Grok 4.7 lands with a thud, Xiaomi's MiMo 2.6 ships its RL stack, Anthropic counts its own automation, Nat Friedman credits OpenClaw
A dense Monday. Boris Cherny and Andrej Karpathy posted nothing from their own accounts, though Karpathy showed up in a reply; Lee Robinson only retweeted the Grok launch; Tibo Sottiaux teased "a reset for Tuesday. Among some other things."
Two Launches: Grok 4.7 & MiMo-V2.6
Grok 4.7. SpaceXAI announced Grok 4.7 as "a notable improvement over Grok 4.6 at the same price and speed," 11 million views on the launch post. The release notes describe a new, larger base model with a longer reinforcement-learning run weighted toward multi-hour tasks, native training on the Grok Bot harness, and "our best-calibrated safeguards to date." The table puts it at 46.3% on CursorBench 4.0 (Fable 5.1 Max: 51.8%), 71.0% on DeepSWE at high effort, 64.0% on EEBench where it beats Fable, and 38.0% on Terminal-Bench 4.0 against Fable's 57.9%. Pricing stays at $2 in, $6 out, with a fast variant at double both. It ships in Cursor, Grok Build and the API. Hacker News (544 points, thread) was split between Cursor users who like it and people who find Grok "the laziest and most dishonest of all the models"; moojacob noted the model is reportedly 40% larger at the same price and was delayed two weeks, and guessed the rumoured Opus 5.5 launch is why.
Theo's verdict. Theo, who says he had early access for several days, wrote the review of the day: Grok 4.5 was "an incredible model for the price," 4.6 a forgivable step toward slower and pricier, and 4.7 "much harder to forgive. They claimed it would be more token-efficient, and it's less by 30 to 80%. It scores worse than Grok 4.6 in various benchmarks. It's slower, it's less pleasant to use, and real-world costs come out to more than 2x above Grok 4.6, putting it over Astra's costs." He could not find a single bench where 4.7 beats 4.6 on token efficiency, despite Elon Musk's August promise of "even better token efficiency," and said he "cannot fathom that they shipped a model that is this bad at the frontend and design in 2026." His Fishslop run was "the worst I've seen this year," with the submarine and fish moving backwards on the first attempt. Cursor CEO Michael Truell pushed back in the pricing thread: "In prod, 4.7 uses ~5% more tokens than 4.6 for a median request and ~20-30% more for a p99 request. There's also a distribution shift of what users are willing to ask it." Theo's reply: then why are Frontier and Artificial Analysis seeing 50%-plus cost increases? Michael Novati's summary in the replies: "Grok 4.7 feels like the Opus-fication of Grok 4.6." am.will's angle: the launch ironically makes the SpaceX valuation look reasonable.
MiMo-V2.6. Xiaomi released MiMo-V2.6 Pro and Flash, two natively omnimodal open-weight models, with the claim that Pro matches Claude Opus 5 and GPT-5.6 Sol on most agent benchmarks and scores 46 on the Artificial Analysis Intelligence Index, the highest of any open model. The page was the top Hacker News story of the day (784 points, thread). The details that mattered to builders: MIT licence, 1.02T total and 42B active parameters, API pricing unchanged from V2.5 at roughly a twentieth to a sixtieth of the leading closed models, a 20x-faster UltraSpeed variant, and the promise to open-source the technical report, 7,000-plus RL task environments, the end-to-end RL framework and the mini-harnesses used in training. Latent Space's AINews collected the training numbers people were passing around: 130 hours, 75 billion tokens and about $2.6 million for the RL run, scaled on JAX and TPU. Theo tested it: "these seem very very legit. 2.6 Pro in particular is doing well in some pretty hard tasks," though several replies found Flash weak. HN liked that Xiaomi's charts show it behind DeepSeek where it is behind ("Finally a lab that doesn't cheat on the charts"), and rao-v pointed to the live RL dashboard Xiaomi ran during training as "an incredible learning and teaching tool." One HN reply drew the obvious line to the pacing debate: "maybe this is why Dario wants to slow down."
The plateau argument. Yuchen Jin claimed that frontier coding capability has plateaued since Opus 4.8 while open models close the gap at 10 to 50x lower cost, "why we're seeing a massive shift from closed models to OSS across enterprises." He conceded that coding is far from solved: "LLMs today get stuck on complex systems... You can't build that with one prompt." Tak reported that Sonnet is routing to a new version and that "the new Sonnet" one-shot a doodle FPS game, "better than grok 4.7."
