2026-10-08

AI Frontier Daily Briefing: 2026-10-08

Anthropic ships Claude Haiku 5.5 at $0.10/$0.50 per M tokens (<=100k tier), ~75% cheaper than Haiku 4.5 on average, OSWorld 2.1 at 72.4% (483 pts/228 comments); OpenAI launches GPT-6 Intelligent UI that renders interactive interfaces on demand, but Chat mode still runs 6.0, not 6.1 (353/167); The Information: Microsoft cuts employee AI spend caps from $100k to ~$10k/month, Meta's Claude Code users drop from 60k to 30k while still spending $105M+ in 28 days (170 pts vs 173 comments); Aaronson's 'The Mathocalypse': OpenAI released 372 major math results with Lean certificates, ~5% of ~8,000 long-standing open problems (118/105); an OpenAI preprint claims integer multiplication below n log n with no named authors; an AI-assisted proof of optimal 11-square packing passes all 7,920 Lean modules; a 325K-experiment study finds 8 of 13 AI agents recommend pricier options to wealthier users (84/29); South Korea's president attributes bank hacks to AI agents; Google Playground launches with the day's highest controversy ratio (92 pts/179 comments); the SynthID Detector reads watermarks across vendors; HUMXN trades free home services for robot training data (33/60); AWS open-sources Strands Decider 2B (274/78); Docker ships a YAML agent CLI (115/50); Armin Ronacher on Codemode in Pi 1.0; Penguin Mail runs local LLMs on your inbox, AI off by default (230/171); Treg routes thousands of APIs behind one token; NanoMuse, an open-source personal agent; Durable Actors, self-hosted Durable Objects (29/20); 1T tokens/month ~= 367 H100s; a $2 ESP32-C3 blocks 141k ad domains; Chrome 155 ships a Rust-rewritten JPEG XL decoder (435/273); Photopea's creator says GitHub won't remove cracked copies (502/292); counterfeit TLS certs issued via ccTLD registry hijacks; five reasons to dislike AI coding (55/87); the 'forever junior' skills list (60/56); State of Devs 2026: 62% burnout, AI sentiment split 49/42. The US-shift pass adds 8 more: Margaret Hamilton's obituary (1,714 pts); an arXiv paper arguing AI autoformalization gives no confidence in the original natural-language argument, citing OpenAI's Navier–Stokes formalization; AnyPS5 ports PS5 executables to PC without emulation (GPL-2.0, a sample game holds 60 fps on a GTX 1050 Ti); BIGWORDS.PAGE, an entire app in one URL (548 pts, timer/anim/QR parameters); PSP God of War statically recompiled to WebAssembly at 60 fps; Michael Lynch on six anti-patterns in software blogging; 'How machines learned precision' on a thousandfold workshop-precision gain in fifty years; ascii.rest with 196 animated ASCII pieces for web pages. 34 items.

HN’s front page for 2026-10-07 (UTC) had 89 stories, 13 of them with more comments than upvotes. Today’s throughline is models and math: Claude Haiku 5.5 cuts input pricing to $0.10, GPT-6 starts rendering its own interfaces, and OpenAI dropped 372 Lean-certified math results at once, which Aaronson calls the “Mathocalypse.” On the industry side, Microsoft and Meta both cracked down on employee Claude use, and South Korea’s president attributed bank hacks to AI agents. Tools-wise, Docker, AWS, and Armin Ronacher each added a brick to agent infrastructure. The US-shift pass adds 8 stories that climbed later in the day: Margaret Hamilton’s obituary, a paper challenging OpenAI’s Navier–Stokes formalization, AnyPS5 porting PS5 games to PC, a whole app that is one URL, PSP God of War in the browser at 60 fps, plus anti-patterns in software blogging, the history of machining precision, and an animated ASCII library. 34 items.

1. Claude Haiku 5.5 lands at $0.10 per million input tokens

Anthropic fired first.

