FOMO Buster #012: The conversation is the inventory

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 launches ads inside ChatGPT

OpenAI published “Building advertising for the way people use AI” this morning — the official launch of advertising inside ChatGPT. The announcement introduces a new visual ad format in ChatGPT alongside expanded measurement tools, attribution partnerships, and brand-suitability controls for advertisers. The feed timing is the story: this lands one week after DevDay (GPT-6 Astra, dots), days after the Atlantic resignation essay, and in the same news cycle as Meta’s Muse topping the App Store — a product Meta’s own design essay describes as “funded by the ad business.” The industry has converged on the same answer to the free-tier question: the conversation is the inventory.

The strategic read: OpenAI spent years as a subscription-and-API company; ads mean it’s now in the attention business, which changes every incentive downstream. Watch three things: whether ads stay visually separate from answers (the trust boundary), which attribution partners get named (that’s the real customer list), and whether the free tier’s economics start shaping model behavior — the moment the answerer and the ad-seller share a balance sheet, “advertising doesn’t influence answers” becomes a claim that needs auditing, not asserting.

Signal: The free AI tier was never going to be funded by goodwill. Meta proved the ad-funded AI assistant tops the App Store; OpenAI just joined the business. The next competition isn’t model quality — it’s who sells the conversation without breaking the trust that makes the conversation valuable. Sources: OpenAI announcement

02 · Lab Watch

Cloudflare’s Web Search API: the pipe, not the product

Cloudflare’s Web Search API went into beta on October 2 (164 points, 91 comments on HN this morning) and the pricing structure is the whole story. It runs through AI Gateway, is billed at each provider’s list API price with no additional markup, and you can bring your own provider API key. The launch providers: Ceramic.ai, Exa, and Linkup — all three committed to Zero Data Retention for requests through Cloudflare and to Cloudflare’s verified bot crawling standards. Read this as the commoditization playbook applied to search-for-agents: when grounding became infrastructure, Cloudflare’s move was to sell the aggregation layer at cost and own the gateway logs. The billable primitive is the gateway, not the search.

Signal: Search is becoming a commodity input, and the winners sell the pipe. Zero markup plus Zero Data Retention is how you win procurement against every “AI search API” startup charging a premium on someone else’s index. Sources: Cloudflare changelog

Gemini 4 Argon lands on Product Hunt with a point to prove

Gemini 4 Argon took #2 on yesterday’s Product Hunt (262 votes) with the tagline “Google’s frontier model for careful reasoning & complex work” — and it arrives backed by Artificial Analysis’s September 30 verdict: Google is back as one of the top three labs in intelligence achieved. The model now has its own AA model page (High variant evaluated September 30). The Product Hunt listing is the distribution move worth noting: frontier models on PH is a genre that barely existed a year ago; now it’s a launch channel. Watch whether Argon converts launch energy into OpenRouter tokens or stays a benchmark darling.

03 · Open Weights

Cloudflare’s clef: the decision model, open-sourced

The #1 and #6 models on today’s HF trending board are Cloudflare/clef and Cloudflare/clef-flash — and they’re the most interesting release of the week. Clef is a 27B multimodal decision model, post-trained from Qwen/Qwen3.8-27B, released under Apache-2.0 with a full Jev/SystemOne-compatible API. It doesn’t generate text: you feed it a state (text, JSON, images, or video) plus a schema of typed questions — true/false, named-choice, or ordered-score — and it returns calibrated probabilities for every allowed option of every question in a single forward pass. No sampling, no output parsing, no regex on prose.

The numbers come from Cloudflare’s own Decision Index 0.2.1 suite (36 benchmarks — take vendor-run suites as directional, not gospel): clef beats Jev on the headline agent-relevant rows — BFCL 98.5 vs 95.8, ToolRet 69.2 vs 65.3, BANKING77 94.2 vs 79.7, ContractNLI 81.4 vs 71.7, CRUXEval 86.7 vs 73.0, ForecastBench Brier 13.9 vs 17.4 (lower is better). The latency row is the product: median 209.3ms for clef, 38.8ms for clef-flash, against 524.1ms for Jev. And on four end-to-end business workflows (invoice processing, customer service, security incidents), clef matches or edges Jev.

