FOMO Buster #023: Open weights, closed frontier

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

Meta’s Muse Spark Open: 816 billion parameters, a license with teeth, and the two-tier future

On October 15, Meta released Muse Spark Open, an 816B-total / 42B-active mixture-of-experts model built for coding and long-horizon agentic work, and published the weights openly on Hugging Face. The specs: 62 layers, 256 experts with 8 routed per token, a 1M-token context window, natively multimodal input (text, image, video, audio), trained on data from more than 100 languages. 880 points on Hacker News, the day’s runaway thread.

Two numbers tell you what Meta is actually selling. First, the hardware story: the model runs on two Mac Studios with M5 Ultra via MLX (NVFP4 checkpoint, roughly 530 to 700GB), which means a four-figure desktop pair now hosts an 816B-parameter model with 42B active per token. Full BF16 still wants about 2TB of VRAM. Second, the license: the Muse Model License reads permissive (use, modify, distill, sell) but carries two strings. Products must display “Built with Muse Spark Open,” and anyone offering the model as a service above $20M/year in revenue must sign a separate agreement with Meta first. That clause is the price of the weights.

Meta’s own benchmark table is the honest part. Spark Open leads the compared open-weight field on agentic benchmarks: JobBench 61.8, MCP-Atlas 89.2, HLE-Diamond 26.1, ZeroBench 50.0, SimpleVQA 73.6, IntPhys2 68.3. But on the Artificial Analysis Intelligence Index it scores 39, behind GLM 5.3 (45) and Kimi K3 (44), and on Terminal-Bench 2.1 it trails at 76.4 against Kimi K3’s 85.0. The safety numbers are the eye-catcher: on StrongREJECT v2 (adaptive jailbreaks) it scores 0.5 against 60 to 86 for every competitor, and it refused 99.2% of cyber jailbreaks where Kimi K3 refused 75.5%. Whether that is careful post-training or benchmark-friendly tuning is the open question.

The strategic read is the two-tier announcement. Chief AI Officer Alexandr Wang said Meta “will generally not open-source its most powerful models”: Muse Wonder stays closed, Spark Open is the open-weight contender, and the line between them is drawn in the license. Four credible open-weight flagships now sit within a few points of each other (GLM 5.3, Kimi K3, Qwen3.8-2.4T, Spark Open) plus Mistral Large 4. The open frontier is a real race again, and Meta just made the Mac Studio a first-class citizen of it.

Signal: The weights are open, but the business model is not: the $20M Model-as-a-Service clause means Meta monetizes the second-order winners. Watch whether the agentic-benchmark lead (JobBench, MCP-Atlas, HLE) survives independent reruns, and whether Spark Open or Kimi K3 becomes the default open agent backbone. Sources: the research announcement · the Hugging Face model card · HN thread, 880 points

02 · Lab Watch

OpenAI takes GPT-Rosalind to academia

GPT-Rosalind is OpenAI’s dedicated life-sciences model (frontier reasoning plus the Rosalind workbench for medicinal chemistry, genomics, and wet-lab assistance). The new Rosalind Academic Plan gives eligible U.S. academic labs up to 30 Business Standard seats free for 12 months, with GPT-Rosalind behind a trusted-access approval: identity and institutional verification, a designated safety contact, ongoing monitoring. Research data shared through the plan is not used for training by default. API access is not included. The buried detail is the comparison table: on an antiviral-therapeutics assessment, Rosalind delivers the candidate comparison matrix and modeling plan while Mainline Astra withholds the actionable parts. OpenAI is drawing the line between “general frontier model” and “the model that will tell you how to design the drug,” and gating the second one behind a PI’s signature.

Signal: Domain models are becoming access products, not API SKUs. Trusted-access science is a new distribution tier, and 30 free seats is priced to make Rosalind the default lab stack before competitors field one. Sources: OpenAI’s announcement

Anthropic embeds its auditors

Anthropic has given employee-like access to a six-person METR team: company hardware, office space, internal tools on par with its own risk-assessment teams. METR publishes without Anthropic’s veto (redactions allowed only in narrow, predefined categories, which METR may publicly characterize). SecureBio is reviewing the chemical-biological chapter of the August Risk Report; Accenture and Hugging Face are in earlier-stage talks. The frame is three tiers: scoped investigations (publish with approval), publish-without-approval, and the embedded tier METR now holds. Anthropic’s analogy is bank supervision: resident examiners inside the largest banks.

Signal: Voluntary embedded evaluation is the strongest governance move any lab has made, and the publication-without-veto term is the part that matters. The test is whether METR’s first published assessment reads like an audit or a press release. Sources: Anthropic’s update

Google’s magic-state breakthrough: 40x fewer errors on real hardware

Published in Science this week: Google demonstrated magic-state cultivation on its Willow processor, cutting errors 40x versus un-cultivated states and hitting a record 99.99% logical fidelity for T-states. The mechanism: inject a noisy state into a small region of the chip, run coordinated measurements to catch and discard flawed attempts, “cultivate” a pure low-error state, then graft it from a distance-3 color code to a distance-5 surface code. Why it matters: magic-state production is the resource hog of fault-tolerant quantum computing, with estimates that up to 95% of a quantum computer’s capacity could go to making them. A 20-year-old bottleneck now has a working hardware demo.