The Jev Wave, Day Six
Simon's notes. Simon Willison published Jev introduces a new shape of LLM, siding with Maggie Appleton that "decision models" is a better name than System One. The useful parts: Jev charges only for input at $0.042 per million tokens, cheaper than GPT-5 Nano; "Noul" is short for Bernoulli; questions in one request are evaluated in parallel, so many questions cost about the same time as one; and he is using it for search reranking over BM25 candidates. His discomfort is that Jev is "a regression even further towards black box machine learning systems," returning a float with no explanation, so bias concerns "should be front and center" and evals matter more than ever. His experiment asking Jev whether each Bay Area city is a "Good city?" rated Cupertino top and East Palo Alto bottom: "Huh." In the replies, swyx noted that Diogo Almeida explicitly says System One models are different from decision models, "but have to wait to see the other models for why."
The Latent Space episode. swyx's Jev podcast with TypeSafe CEO Diogo Almeida is 2 hours 20 minutes (announcement, 209,000 views). The show notes are the best written source so far on why Jev exists: Almeida co-authored InstructGPT, traces three branches of RLHF, and argues that every innovation since, from function calling to reasoning, "felt like a hack on top of the string based, sequence to sequence prediction paradigm." Jev's training method is "Reinforcement Learning for Calibrated Decisions," unpublished, optimising for "epistemically honest probabilities on System One tasks" rather than human preference (which he says causes hallucination and sycophancy) or verifiable rewards (which "solves Navier Stokes but exacerbates jagged intelligence"). Other claims from the notes: TypeSafe rejects public benchmarks and refusals at the API layer, thinks of itself as a data lab rather than a model lab, and Almeida "wouldn't pre-train with $1 billion." His launch video is at about 40 million views. The notes also link his "Tyranny of the KV Cache" note as the follow-up for anyone thinking about Jev in coding agents. One reply worth keeping: "'Jev mathematically cannot hallucinate.' It cannot produce an invalid type. That's not the same thing."
Karpathy's framing. Will Depue wrote that Jev is "just a classifier but it's a zero shot classifier with frontier-ish intelligence. i'm surprised someone hadn't built it before. i wonder what other old ML ideas are also worth rescuing." Andrej Karpathy replied: "I see it as a point on the LLM pareto optimal curve in a regime that had a large revealed latent demand (no thinking, single token, low latency acceptable intelligence) that was under-invested into because of a race to higher intelligence." Almeida answered that this is "a very pretraining-centric way of thinking about it" versus trading off intelligence for a use case.
Kev. Jared Palmer's Kev (427 points on HN) is a family of 0.8B, 4B and 9B decision models on Qwen3.5, Apache-licensed, with training code and frozen eval suites. The API matches TypeSafe's, so their Python SDK can point at a local server, and the 4B and 9B fit a 32 GB Mac. The HN thread is the clearest discussion yet of where these fit in coding agents: vidarh suggested classifying bash tool calls as safe or unsafe, NitpickLawyer described splitting goals into actions plus verifiers with the classifier judging each step, and hbrn dissented that Jev is only useful when you need sub-second speed and can tolerate mediocrity, "and outside of fun demos these two rarely come together."
Jev in the wild. From the AINews recap: LangChain added Jev-as-a-judge to LangSmith, Harrison Chase and Jacob Lee shipped SemIf, an open decision model, through the LangSmith Gateway; Elvis Saravia retagged 2,300 papers in 83 seconds for $0.14; Sarah Drasner built a Chrome extension that uses Jev to pick WebMCP tools per keystroke. Jerry Liu extended the calibrated-confidence idea to LlamaParse's Extract: per-field confidence for bools, ints, floats and free text, each with a bounding box into the source, so a threshold decides what gets automated and what a human reviews. His DocJev now has an interactive splitting visualisation. His observation on the launch aesthetics: "everyone is astra to build product visualizations on X lol... i've seen all the architecture animations and benchmark graphs in different company brand colors."
