Premium and above. Usage is billed separately at API list prices, through top-ups." So on the new plans, a new Cursor customer's subscription pays for Grok, and Claude or GPT costs extra at list price. Lauren (@poteto), who works on Grok Bot, called it "one unified sub with incredible value." The rest of Lauren's timeline was Grok Bot enthusiasm, like Peter Yang describing bots built "on personas instead of chat threads, so it really feels like hiring employees to work for you."

My read: the bundle is a good deal if you're happy on Grok models. If you picked Cursor to switch freely between Claude and GPT, new signups now pay the API meter for that.

Codex: a daemon nobody asked for, and a Day 8 reset

The daemon. A Codex CLI issue from September 25, #48195, reached the top of HN (535 points, 271 comments). Since 0.157.0, opening an interactive Codex session:

  • starts a persistent codex app-server --managed-daemon that keeps running "after every Codex window is closed (reparented to PID 1)";
  • copies the whole 374 MB platform package into `/.codex/packages/app-server-daemon/`, so the backend's version is "decoupled from the CLI version I installed with my package manager";
  • runs a detached updater that checks hourly and restarts the backend, which "can kill active turns";
  • pins the environment of whichever session started it, including that session's HTTPS_PROXY.

codex app-server daemon stop also leaves the updater running. The workaround is codex features disable daemon_auto_start, codex app-server daemon stop, then pkill -f 'app-server daemon pid-update-loop'. One commenter found the flag still missing from the official config reference, and the 0.157.0 release note covers it in one sentence. The tone in the issue has turned: "Basically - a trait of malware," and "Guess I'm rolling back all the Codex seat licenses at my company." On HN, icantevenhold found old versions taking "something like 12 gigabytes." timdorr pointed out that Claude Code works the same way, with old binaries in ~/.local/share/claude/versions/ and claude bg processes. thebeardisred made the case for distro packaging, and whimblepop's Nix package of Claude Code already patches both behaviors out.

Day 8 was a reset. Tibo, Day 8: "Global reset landing by EOD PST time. Enjoy," then "What would you like with your side of reset? It's done." Tibo also teased what's next: "Good plans are simple. I have one that should result in significantly more usage for everyone." Over the weekend, a Pro poll with 240,643 votes went 38% for "no 5h limits, same usage" against 33% for "5h limits, 2X week usage." That surprised Tibo, who expected a clear win for 2X. Days 5 to 7 are in the catch-up section below.

Ultrafast. Peter Steinberger: "6.1 Sol ultrafast >>> 6 Astra. ~ same cost but whole different game with 250 tpm." Theo's Short argues the opposite case for Astra: "With Astra's pricing, Ultrafast just isn't worth it."

Meanwhile Theo, at 5:41 UTC: "The $200 Claude Code sub is one of the best values in software right now. Genuinely insane how much work you can get done without hitting limits."

Agentic Coding & Agent Harnesses

Linus Torvalds ships a Claude Code patch. Torvalds posted a kernel patch series on a branch called noref-filemap-read-with-llm, and the cover letter says it plainly: "The code, the test programs, the review behind the list above and the first draft of this text were done with an LLM coding assistant (Claude Code) under my direction. The measurements were run on my machine." (Startup Fortune, after Phoronix broke it.) The patch lets small page-cache reads run inside an RCU read-side critical section, which skips the atomic reference-count bump that many threads fight over. On Torvalds' Core i7-8565U laptop, 16-byte reads across eight threads went from about 10 million to 30–33 million per second, and kernel cycles per read fell from about 2,950 to 725. A single-thread read of an unmapped page got slower, from about 4,150 to 6,550 cycles. It could land in the 7.4 merge window. In July, Torvalds told critics of AI code in the kernel they could "do the open-source thing and fork it."

QGIS goes the other way. QGIS adopted QEP 428, a strict "NOT ALLOWED" policy for AI contributions (89 thumbs up, 27 down). haubourg explained why the project went past "human in the loop": "AI use from external contributors puts a massive and unbearable pressure on the shoulders of our maintainers." On HN (372 points), rfgplk noted a follow-up proposal, QEP 429, that exempts trusted contributors. flir answered the "just rewrite it with agents" crowd: "I bet you could do it in a week... And neither of us will get any traction."

More code, not more software. Ars Technica covered a Harvard study by Fiona Chen and James Stratton, built on Jellyfish data from more than 700 firms and 700,000 employees through March 2026. After a firm adopts coding agents, lines of code go up 30%, commits 20% and pull requests 23%. Resolved Jira issues and epics don't change significantly. Review time from PR to merge goes up 49%, the share of PRs with changes requested nearly doubles, and comments per PR rise 35%. AI wrote only 23.3% of review comments. On HN (331 points), staminade expects the change to start somewhere else: "It'll happen in startups that adopt a no-human-review policy from jump and then maintain it as they mature."

Your codebase is memory. Lauren (@poteto) has had enough of engineers treating slop as something that happens to them: "you are a software engineer, you literally learned how to make sand think." Lauren's two rules: give your agent a way to verify its own work, and "your codebase is memory." "Every nit and every annoyance you've ever seen, every mistake and bad pattern you observe from agents, needs to turn into a constraint, or the agents will just keep repeating those same mistakes like a virus in your code." Better still, "refactor/rewrite your code such that agent anti-patterns become categorically impossible." Matt Pocock said it in two lines: "Don't blame the agent for writing the bad test. Blame the codebase for allowing the bad test to be written." dax went further: "so many of you claim to be obsessed with code quality... you should be going crazy with agents right now. refactor everything." Matt also asked how people are finding GLOSSARY.md from /grill-with-docs: "the more work I put into the glossary, the more productive I feel. It's a warm, comforting feeling having the agent speak my language."