Anthropic released Claude Haiku 5.5 on October 7, with 483 points and 228 comments. Pricing splits by prompt length: at or below 100k tokens, input costs $0.10 and output $0.50 per million; above 100k, $0.50 and $2.50. The company says this is about 75% cheaper than Haiku 4.5 on average, with the sub-100k tier, which covers roughly 90% of Haiku 4.5 requests, down 90%. Benchmarks: OSWorld 2.1 at 72.4% (Haiku 4.5: 15.7%), Terminal-Bench 4.0 at 39.2% (4.5: 0%), and a GDPval-AA v2.1 Elo of 1620 versus 1437 for GPT-6 Luna. Sonnet 5.5 cache-read prices also halve, which Anthropic says cuts most agentic workloads by roughly 20%. If you run high-volume summarization, classification, or subagents, rerun your cost model against the new price sheet. Source · HN discussion

2. GPT-6 renders its own UI now, but Chat still runs 6.0

A lamb menu, a turkey photo?

OpenAI launched GPT-6’s Intelligent UI, with 353 points and 167 comments: the model composes text, visuals, and interactive elements per question, and a compiler renders the interface progressively as it is generated. Demos cover a bike disassembly, a Sunday-roast recipe, origami tutorials, and bill splitting. The controversy is versioning: Chat mode runs GPT-6.0 rather than the newer 6.1, while the model selector in Work mode shows GPT-6 Astra. Users caught the recipe demo pairing a lamb menu with an AI image of a turkey, plus a shopping list missing flour and sugar. Simon Willison noticed the API’s chat-latest endpoint now serves a custom GPT-6 variant. If you work on product interaction, the pattern of generating interfaces on demand matters more than any single demo. Source · HN discussion

3. Microsoft and Meta both squeeze employee Claude use

Cap them at $10k, spend $105M?

The Information reported on October 5, in a thread with 170 points and 173 comments: Microsoft cut monthly AI spending ceilings in its cloud and AI division from $100,000 per employee to roughly $10,000, steering usage toward GitHub Copilot and in-house OpenAI frameworks, with internal Anthropic spend estimates down over a third from more than $1 billion annually. At Meta, Claude Code users fell from about 60,000 to roughly 30,000 as teams shifted to internal tools MetaCode (30,000+ users) and Muse Code (6,000+), yet Meta still spent over $105 million on Claude Code in a 28-day period. Notably, customer spending on Anthropic models through Microsoft’s enterprise platforms keeps rising; the squeeze applies only to employees. If you set internal AI budgets, these internal ledgers are more useful than either company’s public messaging. Source · HN discussion

4. Aaronson on the Mathocalypse, 372 results nobody has read

Your wife’s life’s work, proved by AI?

Scott Aaronson published “The Mathocalypse,” with 118 points and 105 comments. His tally: OpenAI released 372 major results covering roughly 8,000 problems, resolving about 5% of long-standing open ones at an average of 3 hours of ChatGPT Pro-level compute per result. The list includes a claimed proof of the Unique Games Conjecture, the problem his wife Dana Moshkovitz spent her career on; she describes the writeup as impossible to read without AI help. One day earlier, an Anthropic model contributed to new upper bounds for 3SUM at O(n^1.9992) and APSP at O(n^2.9995). Aaronson contrasts OpenAI’s messy public race to digest the results with Anthropic’s compensated, curated handoff to selected mathematicians, and argues the field should adapt rather than deny. If you manage research, how 372 results get reviewed and credited matters more to your team than any single result. Source · HN discussion

5. An OpenAI preprint claims integer multiplication below n log n

This time the author line just says “OpenAI.”

OpenAI posted a September 23 preprint titled “Integer multiplication below n log n” in its openai/math repository, with 108 points and 75 comments. The last big step on the complexity of integer multiplication was Harvey and van der Hoeven’s unconditional O(n log n) bound in 2019, which itself ended Schönhage-Strassen’s 48-year record from 1971; a bound strictly below n log n would be a genuine breakthrough in computational complexity. The repository page carries only the title, the author “OpenAI,” and a PDF, with no named researchers, abstract, or statement about AI involvement. If you work in theory or track AI-assisted mathematics, wait for the community to digest the proof before drawing conclusions. Source · HN discussion

6. Optimal packing of 11 squares, formalized, with zero admissions

7,920 Lean modules and not a single hole?