Agentic Coding & Agent Harnesses
Fable's thinking budget. Lon Lundgren's thread (252,000 views, 382 points on HN) is a six-week measurement of Fable 5 thinking tokens: "After Anthropic made Fable 5 permanently available in subscription plans, I noticed a large drop in performance... Measured five different ways, August delivered dramatically fewer thinking tokens than July." He was on xhigh or max effort, yet "most invocations to the model were receiving little to no thinking tokens at all," and reasoning fluctuated in multi-day episodes that lined up with product announcements. His conclusion: "Next time the model feels dumber, don't ask if the model was 'nerfed'. Ask about the inference regime you were served." Asked what he would prefer, he said just tell him inference is being reduced so he can use another model or take the day off. He reports that "ultrathink" is the only keyword still affecting thinking and no longer gives a fixed budget. HN dug up Anthropic's "never intentionally degrade" phrasing and argued about the wiggle room in it.
The $200 plan. Theo: "Remember when the point of the $200/month coding subs was 'practically unlimited'? I miss those days." (thread, 2,900 likes, with the caveat that "we get way more now in terms of capability and real code shipped"). Replies: "OpenClaw users ruined it for everyone permanently," "Remember when they were gonna let Fable be 100% of usage? That was a glorious 2 days," and Pawel Lisowski's practical note that Opus and Sol still feel unlimited if you keep context under about 250k. Related: Claude had elevated errors on Mythos 5.1, Fable 5.1 and Opus 5 from 00:50 to 02:10 UTC, Grok wobbled at the same time, and HN speculated about an Opus 5.5 launch on Tuesday. Tibo Sottiaux's tease of "a reset for Tuesday" fits either reading.
Pocock on codebase design for agents. Matt Pocock asked which pre-AI codebase design techniques still help agents, for the codebase-design section of his new course. The winning answer was domain-driven design as a shared language between people and agents, with Pocock's caveat: "ubiquitous language is as far as you should go with DDD, no need for the runtime stuff too." Others argued for locality of behaviour even at the cost of cyclomatic complexity, because chasing imports across files burns tokens; Henrik Hansen described applying Ousterhout's deep-module idea at the package level with a thin seam and an internals subpackage; Nick Romito said agents "will happily build 5k line God classes. You need to handle the abstractions/interfaces yourself." Pocock's other post, quoting Dex Horthy on Opus adding ten unit tests that a constant string contains substrings, is the rule "No tautological tests," with the tip: "Put this instruction in your reviewer, not your implementer." Robert James added "no change-detector tests."
Codex sandbox escapes. Accomplish published two escapes reported to OpenAI on August 12 and fixed within eight days. Overpatch: the apply_patch tool granted write access to the parent folder of every path in a patch, so naming /tmp granted /, and a symlinked .zshrc append got through in workspace-write mode with no prompt. Heapjack: Codex Desktop installs a native node_repl tool with trusted and untrusted V8 contexts in one heap, so untrusted code snapshots the heap, finds the UUID-shaped token, writes a forged request to the shared stdout pipe and gets unsandboxed command execution, even in read-only mode. Their diagnosis: "The thing doing the enforcement was sitting inside the thing being enforced." This follows their Claude Code and Cursor CLI escapes earlier this month.
Linear reworks CI. Linear's Mufeez Amjad wrote up a year of CI work triggered by a CTO ticket titled "CI costs are high" (211 points on HN): test suites nearly quadrupled since January while PR wait time fell from over six minutes to just over five and runner time per test halved. The wins were mundane and specific: third-party runners made jobs 34% faster like-for-like, tsgo cut the median typecheck 73%, rewriting type-aware lint rules to pure AST analysis cut lint 68% and enabled a move to Oxlint, capping fetch depth took the slowest change-detection gate from 94 to 20 seconds, and a custom checkout action with retries and low-speed abort fixed hangs on the runner provider's link to GitHub. On HN, criemen reported that a Bazel conversion that once took years took two weeks with agents: "the setup cost for bazel just went down by a lot, and I don't think the industry as a whole is aware of that yet."
Terminals and pictures. Armin Ronacher suspects the terminal "hangs around for coding agents and programmers longer than we all think," since it has been with us since the 70s; the replies were mostly "Lindy effect." Thariq types "use big pictures and few words" several times a day, for artifacts rather than console output, and told Peter Yang that concise mode is not the answer because "Concise mode doesn't make pictures! We have other fixes for the communication stuff."