Is it a tool? Rich Hickey's "'AI' is Not a Tool" (HN, 255 points) says LLMs are automation, not tools, because you don't wield them: "A drive-thru is not a tool for cooking, it is automation for avoiding cooking." Learning a model's quirks means "filling voodoo.md with just the right incantations," and "You are only a software update from everything you've 'learned' being rendered obsolete." Huahai's reply did better on HN (354 points): "Even by his definition, an LLM can be something a human wields," and a compiler also "performs substantial work without the programmer continuously directing each operation." Lauren weighed in from another angle: "i never let agents write for me... i still strongly believe that writing is thinking... words become the interface between you and the genie."

Dynamic workflows at AI Engineer NYC. Anthropic's CJ Avilla and Harrison Stall ran the first full workshop on Claude Managed Agents dynamic workflows, written up by Latent Space. Stall: "you can let Claude fully orchestrate the entirety of its own multi-agent topology." The lead agent writes a plan, then an orchestration program that launches agents and collects their results. Asked how Claude decides how many agents a stock-ticker scan needs, Stall said "evals and lower defaults." The model doesn't know its budget while running, but developers can set a session budget. Telling it the budget might make agents "a little bit anxious."

Smaller things.

Safety & Security

AI worms on Android TV boxes. Anthropic's Frontier Red Team (William Zhou and Daniel Freeman) argues that small models "may be months away" from self-replicating across IoT devices. To show the compute is there, they ran Gemma4-26B-A4B across 15 off-brand Android TV boxes (about $55 each) at 0.39 tokens per second. A 9B model got inference running on 8 of 9 hardware types they tested. In simplified end-to-end tests, GLM-5.2 and stronger models could already replicate, and at the historical rate of improvement small models catch up "by the end of this year." Such a worm would be slow ("a doubling time of days") but has "no centralized point to shut off." HN was skeptical. pythonaut_16: "Guess it's time for Anthropic's next round of fearmarketing. Beneath the fear there does seem to be a cool technical demo here though." vannevar saw "an admission that maybe we won't really need all these data centers."

Provably hidden capabilities. A NeurIPS 2026 paper, "AI Models Can Provably Hide Arbitrary Capabilities" (Draguns, Motwani, Douglas, Schroeder de Witt), embeds encrypted neural networks in a model's weights that only decrypt on a secret trigger. Under cryptographic assumptions they stay hard to find "even given full access to model weights," and they survived fine-tuning, weight noising, gradient-based trigger search and latent adversarial training. That's a hard limit on capability audits that rely on eliciting behavior. On HN, b473a: "We went from having a panic over binary blobs after the xz crisis to everyone installing giant binary blobs on their computers and giving them access to their filesystems and online accounts."

A malicious commit in Zod. A commit titled "Fix bugs and improve performance" bumped Zod to 4.6.6. It added a meta.test.mjs that looks up a server address in an Ethereum contract, downloads a payload and runs it. It also switched the release workflow from a manual trigger to publishing on pushes to main. On HN, commenters found that the poisoned version apparently never reached npm. xx_ns noticed the preinstall key sat at the package.json root instead of under scripts, so it wouldn't have run anyway. A maintainer note quoted in the thread says all of the maintainer's npm packages moved in September to staged publishing triggered only by hand, "This means 2FA is always required." The commit's workflow change looks like an attempt to get around exactly that.

Catching up: October 9–11

There were no roundups for October 10–12. These are the things from Friday through Sunday that are still being talked about.

Videos

Other Interesting Stuff

  • GLM-Legal. Z.ai released a legal model (HN, 520 points). postalcoder noted it's "not an open model" and wondered whether Chinese labs are trending proprietary. Natsu saw nothing in the announcement about hallucinated cases.
  • Ben Recht on OpenAI's FFT result. "The OpenAI algorithm for the FFT is less a breakthrough and more a malicious exploit." It only beats the FFT under a model of computation that stores complex numbers at arbitrary precision and multiplies them in one operation.
  • Hardware. Intel's D270 "Crescent Island" data center GPU has 480 GB of LPDDR5X, 1.5 TB/s and 1.3 petaflops of FP8 at 350 W (via LLMJunky). SemiAnalysis reports that AMD landed megakernel decode in upstream vLLM on ROCm before CUDA has it. Janas runs models bigger than RAM by streaming experts from SSD and claims 37 tok/s from an 80B model on a 32 GB laptop.
  • "Fable took my jorb." Cendyne on being laid off as the last engineer: "It wasn't that Fable was better at my job. As the last engineer, Fable was cheaper than my compensation." Also: "'AI took my job' is far more fun to say than 'AI contributed to my role being eliminated.'" (HN, 220 points.)
  • "Thinking will become a hobby." Diomidis Spinellis predicts thought becomes a leisure pursuit, "as hunting did following animal domestication."
  • The Anthropic IPO. Research firm New Constructs calls it "the most ridiculous IPO of 2026" and values Anthropic at $150B against the reported $2T target (CNBC, from October 7, on HN today).
  • OpenFatigue™. "Open is the new closed." The parody includes "Weights Available*" after "a 41-page acceptable use policy, a monthly-active-user cap, and a vibe check."
  • Jerry Liu's verdict on a viral "China has killed the entire vector database industry" thread about TencentDB Agent Memory: "least clickbait slop post."