FOMO Buster #014: The 722-proof machine

A newsletter that makes FOMO obsolete.

if i fear missing out on reading this newsletter, did i win or lose

Curated by Maurice.

01 · The One Thing

OpenAI drops 722 math manuscripts on GitHub

On October 6, OpenAI published “Sharing AI progress in mathematics” and released the openai/math repo (Apache-2.0): 722 manuscripts in 372 result families, produced by an unreleased internal frontier model, selected from an evaluation of ~4,000 open problems. The average compute per result: ~3 hours of ChatGPT Pro thinking. Checkability is built into the release: 235 of the 372 families link to Lean formalizations, and a formalization index lists 162 papers with a formalized main result, checkable via Comparator. The claimed results include the quasi-Riemann hypothesis and the Hodge conjecture for CM abelian varieties. Ten abridged reasoning summaries ship with it (irrationality exponent of pi, Mahler conjectures, free group factors). The caveats are printed up front: OpenAI says unformalized results may have issues, and none of it is peer reviewed yet. HN’s #2 thread (1,088 points, 1,149 comments) is still going.

The read: this is the second wave (The Verge calls it “another batch”; Scientific American says the field is “already in shock”). The important shift is the release format: a Lean-verified corpus with reasoning traces and compute budgets per result, shaped with an IAS advisory group. That is math as a dataset, not as papers. If the Lean formalizations hold, the bottleneck in mathematics moves from “who can prove it” to “who can check it,” and a 372-family corpus of formalized results becomes training data for the next model. The debate is ethics and research practice: 722 manuscripts dumped at once breaks the peer-review pipeline and turns mathematicians into verification annotators. Naomi Saphra’s line is the crispest consequence: you can no longer use the duration of an open problem as a signal of its difficulty or importance.

Signal: OpenAI is industrializing mathematical discovery: ~4,000 open problems in, 722 manuscripts out, Lean-checkable, with the compute budget printed on each. The field’s reaction (ethics debate, shock) is lagging the release format, which treats proofs as a verifiable dataset. If the Lean corpus holds, the next model trains on it, and the loop closes. Sources: OpenAI announcement · GitHub repo · HN thread

02 · Lab Watch

Claude moves into Google’s house

Anthropic launched Claude for Google Workspace in public beta on October 6, a first-party add-on available on all paid Claude plans that puts Claude in a sidebar inside Docs, Sheets, and Slides. It reads the open file, sees your selection, and edits directly: targeted rewrites in Docs (larger edits surface as suggestion cards), formulas, pivot tables, native charts, and Python data cleaning in Sheets, and deck-consistent new slides in Slides that even flag overlapping elements or hard-to-read text. The default is “ask before edits” approval cards; “accept all edits” is the alternative. Separately, beta connectors let Claude create and edit Google files from inside the Claude chat. Access mirrors the user’s existing Google sharing permissions, and enterprise controls (Compliance API, customer-managed encryption keys, OpenTelemetry audit export) carry over to the add-on.

The strategic read: Anthropic just walked into Google’s own productivity suite with its own agent and offered it as the alternative to Gemini. The enterprise distribution war is no longer about which chat tab you keep open; it is about whose agent has write access to the artifacts where work actually lives. Pair this with Anthropic’s $100M bet on credentialing the people who install AI (yesterday’s card) and the picture is coherent: they are buying both ends of enterprise distribution, the talent and the surface. Watch whether Google responds by tightening Workspace APIs or by sharpening Gemini’s own agentic editing.