Now hold those numbers next to Red Hat’s benchmark (see The Argument): on guardrails, a fine-tuned DeBERTa-v3-base hits 89.01% at 54.1ms — on a MacBook M1 CPU — beating Jev’s 86.35%, while Qwen3.6-35B as LLM-as-a-judge tops the board at 89.31%. The picture that emerges has three layers: (1) fine-tuned micro-classifiers (fastest, cheapest, need labeled data), (2) decision models (zero-shot, guaranteed schema, one forward pass), (3) LLM judges (most flexible, most expensive). Cloudflare just gave away a best-in-class layer-2 model with permissive licensing, Jev-compatible APIs, and a hosted leaderboard — aimed squarely at TypeSafe’s Jev and every “AI judge as a service” startup. Cloudflare owns the pipe (Workers AI); now it owns a reference model for the pipe. The decision model is the structured-output endgame, and it’s being priced at zero.

Signal: The quantize-and-distill era is over; the distill-and-contract era is here. The valuable artifact is no longer the biggest model but the smallest model that returns a guaranteed contract — a probability, not a paragraph. Cloudflare just made that artifact free and open, which is what the “LLM-as-a-judge” API businesses should be worried about. Sources: Cloudflare/clef on HF · Decision Index · Red Hat: decision models vs guardrails

Strata: the week’s biggest HN thread is a local-inference installer

The #1 thread on HN this weekend (862 points, 387 comments): Strata, running Qwen3.8-Flash-Next (125B) on a gaming PC — now at 12,988 stars, up from ~9.7K when this newsletter first covered it four days ago. The headline number: the README says an RTX 3090 (24 GB) should write 100–140 tokens/sec, and the thread’s benchmarkers are posting 100+ tok/s on RTX 4090s. The mechanism (24,576 tiny experts, ~10 active per token, hot experts on GPU, all of them in RAM, SSD lookup table, speculative decoding at 1.6–1.8x) hasn’t changed — what’s changed is the audience. This is no longer a story for quant hobbyists; it’s a story for everyone who double-clicks a .bat file, lets their AI assistant run the installer, and points their coding agent at localhost:8080 with a Claude Code override. The installer is the product, the quantization is the plumbing.

The board rhymes with it: Cloudflare/clef and clef-flash are #1 and #6 trending. convaiinnovations/laya holds #2 trending (5,204 likes, 11,733 downloads) — the open-source Jev alternative that Red Hat just benchmarked. autotrust/JEV-27B-VL is a new climber (#12 trending) — typed decisions, calibrated probabilities, the SystemOne wave. And CUA-S1 (trycua’s “System 1” decision models for computer use) closes the loop: decision models are now the shared substrate of guardrails, routing, and computer use.

Qwen/Qwen3.8-27B took the all-time like crown last week and extended it to 16,982 (+71 in a day — still climbing, not cresting). Lightricks/LTX-2.5 sits at #11 all-time (6,418 likes): the video models are cementing as gravity wells. New climbers not yet established: PSRben/VisionHOPE (beats vision transformers with adaptive memory, per the RSS feed), ISTA-DASLab’s Qwen3.8-Flash-Next GSQ-RCO GGUF (2.2M downloads — the quant Strata is built on), nvidia/Nemotron-3-Diarization (687 likes, 55K downloads — diarization is having a moment), and TaichuAI/ZDTaichu5.0-9B.