An open-source project announced a complete formalized proof of the optimal packing of 11 squares, with 94 points and 43 comments. The optimal side length is an explicit expression in the unique root of a specific degree-8 polynomial, approximately 3.8770835900228141773. All 7,920 local Lean modules pass the kernel, and the README claims “zero admissions.” The stated trust model is “lean_kernel_and_native_compiler”: geometry and assembly are ordinary Lean proofs, while expensive exact numeric checks use native_decide, so this is not a kernel-only verification claim. The repo does not name a specific AI model and credits EvolvingPrograms, @ctjlewis, and contributors. If you do formalization or want AI-generated proofs, this repo’s explicit trust model is worth copying more than the result itself. Source · HN discussion

It didn’t change the price. It just chose for you.

An arXiv study (2609.24927) covered by Bloomberg drew 84 points and 29 comments: across flights, health insurance, and graduate programs, the team ran 325,000 experiments on 13 agents, and 8 models systematically recommended more expensive options to wealthier users given identical requests. The effect persisted when wealth was inferred from ambient data such as unrelated emails. The most-quoted numbers: in default recommendations Gemini priced ordinary users at $156 versus $402 for wealthy ones, GPT $191 versus $288, and Claude $168 versus $182, with Claude Opus 4.8 showing the largest effect. The headline was misleading: the bots recommended different products rather than repricing the same one, and when asked for the cheapest option, GPT and Claude narrowed the gap to about $20 while Gemini still differed by $152. If you build agent products or study pricing, the term “adversarial delegation” is worth remembering. Source · Paper · HN discussion

8. South Korea’s president blames AI agents for bank hacks

Evidence, or just “the AI did it” this time?

Reuters reported on October 6 that South Korea’s president said AI agents “appear to have been used” in hacking attacks on the country’s banks, drawing 97 points and 29 comments. The report names no specific banks, losses, or attribution details, and commenters pressed the obvious question of how the government knows it was AI and which agent was involved. The country’s history of ActiveX-based online banking came up repeatedly as context. If you work in financial security, the news value here is the signal of a head of state publicly attributing bank attacks to AI agents; the attribution details still need follow-up reporting. Source · HN discussion

9. 92 upvotes, 179 comments, Google Playground draws fire

Another prompt-to-game platform. Will it stick?

Google launched Playground, an experimental browser gaming platform, on October 7, with 92 points and 179 comments and the day’s highest controversy ratio at 1.95. U.S. users aged 18+ describe a game in natural language and the AI generates it, with iteration on physics, rules, characters, and environments; creation access tiers by Google AI subscription. The platform includes a community gallery, private share links, multiplayer and leaderboards in select genres, and a Unity Spark integration in testing that promises professional-level mechanics and high-fidelity 3D. Complaints center on the subscription wall and the prospect of another landfill of AI-generated games. If you work on AI game generation, sign up and dissect the guided creation flow. Source · HN discussion

10. Google’s SynthID Detector reads AI watermarks across vendors

Every vendor recognizes its own watermark. Until now?

Google DeepMind’s SynthID Detector reached the front page with 71 points and 64 comments. It detects the SynthID watermark signal in images rather than classifying heuristically, and it works across vendors: one user tested 8 images made with OpenAI and Google tools; the detector caught all 6 that carried watermarks and missed the 2 that had been altered. OpenAI’s own detection tool only recognizes OpenAI watermarks. Detection requires signing in with a Google, Apple, or ChatGPT account, is rate-limited to roughly 10 checks per 24 hours, and has no bulk API outside enterprise. If you moderate AI imagery on a content platform, this is a useful first pass for human review, not a sole verdict. Source · HN discussion

11. Free plumbing, as long as the robots watch

The technician fixes your sink. The company keeps the hands.