Signal: The agent race just moved from the chat window into the document. The winner is not the best model; it is the one with write access to your spreadsheet. Sources: 9to5Google · VentureBeat

Anthropic opens the cyber cabinet

The expanded Cyber Verification Program (October 6) offers three tiers of reduced-blocking cyber access to qualifying security professionals (Defense, Red Team, and Specialized access), each including Claude Opus 5.5, Sonnet 5.5, Mythos 5.1, and new models going forward. The published numbers: through Project Glasswing, partner organizations found at least 129,000 verified software vulnerabilities between April and July 2026; Anthropic’s own open-source scanning found another 5,500 between April and October. More than 33,000 have been rated critical- or high-severity. The tier-efficacy test on CyScenarioBench (10 multi-stage cyber-operation challenges, 5 attempts each): Defense Access blocked 46 of 50 trials while succeeding at 4; Red Team Access blocked nothing and completed 34 of 50, matching the 67.6% success rate with no safeguards applied.

The read: this is the official answer to yesterday’s Mistral cyber card. Where Mistral ships an open model that does offensive-adjacent security work by default (82% on the reproduce-and-patch test), Anthropic gates the same capability behind vetting and data retention. Two competing theories of distributing powerful cyber models: the capability is the product vs. the capability is the liability. The refusal gap is becoming a market segment, and now it has a compliance tier system. Watch the “Enterprise Frontier Safeguards” zero-data-retention option landing this fall; the vetting and monitoring overhead is the tax Anthropic charges for access.

Signal: Two cyber strategies within 24 hours: Mistral sells the capability open, Anthropic gates it behind vetting. The closed labs’ safety posture is now a product line. Sources: Anthropic announcement

Google opens the SynthID detector to everyone

Google launched the standalone SynthID Detector to the public today (global, in English): anyone can now check whether an image, video, or audio file carries SynthID, covering outputs from Google plus partners OpenAI, NVIDIA, Kakao, and soon Apple. The scale numbers: 180 billion images and videos watermarked to date, plus 240,000 years of audio. Google’s built-in verification features in Search, the Gemini app, and Chrome already handle 1M+ requests daily.

The read: this is the provenance counter-program to OpenAI’s textGrain (researcher-gated detector, EU-only default). Google is going for ubiquity: watermark at production scale, verify everywhere, let partners ride the same rails. Two provenance philosophies: OpenAI treats detection as a compliance instrument under regulatory control; Google treats it as consumer infrastructure. The honest caveat is the same for both (paraphrase and editing degrade every watermark), but Google’s bet is that a detector everyone can use beats one only researchers can apply for. Watch the “soon, Apple” line: if Apple’s outputs join, SynthID becomes the closest thing to a universal provenance layer the industry has.

Signal: Google is turning watermark detection into a public utility while OpenAI keeps its detector behind a researcher’s application form. Provenance is splitting into compliance (OpenAI) and consumer infrastructure (Google); the open question is which one platforms actually build on. Sources: Google announcement

OpenAI’s Decisions API goes public

HN’s #5 thread (366 points, 209 comments). OpenAI’s Decisions API moved from limited preview to public beta: typed answers (predicate, choice, score) on text and images, 10x faster than the Responses API, powered by a single model, gpt-6-luna. GA is expected in the coming weeks. The pricing is the message: $0.10 per 1M input tokens, and that is the whole price (no output, cache, or other charges). It supports Zero Data Retention and HIPAA for eligible customers, with data residency in the US and Europe.

The read, against the decision-model wave (see Open Weights): OpenAI is formalizing the pattern the open-weights crowd discovered. LLM-as-judge is too expensive and too sloppy; a typed-decision primitive is the product. At 10x the speed with input-only billing, the pitch is “stop burning frontier inference on yes/no questions.” The open question is whether gpt-6-luna is a real decision model with calibrated reliability scores or a fast small model with a structured head, and whether developers will trust a closed endpoint for the reliability guarantees they need.

Signal: The “decision model” category now has a first-party API from the biggest lab. $0.10/M with no output charge is OpenAI pricing the judgment primitive like a commodity, not a model. Sources: OpenAI Decisions docs

OpenAI halves the ladder: three API tiers

Five paid usage tiers became three on October 6: Build, Launch, and Grow. The top tier now needs $500 in total API payments instead of $1,000; organizations move automatically. And the consumer repricing continues: the $200 ChatGPT Pro plan’s included usage halves on October 30 (20x to 10x the Plus allowance in Work/Codex, 200 to 100 GPT-6 Pro messages per week), price unchanged, alongside a new $500 Pro 500 tier with Ultrafast speeds.

