Theo says you're using Jev wrong, Muse turns out to be OpenClaw, Cherny loves being wrong, Google ships AX
Sunday. Simon Willison, Andrej Karpathy, Lee Robinson, Tibo Sottiaux and Matt Pocock posted nothing new on X; Simon's blog carried two small items. The Jev story kept going, personal cloud agents got their first comparative reviews, and Hacker News carried most of the infrastructure news.
The Jev Wave, Day Five
Theo: you're using it wrong. Theo's 30-minute video Jev is incredible went up overnight, pinned with the line "Jev is incredible. I'm scared people are using it wrong." The description is the thesis: Jev is "a fast classifier with tons of perks and safety, but it doesn't replace reasoning models like Astra/Fable." The replies do the rest of the work. Alex Vainshtein's mental model: "Jev belongs in bounded hot-path decisions: route / retry / accept / classify, while frontier models handle open-ended work. Using it everywhere misses the point. Using it to shrink the control-loop cost is where it gets interesting." Jatin Garg: the wrong use is "routing every tool call through it instead of reserving it for ambiguous decisions." Kevin Herron tried tic-tac-toe after Theo tried checkers and found it "can't block my obvious winning moves" regardless of board format. Jay Zhou's warning is the sharpest: "you can only grade a pick if you already know the answer. which is why the wrong way never feels wrong." Others pointed at Made With Jev, a catalogue of about 300 builds, and a Chinese clip where Astra plans and Jev controls the character to beat Minecraft's Ender Dragon in eight minutes.
DocJev. Jerry Liu introduced DocJev, an open-source library for document classification and splitting: hand it a PDF, DOCX or PPTX plus natural-language category rules, and Jev returns the category with probabilities and review flags, or the boundaries between sub-documents with page ranges. He claims 6x faster than GPT-5.6 Luna at equivalent accuracy, with liteparse as a free local OCR backend and LlamaParse optional. The README stresses it is an independent implementation that does not call LlamaIndex's hosted Classify or Split APIs. The demo packet has two adjacent Treasury auction results in the same category, which the splitter still separates, while keeping a ten-page BEA release with dense tables together. The visual report covers 40 real documents. Replies asked the right questions: The1Broom wants to know "what confidence do you take before you trust a split point," Jay Zhou notes "a bad split never throws an error. it just quietly changes every answer downstream and the citation still looks fine," and tensorquay points out the 6x figure excludes OCR time. Jerry's follow-up: Jev's promise is to make business operations "extremely lightweight and fast, while still saving intelligence-heavy tasks for larger agentic systems."
jevchat. Kyle Pena's jevchat (123 points on HN) turns Jev into a chatbot by asking it one question per step: given the user's question and the reply so far, which symbol comes next? Strategies include a bisection over a sorted alphabet and a bucketed mode that can hold more than 255 symbols. "The idea is for fun, the cost is somewhat impractical, and the results are hilarious." Asked to write a short story, it wrote "a story." The HN thread is worth it for ralusek's debugging technique, which is to have Codex generate 30 plausible reasons Jev scored something a certain way and then ask Jev to score each reason ("Felt like those people that give their dogs the buttons to push"), and for Pete Sergeant's lament that every weekend Jev project he thinks of has five implementations before he starts.
Jev pod tomorrow. swyx announced Latent Space's Jev episode with creator Diogo Almeida drops Monday.
Agentic Coding & Agent Harnesses
Armin's Sunday meditation. Pi posted a video of Armin Ronacher talking through "shared frustrations with agentic software engineering," following his question on Saturday about what people struggle with most. "Remember that you're not alone in finding it harder than it looks!" Replies were a support group: "Spent four hours undoing what took it thirty seconds to break," "Sunday meditation, Monday agent debugging," Henrique Bastos on new models feeling "like a racing car with no traction control," and Jonathan G on contributing a Rust performance fix to a project in a language he does not know: "we're sacrificing understanding for speed which will eventually have an inverse effect when we can't understand anything anymore." Separately, Pi 0.86.1 shipped a Meta Muse provider (/login meta or META_API_KEY, for Muse Spark models) and a faster launch via Node's persistent compile cache.