Signal: The open-weights leaderboard is reorganizing around a single question: how small can the guaranteed answer get? 125B on a 12 GB card, 27B decisions in 39ms, Jev-style verdicts from a diffusion model. The frontier isn’t bigger anymore — it’s cheaper to ask. Sources: Strata on GitHub · HF trending · trycua/cua

04 · Money Flow

RobCo becomes Europe’s robotics unicorn

227 points, 202 comments. The facts, per Reuters/WSJ: Munich-based RobCo passed a $1 billion valuation — doubling in nine months from the ~$500M January Series C ($100M co-led by Lightspeed and Lingotto). The mechanism of this round matters: it’s a $40M secondary share sale, mostly employees selling to existing investors (Sequoia, Lightspeed, Lingotto, Leitmotif) and new ones (Cherry Ventures, European Tech Collective). The product beat: Alfie, a two-armed, self-learning robot for industrial settings, launching in March; 1,000+ robots already sold, including to BMW. And the geography: CEO Roman Hölzl is relocating to the US — Austin and SF operations, customers in more than a dozen US states. The WSJ context: robotics and physical-AI startups raised $33.4B in H1 2026 — more than all of 2025, per PitchBook. Read the secondary sale as the tell: employee liquidity at 2x in nine months means late-stage robotics is liquid again, which means the exit clock is running. The physical-AI wave is now funded like the LLM wave was in 2024 — the question is whether factory floors adopt like developers did.

Signal: Robotics capital is moving at software speed now — 2x in nine months, employee secondaries, founder-led funds. The constraint isn’t money anymore; it’s whether the robots survive contact with real factories. Sources: Reuters · WSJ · EU-Startups

Huawei and Qualcomm sign a broad patent license

107 points, 66 comments. The smartphone patent cold war thaws another degree: two of the largest wireless patent holders on earth signing a broad cross-license. The comments are doing the math on what this means for the next decade of royalty flows — and for the smaller handset makers who paid full freight while the giants negotiated. Watch: patent peace usually precedes a standards fight, not follows one.

05 · Leaderboard

OpenRouter: the board measures attention, again

The churn continues. Hermes Agent reclaimed a clear #1 at ~1.99T tokens (+6%). Framer retook #2 at ~1.22T, Kilo Code #3 at ~1.21T. Claude Code slipped to #4 at ~1.13T (-10%) — one week after taking the overall #2 spot. Cline holds ~1.04T, Freebuff ~815B, omp ~801B, Strix ~610B, and Codex sits flat at ~522B (#9) — its spike fully digested. Four apps now sit between 1.04T and 1.22T, separated by noise. Nobody has held #2 for more than a few days in weeks. The only stable fact on this board is its instability: the coding-agent tier is a sentiment market, and sentiment re-prices daily.

Artificial Analysis: Grok 4.7 (Low) joins the value tier

The week’s notable new evaluation: Grok 4.7 (Low) — released September 21, scoring 42 on the Intelligence Index against a price-tier median of 26, with $2.00/$6.00 per 1M input/output tokens (blended ~$1.35/M), a 500K context window, and 73.6 tok/s output. The strategic read: the low-reasoning tier is where the value war is being fought — 60% above the tier’s median intelligence at 40% below the median output price. Also in the AA feed: the Gemini 4 Argon evaluations (September 30) behind the “Google is back as one of the top three labs in intelligence achieved” verdict, and Upstage’s Solar Mini 4 (October 1).

Signal: Two leaderboards, two stories: OpenRouter measures the crowd’s attention this week; Artificial Analysis measures where the value is settling. The value is settling in the low-reasoning tier, at $1–2 blended per million. Sources: OpenRouter rankings · Grok 4.7 (Low) on AA

06 · The Argument

Turn off Apple Intelligence on macOS 27 and get its disk space back

687 points, 456 comments — the weekend’s second-biggest thread, and it’s a rebellion in a shell script. RemoveMacAI turns off Apple Intelligence on macOS 27 and deletes its downloaded models: 1,362 stars since September 29. The detail that made the thread: macOS 27 no longer has a single off-switch for Apple Intelligence, and the models stay on disk after the features are turned off. The tool’s method is careful — a configuration profile using Apple’s own restriction keys, model removal through Apple’s asset service, System Integrity Protection stays enabled, nothing under /System touched, and the profile redirects each model’s re-download to a closed local port so macOS can’t fetch it again. Fully reversible. The kill list: Siri, Genmoji, Image Playground, the ChatGPT extension, Writing Tools, Mail/Messages/Safari/Notes summaries, smart replies, inline predictions, Spatial Photos, Photos Clean Up, Xcode predictive completion.