The read: OpenAI is lowering the on-ramp and raising the ceiling in the same week. Cheaper to become a heavy API user (halve the spend to reach Grow), pricier to be a heavy app user (Pro 200 halved, Pro 500 at $500/mo). That is the classic platform squeeze: make it cheap to get hooked through the API, expensive to live in the app. The Pro 200 halving, softened with a one-time 62,500-credit grant (a stated $2,500 value, expiring December 31), reads like a subsidy being wound down because frontier models got more expensive per useful hour. Per SemiAnalysis’s math, Pro 100/200/500 now offer the same tokens per dollar, and Opus 5.5 on any Claude plan offers roughly 4x the API-equivalent value.

Signal: OpenAI is repricing both ends of the stack at once: cheaper to become a heavy API user, pricier to be a heavy app user. The platform wants your code, not your subscription. Sources: OpenAI developer community · ExplainX on the Pro repricing

A process note for the regulars: this week’s Mistral feed went [STALE] again and has nothing from its own lab; the five-lab watch leans on HN and the RSS feeds when the official pipes fail.

03 · Open Weights

The retrieval wars, on-device edition

Google’s EmbeddingGemma 2 is a 740M-parameter open multimodal embedding model (Apache 2.0) that maps text, code, images, audio, and video into one embedding space, optimized to run on a phone. The numbers: modular encoders (270M text-only, +170M vision, +300M audio), 8K context (5.5 minutes of audio, 29 images, 58 video frames), and with quantization the text-only weights fit in ~191MB active RAM on a Pixel 11 Pro. Quality: a 9.92-point jump on MTEB Code to 78.68, matching or beating many models 2x its size on text, vision, and audio. It is #4 on today’s HF trending board (823 likes), and Unsloth already has a GGUF.

Read it next to Snowflake’s Arctic Embed L (335M params, 55.98 NDCG@10 on MTEB Retrieval, ahead of Google’s Gecko 55.7 and OpenAI’s text-embedding-3-large 55.44): closed-API embedding quality is now reproducible in open weights at a fraction of the size. The retrieval layer, the unglamorous backbone of every RAG system, is commoditizing faster than generation. The strategic question: what happens to the embedding-API business when a 740M-parameter Apache-licensed model running on a phone matches it?

Signal: Generation got the headlines; retrieval is quietly going open first. When the embedding layer runs on-device for free, the moat moves to the data you index, not the model you call. Sources: Google announcement · Arctic Embed L on HF

Strands Decider 2B: the decision model goes tiny and open

HN at 239 points. Strands (the AWS agents team) released Decider 2B: a 2B-parameter open-source decision model (Qwen3.5-2B torso, LM head replaced with a ~1M-parameter pointer head, rank-16 LoRA). It answers typed yes/no/choice/score questions in a median of ~115ms on an RTX 3090, ~153ms on an M3 MacBook. Weights, code, training data, and scripts are all public; on JevBench it ranks 3rd of 33 in the 2B class.

The mechanism is the interesting part: decisions are made by scoring hidden states at the option positions against the <answer> position in a single parallel pass. No text generation, so it is fast and always returns an answer from the selected options. Strands’ demo use case is an agent “intervention system”: before any tool call, a 115ms decision model checks whether the tool’s arguments are grounded in anything the user actually said. That is a judgment too cheap to exist in a world where every check costs a frontier LLM call, and it is the open counterpart to OpenAI’s Decisions API in Lab Watch.