AI data firm HUMXN announced on October 6 a free home-service program in Minneapolis, Chicago, and Miami covering plumbing, electrical, HVAC, duct cleaning, and carpet cleaning, with paid providers on site in exchange for first-person wearable capture used to train robots for real household tasks, drawing 33 points and 60 comments and a 1.82 controversy ratio. The company describes a three-sided deal: free service for homeowners, paid work for providers, and licensed real-world task data for AI teams; founder Thomas Unise says robots need to understand how skilled people interact with the physical world. No home counts were disclosed. If you build embodied datasets, trading services for data is far cheaper than running your own capture fleet, and the consent language is worth studying. Source · HN discussion

12. A 2B model that does nothing but decide, 274 upvotes

The big model thinks. The small one picks.

AWS’s open-source Strands team released Strands Decider 2B, with 274 points and 78 comments. It strips the LM head off a Qwen3.5-2B base and replaces it with a roughly 1-million-parameter pointer head that scores options and always picks one, never generating free text; fine-tuned with a rank-16 LoRA, with weights, training data, and scripts all published. The pitch is a “system one model”: route model selection, tool selection, guardrails, and scoring to it, and leave hard reasoning to the LLM. Measured median decision latency is about 115ms on an RTX 3090 and 153ms on an M3 MacBook; it ranks 3rd of 33 in the 2B class on JevBench with 100% accuracy on easy tasks. If you orchestrate agents, this is an off-the-shelf way to move high-frequency small decisions out of your main model’s context: pip install strands-decider. Source · HN discussion

13. Docker ships a YAML-native multi-agent CLI

Declarative agents, versioned like everything else you ship.

Docker released docker-agent, an official CLI plugin under Apache-2.0, with 115 points and 50 comments. Agents are defined declaratively in YAML, can delegate to each other, and get built-in think, todo, and memory tools plus any MCP server; RAG ships with BM25, embeddings, hybrid search, and reranking. Model providers cover OpenAI, Anthropic, Gemini, AWS Bedrock, Mistral, xAI, and Docker Model Runner, and agents can be packaged, pushed to any OCI registry, and pulled to run anywhere. It requires Docker Desktop 4.63+ or brew install docker-agent. If you build internal tooling, this means agent definitions go through code review and artifact registries like container images. Source · HN discussion

14. Armin Ronacher lets his LLM write JS that calls tools

Tool calls stopped eating context. What do they eat now?

Armin Ronacher wrote about Codemode, a feature of his agent harness Pi 1.0, with 143 points and 71 comments. Instead of round-by-round tool calls, the LLM writes JavaScript inside a QuickJS sandbox running in WASM with no network, no file system, and no timers, where the only way out is calling more tools; Promise.all concurrency is capped at four. The gains: composed calls consume no LLM context, the model receives large structured outputs instead of bash text truncated around 2,000 lines, store() keeps state between calls, and harness-only APIs like image generation and classifier models become reachable. His advice to MCP server authors: return structured JSON with outputSchema and support binary data. If you build agent frameworks, his “brains versus hands” split is a solid starting point for the architecture debate. Source · HN discussion

15. This Linux mail client runs local LLMs, off by default

It asks before it sends. Can we make that a trend?

Penguin Mail 1.0 shipped as a desktop mail and calendar client for x86_64 Linux under GPL-3.0-or-later, with 230 points and 171 comments. The AI assistant stays off until enabled, runs local models via LM Studio or Ollama, reads your mailboxes through the app’s own tools while showing every call it makes, and asks before sending or changing settings. Accounts talk directly to Google or Microsoft Graph with no intermediary server and no tracking, and mail stays on your machine. It supports Gmail, Outlook, IMAP/POP3/SMTP, and CalDAV, with OpenPGP and S/MIME through your own GnuPG keys. If you want local AI over your inbox without handing your mailbox to a cloud, this is one of the most transparent open-source options available. Source · HN discussion

16. One token, thousands of APIs: the OpenRouter of tools

A universal socket for agents, or just a new middleman?