The thread is 456 comments of disk-space grievances and privacy audits — the debloat genre, but aimed at AI specifically. The read: Apple ships AI as infrastructure you can’t see and can’t remove, and the response is a market. Yesterday’s RemoveMacAI is tomorrow’s business: every OS vendor that removes the off-switch creates demand for someone else to sell it back.

Signal: The AI tax is now measurable in gigabytes. When the platform owner ships intelligence as undeletable infrastructure, the delete button becomes a product — and 1,362 stars in a week says it’s a product people want. Sources: RemoveMacAI on GitHub · HN thread

Decision models meet their skeptics

Red Hat’s AI Safety team did the experiment the decision-model hype needed: benchmarking Jev against traditional guardrails (129 points, 46 comments). Nine guardrails, two risk domains (prompt injection, content safety). The results: Qwen3.6-35B as LLM-as-a-judge wins at 89.31%; a fine-tuned DeBERTa-v3-base classifier hits 89.01% at 54.1ms — running on a MacBook M1 CPU; DiffusionGemma-as-Jev 87.72%; Jev itself 86.35%; Laya 85.44%. So the headline holds — decision models don’t beat the judge, and they don’t beat a purpose-built classifier either. But the footnote matters: Jev-style models beat Mistral’s Shieldstral (72.02%) and run zero-shot, no labeled data required. The honest read: decision models own a real niche — zero-shot, fast, schema-guaranteed — and Cloudflare’s free clef just made the niche free. Read this next to the clef card in Open Weights: the skeptics are right about the ceiling, and Cloudflare is right about the floor.

07 · Papers

Finetuning with Sampling: SFT Learns Better Than You Think

The post-training orthodoxy says RL generalizes, SFT memorizes. This paper (Karan, Chen, Du — submitted October 1) says the orthodoxy is measuring the wrong thing. The method: an MCMC sampling algorithm that progressively transforms off-policy expert traces to be more on-policy for a reference model — instead of changing the learning objective to fit the data, it reshapes the data distribution to fit the learner. Across scientific skill acquisition, mathematical reasoning, and open-ended expertise, this lets SFT rival strong on-policy baselines — generalizing better and forgetting less. The conceptual line: sampling as “a model-native operator that shapes data for learnability,” a primitive for the whole post-training stack. If the result holds, the cheapest lever in post-training isn’t the RL algorithm — it’s the data’s distance from the policy.

Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents

Wavestone’s research team did what nobody else had: a source-code anatomy of eleven production coding harnesses — Claude Code, Codex CLI, Gemini CLI, Mistral Vibe, OpenHands, Aider, Mini-SWE-Agent, Hermes, Pi, OpenCode, OpenClaw — roughly four million lines of Python, TypeScript, and Rust, plus a longitudinal quarter-over-quarter re-pin of the original eight. The findings: seven canonical subsystems (loop, tools, context management, safety, orchestration, extension, and one more the paper names), 29 recurring design patterns, and two stunning absences — no agent runtime imports a general-purpose agentic framework, and none retrieves code with vector embeddings; the field runs on hand-rolled async loops and deterministic retrieval. Skills lead MCP in adoption (9/11 vs. 8/11), and behavioral policy is migrating from prompt prose into configuration. The thesis: in H1 2026 the coding harness completed its turn from tool to platform. The paper closes with 18 design recommendations and a 90-line minimum-viable-harness scaffold — the whole discipline, distilled to a page.