Read against the HF board: Cloudflare’s clef is still #1 trending (1,589), autotrust’s JEV-27B-VL #2 (1.53M downloads), GEV-26B-Decide #6, clef-flash new at #8, convaiinnovations/laya (calibrated decisions via RLCD) at #10, and vLLM’s Vela 2.0 collapsing nine safety checks into one forward pass. Meanwhile Nous Research’s new Hermes Index scores frontier models on four agentic task suites and reports dollar cost per task: the leaderboard that matters now quotes accuracy per dollar. The decision-model wave is now a full stack: 2B to 27B, open and closed, API and local.

Signal: Three weeks into the decision-model wave, it is complete at every scale: OpenAI’s $0.10/M API, Strands’ 2B open model, Cloudflare’s edge deployment. “LLM as judge” is being replaced by typed decisions at ~100ms; the cost of a judgment is falling ~100x, which is what it takes to put one inside every agent loop. Sources: Strands announcement · HN thread

The board this morning

Qwen/Qwen3.8-27B extended the all-time like crown to 17,170. Lightricks/LTX-2.5 holds #9 at 6,745. New climbers not yet established: deepseek-ai/DeepSeek-V4.1-Flash (4,210 likes, 1.26M downloads), prism-ml/Ternary-Bonsai-2-27B-gguf (2,510 likes, 4.27M downloads), and google/embeddinggemma-2 (debut at #4 trending, 823 likes). Losing trending momentum vs their like counts: the classic image-gen gravity wells (FLUX.1-dev, SDXL) are stable but absent from the trending board; image generation’s likes are mature, while decisions and retrieval are climbing.

04 · Money Flow

Anthropic courts founders with a $50K+ bundle

Anthropic expanded Claude Startups on October 6: approved founders get $1,000 in API credits, a free year of Claude Team (five Premium seats), office hours with the Applied AI team, and the new Claude Startup Stack, partner offers from Linear, Lovable, ElevenLabs, Granola, and Hex worth up to $45,000. That is a $50K+ package per startup; eligibility is startups founded in the last five years or funded in the last two, and the program already counts thousands of members.

The read: this is the land grab for the next generation of AI-native companies before they commit to a platform. Anthropic’s own line (per its announcement): most people will get AI’s benefits through the companies built on top of models, not the models themselves, so partnering with founders is central. Read against OpenAI’s tier repricing in Lab Watch: both labs are racing to make their stack the default for new companies, one with cheaper on-ramps, the other with subsidized tooling. The founder-credit wars are the new cloud-credit wars, and the winners get baked into companies that outlive the subsidy.

Signal: The startup-credit wars are the cloud-credit wars replayed: subsidize the tooling, own the default stack. The $45K partner bundle is Anthropic admitting the moat is the ecosystem, not the model. Sources: Unite.AI

05 · Leaderboard

OpenRouter: the board reshuffles again

Sorted by tokens this morning: Command Code surged to #1 at ~2.33T tokens (up from ~263B yesterday, a 9x jump that reads like a window or attribution change), Hermes Agent ~1.52T, Claude Code ~1.25T, Framer ~1.20T, ZCode ~673B, Kilo Code ~568B, Pi ~528B, Codex ~508B, Cline ~475B, Freebuff ~470B, Proto Agent ~398B, Nous Research API ~348B, omp ~243B, OpenClaw ~194B, OpenHands ~179B, DeepSeek Harness ~159B, Strix ~104B, ISEKAI ZERO ~78B.

Signal: The only stable fact on this board remains its instability: #1 changed hands again, and this time via a suspicious 9x jump. Watch the two-week trend, not today’s headline. Sources: OpenRouter rankings

Artificial Analysis: Mistral Large 4 under the microscope

Artificial Analysis published its Mistral Large 4 Preview evaluation on October 6, the follow-up to yesterday’s One Thing; the Cyber Index detail (top five globally, 82% on the reproduce-and-patch test) is the number to keep. Yesterday’s card stands: the refusal gap is the story, and the independent evals keep filling in the sheet.