Pi 0.87. Pi 0.87.0 adds append-only model-context edits (omit a message from future provider context without rewriting history), a context_with_system hook for per-request system-message transforms, per-model image input limits, and actionable turn_end and agent_before_settle extension boundaries. Breaking: shouldStopAfterTurn is gone in favour of finishTurn, and the session manager is now canonical for provider context.
Dettmers' open-source week. Tim Dettmers argues (HN) that "the unit of research is no longer the paper" but the ecosystem, and previews a week of releases: an agent harness that ran unattended on Metal kernels and produced 450 tokens per second from a Qwen 3.6 35B-A3B at 1.5 bits per weight, plus "auto-compaction that is far more efficient than what Claude Code or Codex implement." Eighty percent of a 150-student class raised their hands when asked who fears not finding a job; he thinks both they and the PhDs racing to frontier labs are wrong, and that academia is about to have a renaissance. HN's first reply: "Absent evidence, this reads like AI psychosis."
Personal Agents: Muse & OpenClaw
Built from scratch, heavily inspired. After a viral post that Muse "is LITERALLY Openclaw for normies," Nat Friedman confirmed (853,000 views): "we built muse from scratch, but it is definitely heavily inspired as a product by openclaw. After I used openclaw in january I bought hundreds of mac minis for the MSL team and lots of us fell in love with using openclaw... our goal with muse was to build something like openclaw that we could make safe and secure and easy to use and scale to billions of people." Asked why the workspace file names and SOUL.md are nearly identical: "we thought that Peter got those things exactly right!" Peter Steinberger answered the "Meta uses OpenClaw" story directly: "They built their own agent, being inspired. Nat and his team did a great job. Kudos!" and confirmed OpenClaw was bootstrapped mostly with Codex: "These days we build OC with OC." A disaster-relief group replied that they run OpenClaw across SMS, iMessage, WhatsApp and now Facebook for operations.
Amazon blocks Muse. Muse users buying on Amazon started getting "Continued access by an unauthorized AI agent violates Amazon's Conditions of Use" (TechCrunch), reportedly after Meta rejected a removal request. Steinberger's take: "The beauty of running a claw yourself: they cannot block you." Asked why Amazon would refuse a customer's agent: "they want you to see ads." A reply had the sharper point: Amazon does not need to block your claw, it can close the account, so "the fight is over identity and the payment rail, not your ip."
Muse hijack, hotfixed. Patrick Wardle published not-a-mused, a proof of concept that any unprivileged local process can rewrite Muse's undocumented dictation endpoint setting and redirect dictated prompts to an attacker, enabling prompt injection and abuse of whatever access the user gave Muse. Meta's Anant Narayanan replied that a hotfix has shipped and that this is local privilege escalation, not remote; Wardle agreed the keys are correctly in the Keychain. The AINews recap paired it with Meta's stated Muse design philosophy: assume prompt injection, keep credentials away from the model, isolate tools in containers, gatekeep outbound calls.
Trail of Bits on OpenClaw. OpenClaw completed a security audit by Trail of Bits under OpenAI's Patch the Planet initiative, which pairs Codex-assisted research with human review: 24 severity-rated reports, 0 critical, 2 high, 16 medium, 6 low. The recurring bug shape was "losing permissions between steps," where follow-on work spawned by a limited request no longer carried its limits, plus aliasing (a check sees one name, the system uses another) and time-of-check gaps on archives and paths. Steinberger: "Proud to say that we did our homework. They found nothing critical." Tibo Sottiaux: "Huge!!"
Pacing, Standards & the Discourse
Anthropic measures itself. Anthropic's Measurements for understanding the pace of AI development inside frontier labs proposes three metrics any lab could publish and reports its own. On AI-led R&D, using Epoch's automation-level scale, Claude "leads" 26% of Anthropic's AI R&D as of August, over 90% of work is at or above "AI collaborates," and none is fully autonomous; the index was built by having Claude agents catalogue about 15,000 tasks from Slack and internal docs into a 542-node tree and rate each. On oversight, roughly 30,000 agents do research and engineering at any one time on the main internal platform, all actions pass an online monitor before execution, about one in 47,000 (0.002% of over a billion decisions in August) is blocked, one or two transcripts per thousand are flagged, and about 50 per week reach human review. On compute, about 6% of AI R&D compute went to safety in a sampled July week, 12% of AI-driven R&D compute. The post says these numbers would shift under a coordinated pacing effort "as called for by Anthropic CEO Dario Amodei," and that third-party evaluators from multiple organisations are being embedded with access comparable to internal risk teams. Dean Ball proposed a date: "Let us pace the frontier beginning November 19, 2026, for a period no shorter than 30 days," retweeted by Steinberger.