Signal: The paper layer keeps rhyming with the product layer within weeks: skills beating MCP in real harnesses is the empirical version of every “skills are the new API” argument. And “no vector embeddings for code retrieval” should end a thousand architecture debates.

Trust the Critic More

The RL orthodoxy says learned critics are too inaccurate to trust — use them for baselines only, roll every trajectory to its terminal reward. Actor-Critic with Action Chunking (AC2) disagrees: assign credit to action chunks — short continuations of past-trajectory prefixes — and let a learned critic score the state at the end of each chunk, so the policy updates without ever seeing a terminal reward. Three reliability tricks: local readiness (critic updates only on problems where the critic is accurate), reference solutions from past successful rollouts, and 10k-token chunks instead of token-level credit. The result: training Qwen3-4B on FineProofs-RL, AC2 beats GRPO’s 18.5% peak on IMO-ProofBench using 2.5x fewer decoding FLOPs — 25% fewer steps, fewer tokens per step. The read: the critic was always the cheaper oracle; we just didn’t trust it at the right granularity. Chunk-level credit assignment is a design space that was sitting there unexplored.

Signal: Two papers this week say the same thing from opposite ends: stop fixing the algorithm, fix what the algorithm sees. SFT-with-sampling fixes the data; AC2 fixes the credit horizon. The post-training frontier is observability, not objectives.

08 · GitHub Trending

image-blaster: a photo into a 3D world in five minutes

9,384 stars for the newsletter’s top story: an image-to-world skillset for Claude. Drop a photo in input/, tell Claude to “blast it,” and in under five minutes you get 3D models (.glb, .obj) of the dynamic objects via Hunyuan 3D, a Gaussian splat (.spz) of the static environment via World Labs’ Marble 1.1, and ambient + physics SFX (.mp3) via ElevenLabs — orchestrated entirely through Claude skills. The output embeds in Unity, Unreal, Godot, Blender, or Three.js. The interesting architecture: the product is the skill package — the orchestration layer is text files for Claude, the heavy lifting is APIs. “IMAGE-BLAST your childhood bedroom” is the demo line, but the real story is jumpstarting 3D work: environment art’s cold-start problem solved by a skill pack.

AgentCraft: your coding agents, inside Minecraft

Created October 3, already at 349 stars: a team of Claude agents doing real work on your code, inside a Minecraft studio you can walk around in. You type a goal; a lead agent (Marlow) plans and pins tasks to a wall; workers walk to their desks and code in their own git worktrees while monitors stream their logs; when they need you, an agent walks to the podium with a question. You review the real diff and press Merge — nothing touches your branch without it, nothing is ever pushed (enforced inside git itself, not a command filter). Yesterday’s Product Hunt had Cubicle — “a live office for your agents, read-only by design.” This is the other answer to the same question: not a dashboard, a place. Permission prompts as gameplay. A measured small run cost about $6. The read: agent observability is becoming a UX genre, and the genre is splitting — passive monitoring (Cubicle) vs. walkable command centers (AgentCraft). The winner is whoever makes “what is it doing in there” feel safe instead of surveillance-y.

cua: the full computer-use stack

28,121 stars and the week’s sleeper: trycua/cua ships the whole computer-use ladder — Cua Spaces (full desktops for agents, macOS VMs on your Mac, free for individuals), Cua Driver (inspect and operate native apps on macOS/Windows/Linux), Lume (local VMs on Apple Silicon), the Cua SDK, Cua Bench (tasks, evals, trajectory export), and CUA-S1 — “System 1” decision models for computer use, i.e. the decision-model thread from Open Weights applied to clicking buttons. Read it next to the Latent Space episode in Listen (“How OpenAI shipped its Jev competitor in 1 Week”): the computer-use race now has an open stack, a benchmark, and a decision-model layer. OpenAI’s dots made computer use a product; cua is making it a platform.