06 · The Argument

Boston Dynamics hires the man who built Alexa

Boston Dynamics named Rohit Prasad CEO, effective today. Prasad spent 12 years at Amazon as SVP and Head Scientist for Alexa and AGI: he built Alexa from its earliest days into a service in hundreds of millions of households, then led the Amazon Nova foundation model family. Before Amazon, 14 years at Raytheon BBN on ML research. His mandate, per the company: accelerate the integration of AI and robotics and expand real-world Physical AI applications.

The read: this is the clearest signal yet that Physical AI is becoming the next platform war, and the leadership market agrees. Boston Dynamics did not hire a robotics lifer to run robots; it hired the executive who turned an AI research breakthrough into a mass-market consumer product and then into an enterprise model family. That is exactly the trajectory Boston Dynamics wants for Atlas and Spot: research darling to industrial product. Note the timing context: the company reportedly plans an IPO next year and Atlas deployments at Hyundai plants starting 2028. The robot shop is dressing for the public markets with an AI-platform CEO.

Signal: When the world’s most famous robot company hires the man who shipped Alexa to hundreds of millions of homes, the message is that robots are now a distribution problem, not a research problem. Sources: Boston Dynamics announcement

Google Playground: type a game, play it

Google Labs launched Playground today, an experimental browser platform that turns plain-text prompts into playable 2D and 3D games, no coding required. The read: this is vibe-coding crossing into a new medium. Websites were the first prompt-to-artifact success; games add state, rules, and real-time interaction, a meaningfully harder generation target. The question it poses: when the game is generated on demand, who owns it, who moderates it, and what does “shipping a game” even mean? Watch whether this becomes a creator economy (shareable games, the scriptless Roblox) or a tech demo that teaches Google’s models what playability means.

Signal: Prompt-to-website was the demo; prompt-to-playable-game is the thesis test. Games require rules, state, and fun, and you cannot hallucinate your way through a physics loop. Sources: Google announcement · HN thread

07 · Papers

Agent in a Bottle: Can LLM Agents Turn Their Capabilities Into Cheap, Scalable Artifacts?

Sonthalia et al. ask the amortized-intelligence question directly: LLMs can solve many narrow tasks, but querying them separately for millions of related instances is prohibitively expensive. Can agents autonomously create cheaper solutions, bottling their own capabilities into artifacts (classifiers, decision rules, distilled models)? Read against today’s decision-model wave: this is the academic statement of what Strands and OpenAI are productizing. The frontier pattern: solve it once with the expensive model, bottle it into the cheap artifact, run it forever.

AdvSim2Real: Training Web Agents Against Adaptive Prompt Injection in a Web World Model

Hashmi et al. train web agents against adaptive prompt injection inside a web world model: the injection problem treated as an adversarial training environment rather than a filter bolted on afterward. Read against the prompt-injection discourse: the field is converging on “train in the attack” instead of “guardrail the attack.”

ScienceClaw: Benchmarking Continual Self-Evolution of AI-for-Science Agents

Zhang et al. benchmark whether AI-for-science agents turn verified executions into persistent program-level improvements: continual self-evolution across the natural and social sciences. Read next to the One Thing: the math-paper flood is the artifact; ScienceClaw asks whether the agents that produce it get smarter from doing so.

When Forgetting is not Catastrophic: On the Mechanics of Spurious Forgetting

Palit et al. on spurious forgetting: knowledge a model appears to forget during finetuning often remains stored and recoverable. Finetuning on new facts can even produce forgetting that looks catastrophic but is not.

Signal: Two papers this week attack the economics of intelligence from opposite ends: Agent in a Bottle asks how to amortize expensive capability into cheap artifacts; AdvSim2Real asks how to train the attacks away instead of filtering them. The field is moving from “bigger model” to “better amortization.” Sources: Agent in a Bottle · AdvSim2Real

08 · GitHub Trending

rea: reverse-engineer anything with agents

Now 12,521 stars, +4,666 today, the week’s runaway. Reverse-engineer anything, from app behavior down to native binaries, with agents. The pitch keeps landing because it inverts the usual agent story: instead of agents writing new code, they dissect existing code. In a world of inherited codebases and opaque binaries, “explain this to me” is the more valuable direction.