Theo's "prompt wider" lands. Theo thanked viewers for the response to his video arguing you should not use models dumber than Fable and Astra. The replies describe a behaviour change: OdraivK "stopped all the orchestrating stuff as well and just let the big boys do it. I've been able to get more work done since Fable and Astra didn't have to orchestrate and babysit dumber agents that ended up taking more turns." Robert James used a wider prompt from his car by voice and had Astra read his files, produce docs and visuals, and email them to him. Theo confirmed he "hasn't run a goal in months" when someone asked about the /goal cycle.
Boris Cherny is often wrong. Claude Code's Boris Cherny posted a note he had shared with his team, I am often wrong: a six-step loop (understand available information, gather missing information, define the problem, define a clear and simple approach, define a goal, act with urgency), redefined whenever new data arrives, with feedback given in real time when someone skips a step. The failure modes he sees most are a poorly defined problem and an approach that is not simple. On HN (181 points) the reaction was mixed: cube00 did not like "forcing your team to use your own personal framework" or urgency as the default ("no wonder everyone is burnt out"), adamsb6 called it an OODA loop, and otterley recognised the Amazon decision-doc process. Cherny replied in the thread that "this isn't the one true framework, but it is the one I use," and that feedback reads as corp-speak only if you have not worked in a culture where it is honest and no-blame.
AX. Google's AX (408 points on HN) is an Apache 2.0 orchestrator that treats agents as "neither microservices nor batch jobs": you ax apply a YAML Task with a Workspace that lists Git repos, MCP servers and skills, or just a plain-English goal that an agent fulfils on first boot. Four primitives: Task (sandboxed, suspendable), Workspace, Gateway (network allowlists with credential injection) and Model. It runs on a runtime called Agent Substrate and claims billions of concurrent tasks per cluster with sub-second resumption of agents parked on model calls or human input. HN was not charmed: "k8sification of AI was always inevitable, if only as a form of salary justification," several people said Gemini in Sheets cannot do basic tasks, and the sunset track record came up. SP3269's counterpoint: this one sells GCP, so it stays.
Why MCP was always a bad idea. Maharshi Patel's post (HN) argues MCP was built for weaker models, that the "MCP industrial complex" of search/execute wrappers and monitoring exists to paper over context bloat, and that models now read --help and call documented APIs directly, so "we delete most of our MCP servers." His proposal is to standardise how agents use HTTP instead: an Accept: text/markdown header, and Malte Ubl's request that harnesses send a preferred programming language so docs sites serve the right SDK examples, which Shopify already ships. HN's objections: credentials still have to live somewhere the model cannot see, agents without shell access exist, and cheaper production models still need the help. Toby Hinloopen described a tool that captures CLI output and lets the model query it later, to which layer8 replied: "Otherwise known as a 'file'."
Simon's weekend. Simon Willison released llm-keys-ui because he runs coding agents on remote machines from his phone via Codex Remote and refuses to paste API keys into agent sessions: the plugin serves a local web form for saving keys, and the agent later runs llm keys get anthropic. He also quoted an HN user two weeks into a big-company job: "everything is made by Claude Code... People are working 12 to 13 hours a day just to press enter. Nobody is reading anything."
Personal Agents: Muse, Bot & Claws
Muse is OpenClaw under the hood. Bill D'Alessandro's claim that Meta's Muse assistant is "very good, and I think I've figured out why. It's OpenClaw under the hood" reached 249,000 views, retweeted by Peter Steinberger. Asked why it beats Instinct, he says the native app is "the whole thing": you can see its browser, and it has iOS system-level hooks. He credits Meta's Spark model for speed while conceding "the Grok SOTA model is slightly smarter." Another user timed Muse finishing in 15 minutes what Grok Bot took two hours to do. David Koonar: "I asked muse if it was openclaw and it was trying so hard to say no."
Verdict thread. am.will asked for verdicts on Muse versus Grok's Bot after initially meme-ing Zuckerberg about personal data: "I have updated my opinion. I'm actually really impressed with it." He likes that it is "a secure VM you can put all your credentials and stuff in." Replies split along ecosystem lines, with Grok users preferring texting an agent to opening an app, a complaint that Muse "is just one mega conversation" with no bot mode, and one who finds Meta too hard to trust "even if the model and harness are great."