On the boards: tester-army/e2e holds daily #1 (4,020 stars, +1,430 today) — the e2e-testing reinvention continues. DuarteSantos8/openGym (+1,444/day): a self-hosted gym tracker doing real numbers. boykopovar/AnyPS5 (+994/day): porting PS5 executables to Linux and Windows. Panniantong/Agent-Reach (91,569, +1,156/day): eyes for agents across the whole internet. Weekly: VoiceStudio (fully-local ElevenLabs alternative, +14,689 this week), paperclip, vectorize-io/hindsight (agent memory that learns, +10,623), heygen-com/hyperframes, and DietrichGebert/ponytail (155,662 stars — the lazy-senior-dev agent).

09 · Product Hunt

Yesterday’s top 10 launches

1. CoreSpeed — One MCP for everything your agents need: apps, memory, tools (317 votes)

2. Gemini 4 Argon — Google’s frontier model for careful reasoning & complex work (262 votes)

3. Smooth Recorder — Mac screen recordings that look edited when you hit stop (222 votes)

4. Blume 2.0 — The open-source docs framework for humans and agents (193 votes)

5. DocsAlot MCP Connector — Maintain your help-center by asking Claude, Cursor or Codex (160 votes)

6. ChatGPT Space — Create, collaborate, and build together with AI in Space (153 votes)

7. opensend.cc — The open source email platform that runs on your server (150 votes)

8. Octri.dev — Generate customizable Docs, SDKs, MCP server, and monitoring (123 votes)

9. Pass Designer — Design Apple Wallet passes and see the real card (109 votes)

10. WikiFix for Confluence — Find and fix issues in your knowledge base (93 votes)

Three worth a click: CoreSpeed — “one MCP for everything” is the aggregation land grab; whoever owns the agent’s tool registry owns the agent’s wallet. Smooth Recorder — “looks edited when you hit stop” is the Loom-killer thesis in one line; the screen recorder is the last un-AI’d creative tool. opensend.cc — self-hosted open-source email, 150 votes on day one: the subscription-fatigue vote, expressed as a product.

10 · Hacker News

Run Qwen 3.8 Flash Next (125B) on consumer hardware at 100 tokens/sec

862 points, 387 comments — covered in Open Weights. The thread’s velocity is the story: four days, 862 points, 12,988 stars.

Turn off Apple Intelligence on macOS 27

687 points, 456 comments — covered in The Argument. The debloat genre goes AI-native.

Europe’s new robotics unicorn: RobCo hits $1B

227 points, 202 comments — covered in Money Flow.

Cloudflare Web Search API

164 points, 91 comments — covered in Lab Watch.

Apple and a Hacker’s Future

121 points, 105 comments — covered in Read. Pairs with RemoveMacAI: the essay the thread needed.

Also notable

Denmark’s CPR registry breach (304 points, 237 comments): 8.8M people’s data exposed — a national-registry breach, real news, wrong newsletter. A browser-native VB6 IDE (326 points): retro joy, not news. Doom ported to SQL (220 points): the database-joke genre, thriving. Tippett Studios’ digital archive (188 points): VFX history, not AI news. Nobel Prize in Medicine 2026 (83 points): congratulations to the laureates; not this newsletter. GrapheneOS vs. the Pixel 11 (130 points): Android security standards, not this newsletter.

11 · OpenRouter Apps

The full board this morning

Hermes Agent (~1.99T, up — clear #1), Framer (~1.22T, retook #2), Kilo Code (~1.21T, up), Claude Code (~1.13T, slipped to #4), Cline (~1.04T, flat), Freebuff (~815B), omp (~801B), Strix (~610B), Codex (~522B, rank #9 — the spike fully digested), Command Code (~419B), Pi (~389B), DeepSeek Harness (~279B), OpenClaw (~192B), OpenHands (~141B), ZCode (~108B). Four apps between 1.04T and 1.22T, separated by noise; #2 has changed hands three times in two weeks.