The boards

Daily: boykopovar/AnyPS5 (+2,725, 8,885 total), automatic PS5 executable porting to Linux and Windows. mattpocock/skills (279,062, +1,406 today), the skills genre is now the biggest star magnet on the platform. ayghri/i-have-adhd (54,822, +620), a skill to stop your coding agent from burying the answer. cathrynlavery/diagram-design (44,607, +828), editorial diagrams for agents (“no Mermaid slop”). cloudflare/security-audit-skill (25,737, +538). addyosmani/agent-skills (102,486, +453). tester-army/e2e (7,052, +1,391). DuarteSantos8/openGym (6,519, +1,494). Weekly: Panniantong/Agent-Reach (+6,454, 93,025), eyes for agents across the internet. NVIDIA/OpenShell (+5,228, 15,231), the safe private runtime. heygen-com/hyperframes (+3,614, 58,336). mvschwarz/openrig (+3,327, 5,656), agent teams with roles and shared context. pablostanley/yoinks (+2,947, 5,017), yoink any video from the terminal. cursor/plugins (+1,106, 10,199).

09 · Product Hunt

Yesterday’s top 10 launches

1. OpenBot · Grok Bot alternative: free, local, open-source, multiplayer (331 votes)

2. AUDR by Chargebee · Open standard for tracking agent run costs (319 votes)

3. Fuse AI · Build a Custom GTM Stack. One SDK. One MCP. (299 votes)

4. Incredible · Vibe computing. Control your computer with your voice. (272 votes)

5. Ghostifier · Companies have your data. Ghostifier gets them to delete it. (109 votes)

6. NoteWorthy · Notes supercharged with AI, all on-device (109 votes)

7. iphone-use · Let AI agents drive a real iPhone, even apps with no API (108 votes)

8. ruOS · A cloud desktop where AI agents do the work for you (100 votes)

9. Scumble · The open-source editor for AI inpainting (100 votes)

10. Doco · Good Company. Music that fits. (95 votes)

Three worth a click: AUDR by Chargebee · an open standard for tracking agent run costs is the cost-observability layer every agent company eventually builds; Chargebee is trying to own it before the billing gets weird. OpenBot · free, local, open-source Grok Bot alternative; the open clone of the closed hit keeps winning votes. iphone-use · agents driving real iPhones, even apps with no API; computer-use is going mobile.

10 · Hacker News

Mistral Large 4

1,936 points, 1,159 comments, still the #1 thread. Covered in yesterday’s One Thing; the comments are still pricing the 4-GPU inference claim and arguing about whether the cyber benchmarks are the real story (they are).

Sharing AI progress in mathematics

1,088 points, 1,149 comments. Covered in The One Thing.

EmbeddingGemma 2: An open, lightweight multimodal embedding model

387 points. Covered in Open Weights.

Decisions API is in public beta

366 points, 209 comments. Covered in Lab Watch.

OpenTPU: an open-source AI accelerator, developed by AI

326 points, 371 comments. An open-source AI accelerator designed by AI itself. The silicon loop is closing: AI designing the chips that AI runs on, and the design is open.

Claude Code’s suggested message feature

247 points, 145 comments. The argument: the real customer of the suggested message is the model, not the user; every nudge shapes the next prompt into something the model can complete. Read against the agent-attention theme in The Argument.

Strands Decider 2B: a small, open-source, decision model

239 points. Covered in Open Weights.

Also notable

Shipping JPEG XL in Chrome (developer.chrome.com, 248 points): the codec wars end with Chrome shipping JPEG XL. Big web news; wrong newsletter, mostly. Nobel Prize in Chemistry 2026 (nobelprize.org, 169 points): Henri B. Kagan and Kenso Soai. Congratulations; wrong newsletter. Write Like It’s 1866: LLMs Relearn Telegraphese (44 points): a fun experiment; the comments are the delight.