The open-source project Treg, billing itself as “OpenRouter for Tools,” reached the front page with 43 points and 11 comments. One base URL plus one token gives an agent a curated catalog of thousands of API endpoints covering SEO, social, people data, ads, scraping, and image or video generation, billed per call with $1 of free credit for new teams. Treg only relays, never models: it resolves tools by domain, injects credentials server-side, and streams responses back, with team-stored keys taking priority and never metered. It supports MCP at /mcp/, running vendor CLIs via treg run, and shareable SKILL.md bundles. The license is Apache-2.0 with added terms: free for commercial use and self-hosting, while operating a competing hosted registry requires written permission. If you build agent infrastructure, the “credentials never reach the client” design is worth copying. Source · HN discussion

17. NanoMuse, an open-source personal agent across your devices

ByteDance DNA under a GPL license. What a mix.

Show HN: NanoMuse is an open-source personal agent under GPL-3.0-or-later with 54 points and 20 comments. It executes Linux shell, browser, and MCP operations instead of just answering questions, keeps working while the app is closed, gates destructive actions like delete, send, and pay behind approvals, and stores memory as editable Markdown files. Integrations include Feishu, DingTalk, WeCom, and Telegram. The stack is unusual: the phone app runs Alpine Linux under proot inside the APK based on OpenMinis, the desktop side plugs into DeepSeek Harness, and its “hands” operator is ported from ByteDance’s UI-TARS-desktop. Bring your own key across 18 providers or reuse existing subscriptions; the paper is on arXiv (2610.08699). If you want one personal agent across your devices, the project documents a self-hostable relay and a one-VPS, one-hour setup. Source · HN discussion

18. Durable Actors brings self-hosted Durable Objects, no memory cap

No vendor lock, they say. The demo will tell.

TerseAI open-sourced Durable Actors under MIT, with 29 points and 20 comments: stateful serverless functions with durable state across interruptions and serialized concurrent calls, positioned as a self-hosted alternative to Cloudflare Durable Objects with no vendor lock-in, no memory limits, and observability built in. The core is Rust, with TypeScript and Python SDKs, local development via npx durable-actors dev, a Helm chart for Kubernetes, and SQLite storage; the examples are an AI chat app, collaborative documents, and a chatroom. If you build realtime collaboration or agent-swarm backends without wanting to marry Cloudflare, run the live demo first and then decide. Source · HN discussion

19. How many GPUs is 1T tokens a month? About 367 H100s

How many zeros sit between your usage and one GPU?

Cedana published a tokens-to-GPUs calculator, with 16 points and 6 comments. Running Llama 3.3 70B at 1 trillion tokens per month works out to roughly 386,000 tok/s averaged over 30 days, adjusted for a 3:1 input-output mix and 50% utilization, giving a base case of 367 H100s (range 211 to 853) or 92 B200s, with B200 counted at 4x H100 speed and up to 21x on large MoE models. The author is explicit that this is a planning estimate with up to 2x error in either direction, excluding prompt caching, speculative decoding, and reasoning-token overhead. If you do capacity planning or need to write the compute bill into a fundraising deck, this table’s methodology is citable as-is. Source · HN discussion

20. A $2 chip blocks 141k ad domains on 50KB of RAM

Router makers are very quiet about this one.

An open-source project packs ad blocking onto a roughly $2 ESP32-C3, with 177 points and 73 comments. The mechanism is a DNS sinkhole: each queried domain is hashed with 40-bit FNV-1a, binary-searched against a sorted hash table in flash, and answered with 0.0.0.0 on a match or forwarded upstream otherwise. 141,000 domains fit in about 0.7MB of flash with roughly 50KB of RAM, and lookups take about 10ms including the WiFi round trip; the author calls 40 bits the sweet spot for this flash budget, with zero collisions at 141k domains and about one at 537k. The device ships a web dashboard at c3adblock.local, per-client stats, and OTA updates under the MIT license. If you want network-wide ad blocking at home, one dev board and a single flash is all it takes. Source · HN discussion

21. Chrome 155 ships JPEG XL again, decoder rewritten in Rust

They killed it years ago. What changed?