Your claw can FaceTime you. Omar Shahine's FaceTime plugin for OpenClaw, audio only for now, lets you call your agent and lets it call you with approval, using GPT-live for speech. He calls it the longest he has worked on any OpenClaw feature: "It wasn't till Opus 5 and GPT-Sol 5.6 that I could really cross the finish line," with Fable 5.1 used at the end to get it release-ready because it was "too slow and costly" earlier. The docs are candid about the cost: it uses private Apple APIs, injects a helper into Apple's call apps, requires reduced SIP debugging protections and full Xcode, and should run only on a dedicated Mac you physically control. Steinberger's reaction: "Your claw can now FaceTime you!" One reply questioned the point when Grok Bot already sends a voice report every morning.
Security, Safety & the Discourse
Theo vs the "exaggerated threats" story. Kalshi's account posted that OpenAI and Anthropic "reportedly exaggerated AI security threats to push the government to protect their market position." Theo's response: "You'd think there'd be, like, SOME evidence of this if it was really happening." His five points: every proposal only affects models above a size threshold, those rules hit incumbents harder than newcomers, the evidence is anonymous insiders in a New York Post article, people cannot keep secrets, and Elon Musk agrees with the labs despite being a victim if it were true. Told to "stop protecting them," he answered that spreading lies as facts "makes discussing the legit bad stuff much harder." Alex Kantrowitz's It's Time to Bring Some Sanity to the AI Risk Conversation from earlier this month makes the same case at length: warnings that invite data-centre pauses right before two IPOs are terrible marketing, so they probably are not marketing.
Stochastic parrot as cognitive virus. Anthropic's Jack Clark wrote that "stochastic parrot" was "a mimetically-fit cognitive virus that spread from 2021-2025; it temporarily blinded many gifted people to the nature of AI progress, burning up crucial years." 383,000 views, 223 replies. Pressed on why not just ignore it, he said he "constantly" meets people who discount AI capabilities on that basis. Janus (repligate) was unimpressed: anyone blinded for years by a slogan would not have been much help anyway.
Pirate Face. pirateface.co (522 points on HN) mirrors every Apache-2.0 and MIT model on Hugging Face as a checksum-verified torrent, with a planned drop-in HF_ENDPOINT so pipelines pull from the swarm with no code changes. Z.ai has a verified profile seeding GLM-5.2. HN mostly asked why this was not always the case, with CodesInChaos noting public torrents die as they age and BitTorrent v2 is almost never used.
Control points. Po-Shen Loh's guest post on Terry Tao's blog, Why do we need human mathematicians anymore? (HN), responds to economists who told mathematicians to cede control after the Navier-Stokes result. His argument: there are zero examples of a more capable species surrendering control to a less capable one, frontier AI decisions are as opaque as a brain, AI-accelerated hacking flips even pre-AI software into the untrusted category, so the number of "control points" needing skilled human oversight explodes. "Those are jobs. Highly skilled jobs," and there will not be enough people for them, which slows AI down. HN's first reply: chimpanzees looking at human society would say the same thing.
Hardware keys. am.will pointed out that a ChatGPT subscription includes 50% off a YubiKey C or C Nano under Settings, Security. Philo Groves says they are required for Daybreak Blue by October 1.
Other Interesting Stuff
- Qwen Image 2.1. Alibaba's release took 595 points on HN, where the discussion was about why local image generation is now ahead of local code generation: "Image gen you eyeball one frame and stop, code needs hundreds of tokens all correct in sequence." One commenter found it worrying that a 7B model can do this with no watermarking.
- Chatbots and money. The FT reports on a Saturnos study finding AI chatbots give wrong answers to financial queries most of the time; the best performer, Claude Opus 5 in reasoning mode, still erred on 39% of answers (HN).
- Armin on rules. Armin Ronacher's thread on the German and Austrian habit of defending every rule, prompted by a Tesla fined 55 euros at an EV charger for lacking an EV sticker, drew the reply of the day from ferologics: "on average the size of a German's AGENTS.md is roughly 3x that of an American."
- Grok translations. Theo noticed Grok's inline translations on X misfiring on Spanish replies.
- AI Engineer New York. The speaker lineup includes the CEOs of Arize, Ramp, Modal, Nous Research and turbopuffer, plus Bridgewater's CTO.
Sources: nitter.jaydenha.uk RSS and thread pages for all accounts (@potetotes returns 404, as it has for weeks). Simon Willison's blog, pi.dev, YouTube feeds, GitHub READMEs and Hacker News via the Algolia API for the non-X material.