Signal: The coding-agent tier is now a four-way statistical tie measured in attention, not adoption. The only number that compounds on this board is Hermes’s — and it’s compounding on persistence (memory across sessions), not launch spikes. Sources: OpenRouter rankings

12 · Watch

TwoMinutePapers: “The Billion Dollar AI Advantage Is Disappearing”

TwoMinutePapers · Oct 5 — the efficiency thesis in video form. Watch after Open Weights: clef, Strata, and the decision-model wave are the receipts.

WesRoth: “OpenAI’s employee’s WARNING”

WesRoth · Oct 4 — the resignation-essay discourse, continued. Watch against The One Thing: the official story (ads!) and the exit interview, same week.

TheAIGRID: “Googles New Gemini 4 Argon is Now The Worlds Smartest AI”

TheAIGRID · Oct 1 — the Argon hype episode. Watch against the AA verdict in Lab Watch — hype vs. measurement, the eternal double feature.

13 · Listen

Why Dwarkesh is Wrong about Computer Use + How OpenAI shipped its Jev competitor in 1 Week — Latent Space

Latent Space · Sep 30 — the computer-use episode of the week, and the title is doing real work: OpenAI’s dots as the Jev competitor. Listen against Open Weights (clef, CUA-S1) and GitHub Trending (cua, AgentCraft): the decision-model layer and the computer-use layer are converging, and everyone is shipping fast.

Academia is for Ambition — Alex Zhang, MIT — Latent Space

Latent Space · Oct 2 — the compute-for-academia conversation. Pairs with the Wavestone paper in Papers: the harness discipline is being named in industry, but the training runs that feed it still live or die on academic compute access.

14 · Read

Apple and a Hacker’s Future

121 points, 105 comments. Stratechery on Apple’s platform control and the hacker counterculture — read it as the essay behind the RemoveMacAI thread in The Argument. When the platform ships AI as undeletable infrastructure, the hacker ships the delete button. The two stories are one story.

Clef decision models — Cloudflare blog

The announcement behind the clef release in Open Weights. Read the vendor’s own framing against the Decision Index numbers and Red Hat’s skepticism — the full decision-model debate, primary sources only.

15 · Safe to Ignore

Permission to skip

Skip: The Denmark CPR breach. 304 points of genuine news; the FOMO is not. Skip: The Nobel Prize in Medicine. Congratulations, laureates; not this newsletter. Skip: Doom in SQL. 220 points of database humor. The genre thrives; the news does not. Skip: The VB6 browser IDE. 326 points of retro joy. Admire, then close the tab. Skip: The Tippett Studios archive. VFX history is not AI news. Skip: GrapheneOS vs. the Pixel 11. Android security, wrong newsletter. Skip: The ceiling fans. Last week’s rule stands: real science, wrong newsletter.

16 · Build on This

OffSwitch: the AI-feature auditor your OS won’t ship

This week’s 687-point RemoveMacAI thread proves the demand: Apple removed the single off-switch for Apple Intelligence, the models sit on your disk eating gigabytes, and a community script is doing the platform’s job. OffSwitch turns that script into a maintained product for a general reader: an auditor that inventories every AI feature running on your machine — on-device and cloud — and shows you, per feature, what it costs in disk, what it costs in battery/CPU, and what data it sends where. Then it gives you one safe, reversible off switch per feature, with changes that survive OS updates. The receipts are the product: before/after disk numbers, per-feature network logs, plain-language explanations of what each toggle actually kills. Charge $12 one-time, per major OS version — an off-switch sold as a subscription is a contradiction. Ship Mac first (the thread is the market), Windows second (Recall is the same wound), and let every surprise-gigabyte postmortem become your landing page. The platform vendors keep removing the off-switch; the business is selling it back, maintained, audited, and honest.