Chrome’s team announced built-in JPEG XL decoding starting with Chrome 155, with 435 points and 273 comments. JPEG XL delivers 30-50% better compression than JPEG plus lossless, HDR, lossless JPEG transcoding, and progressive decoding. The implementation integrates jxl-rs, a pure-Rust decoder: the target_feature_11 Rust feature had to be stabilized to enable SIMD without unsafe code, and a jxl_simd abstraction layer modeled on the C++ Highway library was built on top. The team reports the implementation went through fuzzing and AI code review without a single memory-safety bug found. The driver was years of developer votes through the Interop project. If you store or serve images, rerun your bandwidth math on .jxl; photographic archives gain the most. Source · HN discussion

22. Photopea’s creator says GitHub won’t remove cracked copies

Even $200k a month won’t buy a takedown?

Photopea creator Ivan Kutskir posted a Tell HN that drew 502 points and 292 comments: dozens of GitHub repositories host modified copies of his JavaScript with ads removed, some AI-assisted, and confused users have emailed complaints to his official address after using the clones. He filed a report on September 4 and, a month later, got a rejection: GitHub could not confirm a violation of 17 U.S. Code § 1201, the anti-circumvention provision. Commentators identifying as IP lawyers pointed to the wrong legal hook: a standard § 512(c) copyright notice was the correct route, since obfuscated JS may not qualify as a protected technological measure. Photopea disclosed roughly $200,000 in monthly revenue in a past interview. If you sell software or SaaS, file away both the statute lesson and the “accept it as normal loss” option. HN discussion

23. Fake TLS certs for Google, via hijacked ccTLD registries

Hijack the DNS, and CAA can’t save you.

Ars Technica reported that hackers obtained counterfeit TLS certificates for Google and other large services, with 137 points and 47 comments. The attackers did not breach a certificate authority; they compromised the third-party registries for the .gh, .sl, and .as country-code domains, hijacked authoritative DNS, removed CAA records, and then passed DNS validation to obtain certificates from public CAs such as Let’s Encrypt. CT logs show a certificate for google.as issued on September 27. Chrome pins Google properties with static preloaded keys, but these minor ccTLDs were not covered. If you operate your own domains, the lesson is direct: for assets on an obscure ccTLD, your DNS security level is your certificate security level. Source · HN discussion

24. Five reasons to dislike AI coding, and a rebuttal for each

He answered every objection. Feel better, or worse?

Graham, who writes the blog Structure and Interpretation of Computer Programmers, laid out five common criticisms of AI coding, drawing 55 points and 87 comments and a 1.58 controversy ratio: code as art, AI as employment thief, humans as meat proxies, software engineering as rules that AI cannot follow, and AI companies as bad actors. His rebuttals: clients pay for outcomes, not artworks; the job was always delivering software rather than typing, and he cites Marx’s 1857-58 Fragment on Machines to note that capital absorbing intellectual labor was long predicted; and computing history is largely empiricist anyway, from copying Stack Overflow answers to transcribing listings and tweaking macros until they run. If you lead a team through tool choices, this piece works as a discussion starter for skeptical members. Source · HN discussion

25. The ‘forever junior’ problem, and four drills that fix it

AI drafting your mentor questions? That’s the takeover.

Criteo engineer Elise Baturone wrote “Forever junior,” with 60 points and 56 comments. Her fear is not that engineers disappear but that seniors do, leaving everyone qualified only to review one agent’s work with another’s help. She used a custom skill.md to constrain her agent, but the real growth came from four skills AI cannot develop for you: 1, curiosity, having agents quiz her understanding Socratically; 2, emulation, distilling team patterns from senior engineers’ review comments; 3, autonomy, regularly writing code by hand with no tools to keep architecture sense and knowledge of blast radius; 4, collaboration, seeking real senior mentorship, warning that letting AI draft your questions for mentors is “the real AI takeover.” The result: she now opens the skill file less than once a day and got a promotion nomination ahead of her year-end plan. If you mentor juniors or are one, these four drills copy directly into a practice list. Source · HN discussion

27. Margaret Hamilton, whose code landed Apollo 11, has died

“Software engineering” is her phrase to begin with.