11 · OpenRouter Apps

The full board this morning

Command Code (~2.33T, new #1), Hermes Agent (~1.52T), Claude Code (~1.25T), Framer (~1.20T), ZCode (~673B), Kilo Code (~568B), Pi (~528B), Codex (~508B), Cline (~475B), Freebuff (~470B), Proto Agent (~398B), Nous Research API (~348B), omp (~243B), OpenClaw (~194B), OpenHands (~179B), DeepSeek Harness (~159B), Strix (~104B), ISEKAI ZERO (~78B). Four apps between 475B and 1.25T, separated by noise; the top slot changed hands again, this time via a 9x jump that deserves a skeptical eye.

Signal: The coding-agent tier is a sentiment market with a memory: Hermes keeps compounding on persistence across sessions while the launch-spike apps trade places weekly. Attention is the commodity; retention is the position. Sources: OpenRouter rankings

12 · Watch

WesRoth: “OpenAI’s secret model just BROKE math…”

WesRoth · the math episode. Watch against The One Thing: the numbers behind the 722-proof release, hype-filtered.

TwoMinutePapers: “DeepMind’s New AI Just Cracked The Code Of Life”

TwoMinutePapers · the AI-for-science episode. Watch against the One Thing: two “AI does science” results in one week, and the verification standards differ.

Fireship: “PewDiePie is setting AI free… and OpenAI is furious”

Fireship · the open-weights culture war, YouTube edition. Watch against Open Weights: Strands Decider and EmbeddingGemma are the technical version of this argument.

13 · Listen

Lenny’s Podcast: “OpenAI’s Head of ChatGPT: We’re entering a new era of AI (again)”

Lenny’s Podcast · Oct 4 · the ChatGPT leadership conversation. Listen against The One Thing: the official story (new era, again) and the 722-proof release, same week.

Hard Fork: “A.I. Agents: Cute, Cuddly and Maybe Catastrophically Dangerous?”

Hard Fork · Oct 2 · the agent-safety episode. Listen against The Argument: cute, cuddly, and now editing your documents.

14 · Read

Anthropic’s Cyber Verification Program, in full

The 129,000-vulnerability numbers and the CyScenarioBench tier test are worth reading against yesterday’s Mistral cyber card: two theories of distributing cyber capability, published 24 hours apart.

The API tier announcement

The Build/Launch/Grow details and the $500 top threshold, straight from OpenAI’s developer community.

15 · Safe to Ignore

Permission to skip

Skip: The GitHub DMCA thread. 415 points of cracked-software takedown drama. Not AI news. Skip: Nobel Prize in Chemistry. Congratulations to Henri B. Kagan and Kenso Soai; not this newsletter. Skip: JPEG XL in Chrome. 248 points of codec joy. Admire, then close the tab. Skip: The Commodore 64 keycap font. 229 points of retro delight. Wrong newsletter. Skip: The 15-meter underground tunnels house. 300k for tunnels. Wrong newsletter. Skip: Rust’s derive often implies inline. Compiler trivia. Wrong newsletter.

16 · Build on This

Provenance Receipts: the audit layer for the AI-content era

This week’s theme is provenance going mainstream: Google opened its SynthID detector to the public, OpenAI watermarked textGrain for EU compliance, and publishers are caught in the middle. The labs sell the watermark; nobody sells the paperwork. Provenance Receipts is the micro-service in between: drop in any file or article, and get back a report (which detectors flag it, what the confidence is, which model families claim it), plus a signed, timestamped receipt and an embeddable “how this was made” badge for your pages. The insight from this week is the pricing logic: detection degrades under paraphrase, so do not sell certainty; sell the audit trail. Price it pay-as-you-go at $0.05 per check, no subscription; trust and compliance budgets are episodic, and a subscription for a checkbox is how you lose to the checkbox. Start with newsrooms and agencies (they have EU exposure and lawyers), expand to marketplaces and anyone selling content to a skeptical buyer. Regulation created the demand; the product is just the receipt.