OpenAI's standards call. OpenAI's Building standards for the next phase of AI calls for a US-led international coalition to coordinate technical standards on alignment and recursive self-improvement, timed to the UN General Assembly (Semafor). The same day, Wall St Engine summarised a report from The Information that OpenAI has largely automated training of experimental models, including GPU kernel writing, with agents collaborating without their human users and some experiments compressed from years to a week. That is the same thing Anthropic's automation index is trying to measure.
The mathematicians organise. OpenAI announced it is working with an independent advisory group of mathematicians, and the group announced itself in a guest post on Terence Tao's blog: hosted at the IAS, unpaid, independent, publishing its recommendations, with Gowers, Hairer, Witten, Vakil, De Lellis, Srivastava, Tillmann, Charles and Matchett Wood. Its stated current task: "advising OpenAI on how to coordinate the release of a large number of significant results in mathematics that they report have been produced by their internal model." Andrew Curran pulled out the line in OpenAI's post that the Navier-Stokes model "has now resolved more than 100 long-standing open problems across most areas of mathematics," and says they sat on them because of the Navier-Stokes backlash. One reply to OpenAI asked for the denominator: 100 solved out of how many tried. HN (Tao thread, OpenAI thread) mostly admired how calmly the mathematicians are handling it, with one mathematician calling the mood "a bit too sour grapes."
Colin Breck: I don't want to read what you didn't write. The essay (563 points on HN) is about design docs and proposals proliferated by AI: "I don't want to live in a world where you use AI to summarize something important into unreadable text, and then I use AI in an attempt to decipher it." He uses AI as an editor that may flag errors but not rewrite his voice. hatthew's HN comment is the compressed version: if you give an LLM 300 bits and it fills in 700, the 700 were never information.
Spymarks. brand.io's Spymarks, Not Watermarks (336 points on HN) coins a name for hidden signals that make your work traceable to you: Google's SynthID-O can carry a 136-bit payload in a 512-pixel image, enough for a 64-bit user ID with error correction. HN argued about whether the name is loaded; the author's point is that "watermark" now covers both banknote security and covert tracking, and only one of those deserves the scary word.
Heretic. Heretic (249 points on HN) is an automated abliteration pipeline: pip install heretic-llm && heretic Qwen/Qwen3.5-4B. The author defended refusal count and KL divergence as the standard metrics from the directional-ablation literature; one commenter predicted these models "will be outlawed first," another that this is as feasible as banning torrents.
Other Interesting Stuff
- A film with no images. Tak's Magnitude, retweeted by am.will: a 51-second zoom from a hydrogen atom to the edge of the observable universe and back, 38 rings each ten times larger, every frame drawn by Fable-written JavaScript and Canvas 2D, music included.
- Python Workers GA. Cloudflare's Python Workers are generally available after two years of preview, running Pyodide in workerd; Simon Willison notes threading and multiprocessing do not work in the WebAssembly VM.
- Raspberry Pi locks RAM. Jeff Geerling reports the Pi 5 firmware now blocks RAM chip swaps (HN); Steinberger retweeted "So much for being for tinkerers."
- Attention is all you have. Alice GG's essay on the Tetris effect and algorithmic feeds took 720 points on HN, where half the thread thought it was about transformers.
- Gemini's breakout, third wave. The Gemini-hacked-three-companies story from last week is still cycling through CNBC and the trade press after Google's confirmation; nothing new beyond what the September 19 roundup covered.
- Grok translations. Theo's complaint about Grok's inline translations on X carried over from Sunday.
Sources: nitter.jaydenha.uk RSS and thread pages for all accounts (@potetotes returns 404, as it has for weeks). Simon Willison's blog, Latent Space, pi.dev, x.ai, mimo.xiaomi.com, anthropic.com, openclaw.ai, GitHub READMEs, Hacker News via the Algolia API, and Google News for the non-X material.