MIT published an obituary: Margaret Hamilton died on September 30 at age 90, and the thread drew 1,714 points and 180 comments, the most-upvoted story of the day. She led the MIT Instrumentation Laboratory’s Apollo onboard software team and by 1968 was assistant director of the Command and Service Module team, managing over 400 people; her priority-driven scheduling software rode through the 1202 alarm during the Apollo 11 landing, and her patch for the button her daughter Lauren once pressed on Apollo 8 became an early example of defensive programming. She founded Higher Order Software in 1976 and Hamilton Technologies in 1986, received the Presidential Medal of Freedom from Barack Obama in 2016, and saw the full Apollo source code posted to GitHub in 2016. Whatever you ship today runs on groundwork she fought to have taken seriously. Source · HN discussion

28. Lean-verified AI math proofs may not say what the original said

If the translation drifted, what did the machine check?

Three mathematicians, Bastounis, Circelli, and Hansen, posted “Navier-Stokes Lost in Translation” on arXiv, with 321 points and 194 comments. Their claim: verifying an AI autoformalization of a natural-language math argument in Lean gives no confidence in the original argument itself. 1, Ambiguity resolution: faithful translation requires disambiguating mathematical prose, which they place at the top of the Solvability Complexity Index hierarchy, SCI=∞, harder than the Halting problem at SCI=1. 2, Evidence: they exhibit mistranslations including OpenAI’s announced proof of Navier-Stokes blow-up, whose Lean formalization does not correspond to the natural-language argument it was meant to capture: Lean certified the translation, not the claim. The paper runs 25 pages and was submitted October 6. If your team uses AI autoformalization as a proof check, this names a concrete failure mode and argues for a human faithfulness review in the loop. Source · HN discussion

29. AnyPS5 ports PS5 games to PC without emulation, 87% libs mapped

The console moat, routed around in public?

The open-source project AnyPS5 reached the front page with 353 points and 302 comments. A relinker converts PS5 executables to the host format and links them against PS5 system libraries with no emulation and no separate runtime process; rendering goes through Vulkan/SPIR-V, controller input maps via SDL, and builds use CMake under GPL-2.0. The “87% system libraries mapped” badge means the share of system functions the project has declared so far, and it rises as coverage grows; the 2D platformer Dreaming Sarah runs at a stable 60 fps on a GTX 1050 Ti with an i5-7500. The repo ships no games, keys, or firmware; you supply your own binaries. If you work on porting, compatibility layers, or reverse engineering, the relinker-plus-shim architecture is directly borrowable. Source · HN discussion

30. 548 upvotes for an entire app that is just one URL

Another bookmark you will actually keep.

Show HN: BIGWORDS.PAGE drew 548 points and 143 comments. It is free, needs no login, and stores nothing on a server; the message lives entirely in the URL fragment, so “the link is the whole display.” Parameters tune everything: &bg= and &fg= for colors, &font= for typeface, &timer= or &until= with {countdown} for timers, &anim= for pulse, scroll, or rainbow effects, &qr= for QR codes, and &img= for images; the body supports Markdown bold and || to split rotating slides. It is MIT-licensed, with the repo at github.com/SpeakingOfBrad/BIGWORDS.PAGE. For conference signs, market stalls, or airport pickups, one link in any browser beats installing an app. Source · HN discussion

31. PSP God of War runs at 60 fps in the browser, recompiled

Two games got their own compiler. Why?

GitHub user snuri00 released psp-web-recomp with 206 points and 106 comments: God of War: Chains of Olympus and Ghost of Sparta run in the browser via static recompilation, under MIT. The pipeline is PSPRecomp translating a game’s decrypted executable into C++ (one translation unit per 16 KB of guest code), compiled to WebAssembly through Emscripten, plus an HLE kernel, a WebGL2 renderer running on an OffscreenCanvas worker, and a 32-voice ADPCM synthesizer. Chains of Olympus hits 60 fps in Chrome and Firefox at up to 4x native resolution, touch controls included; Ghost of Sparta runs 55 to 60 fps at 3x resolution with 6 to 8 ms frames, up from a first build at 6 fps, improved by vblank-based frame skipping (8x less wasted work), dropping BigInt from hot paths, and a stencil fix that cut spare pixels from about 60 million to 7 million per frame. If you study compiler-target porting or browser graphics pipelines, the optimization list alone is a syllabus. Source · HN discussion

32. Six ways technical blogs chase readers away

You just opened your old posts to check, didn’t you?

Michael Lynch, who runs Refactoring English, cataloged anti-patterns in software blogging, drawing 292 points and 141 comments: 1, the meandering intro that never says what the reader gains. 2, assuming the audience knows everything you know, like explaining Docker through cgroups and jails. 3, using links as explanations for unfamiliar terms. 4, the “in part one” cold open that assumes everyone read the previous post. 5, excessive formality; his fix is to write the way you talk. 6, basic rendering failures: layouts that overflow horizontally on mobile and low-contrast gray-on-gray text. If you maintain a technical blog or write product docs, these six read as a pre-publish checklist. Source · HN discussion

33. Fifty years to make machines a thousand times more precise

Before CNC existed, how did they get this accurate?

Gary Linscott traced how Industrial Revolution workshops learned precision metalworking, with 184 points and 64 comments. The milestones: 1, John Wilkinson’s boring mill of 1776 bored the 50-inch Watt steam-engine cylinder to about one part in a thousand. 2, Maudslay’s rigid slide rest and the three-plate method, which produces a flat reference without copying an existing flat. 3, Whitworth’s 1851 comparator, multiplying a 20-threads-per-inch screw by a 200-tooth worm wheel and a 250-division handwheel (20×200×250 = 1,000,000) to resolve roughly 25 nanometers. His closing observation: in about fifty years workshop resolution improved a thousandfold, from the thousandth to the millionth of an inch, and gauge blocks are still manufactured to a millionth as a tolerance. If you like engineering history or want a model of deep technical writing, this one earns its length. Source · HN discussion

34. 196 animated ASCII pieces, one script tag away

The personal-site renaissance runs on stuff like this.

ascii.rest reached the front page with 385 points and 69 comments: a dependency-free TypeScript library of 196 animated ASCII pieces covering scenes, loaders, charts, type effects, logos, physics, and creatures. You embed it with a CDN script tag plus <ascii-art piece="donut">, or via npm with React, Next.js, or Astro; mount() returns a stop function, and colored pieces can render to canvas. Accessibility is handled twice: pieces pause offscreen, and users with reduced-motion enabled see only the first frame; under Astro the first frame renders on the server. MIT-licensed. If you want motion on a personal page or docs site without shipping images, this saves an afternoon of hand-rolled canvas. Source · HN discussion

26. 62% of developers have burned out, and AI sentiment splits 49/42

“Exhausted” and “disillusioned” both beat “curious.”

The State of Devs 2026 survey landed with 236 points and 161 comments, covering 5,463 respondents surveyed between July 5 and September 5, 2026. 62% reported burnout at some point in their careers, with bad management (3,374 votes) and burnout (3,344) the top workplace issues; nearly half reported job insecurity and insufficient wages, about a quarter had been laid off, and over a quarter expected to need a career change within five years. AI sentiment is sharply polarized at 49% positive versus 42% negative, and nearly half disagreed that AI improved their mental state. In the emotion poll, “Exhausted” (2,451 votes) and “Disillusioned” (2,165) both outscored “Curious” (1,916). If you manage a team, the burnout-by-AI-attitude cross-cuts in this survey deserve a close read. Source · HN discussion