Signal: This is the quantum equivalent of a process node: not a bigger qubit count, a better yield on the expensive part. Watch for the scale-up attempt; cultivation at distance-3 is a proof, not a product. Sources: Google’s post

Also on the lab feeds

Anthropic’s Code Review is now generally available for Claude Team and Enterprise: the multi-agent PR review system (parallel find, verify, rank) exits research preview with reviews running 3x faster (20 minutes down to 6-10), Quick/Standard/Deep depth tiers per repo, and at least $10,000 in review credits for eligible Enterprise orgs. Anthropic runs it on nearly every internal PR. Google published a crisis-resilience white paper on preparing for natural disasters. OpenAI added a Sullivan & Cromwell legal customer story. Meta and Mistral were quiet beyond Spark Open.

Signal: Code Review GA is the first time a lab has productized “the model reviews the model’s code” as a managed service with credits attached. The 3x speedup is the number to watch; 20-minute reviews were a demo, 6-minute reviews are a workflow. Sources: Unite.AI on Code Review GA · Google crisis resilience · OpenAI x Sullivan & Cromwell

03 · Open Weights

NVIDIA’s Nemotron 3.5 Super VL: the open model built for agents that never sleep

NVIDIA’s new open model is explicit about its job: enterprise agents running continuously. 120B parameters with only 12B active per token (MoE), distilled from Nemotron 3 Ultra to match its accuracy at roughly a fifth of the parameters. The numbers: up to 5x the throughput of similar-size open models on agent tasks (per the AA leaderboard), 630 tokens/s at 100k input tokens, CharXiv 80.8% (84% with Python), and with NVIDIA’s Dynamo serving stack, 8x the concurrent users with 85% lower time to first token on 16 B200s. Trained on 25 trillion tokens; weights, datasets, training recipes, and RL environments all published. Two Bolt siblings ship alongside: a 74B-A9B squeezed into 24GB of local memory via 2-bit quantization, and a 4GB voice-assistant model coming later.

The read: this is NVIDIA selling hardware with open weights as the marketing. Every Nemotron release is tuned for NVIDIA silicon (NVFP4, Blackwell, DGX Spark), and the coalition behind it (Mistral, Thinking Machines Lab, Perplexity, Black Forest Labs) reads like a buyer list. The enterprise pitch is the fine-tuning story: your data, your servers, no provider ToS.

Signal: The open-model pitch has flipped from “free weights” to “your weights, our GPUs.” NVIDIA publishing the datasets and recipes is more open than most open weights, and every one of them is a coupon for a B200. Sources: The Decoder

FLUX.3 Image Klein: 5.6B open weights for text-to-image and multi-reference editing

Black Forest Labs shipped a 5.6B open-weights image model (transformer; text encoder separate) covering text-to-image, single-reference edits, and multi-reference compositions (up to 10 reference images) in one model, with Qwen3-VL-4B as the text encoder. Two signatures: the distilled checkpoint runs in 16 steps with no guidance, and prompts take bounding boxes on a 0-1000 grid for layout control. Hardware: 31GB BF16 / 26GB FP8 / 25GB NVFP4 on a B200 at 1024px, plus an 18.8GB prompt upsampler (a separate 9B model) that expands short prompts into full scene captions. Licensed under the FLUX Kommunity License (non-commercial, with a commercial-output exception under $5M revenue). The ComfyUI variant is trending alongside it.

Signal: BFL is productizing the pipeline, not just the weights: upsampler, bbox layout language, FP8/NVFP4 variants, ComfyUI build. The 16-step distilled default is the Turbo-checkpoint pattern, now standard. Sources: the Hugging Face model card

The board this morning

google/embeddinggemma-2 holds #1 trending (MTEB Code 78.68, the semantic-everything model). jialinyyzz/humanizer is #2: a text-generation model with 1,796 likes and 85k downloads climbing fast. Qwen3.8-27B (17,830 likes) and deepseek-ai/DeepSeek-V4.1-Flash keep their gravity-well status; Qwen3.8-Flash-Next (6,271 likes, 2.4M downloads) is the download monster. meta-models/Muse-Spark-Open-816B-A42B debuts at #10 trending with 390 likes and 815 downloads on day one. The uncensored-quant lane persists (Qwen-Image-2.1 quants at #6 and #8). Lightricks/LTX-2.5 and fishaudio/drama-3 hold video and voice. LiquidAI/d1-3B and Cactus-Compute/whistle (ASR) round out the movers.

Signal: The trending list now reads as a stack: embeddings, humanizers, image/video, voice, ASR. The model is becoming a component, not a destination. Sources: Hugging Face trending

04 · Money Flow

CoreWeave’s credit market is pricing a coin flip; Huang says relax

CoreWeave’s five-year credit default swaps hit a record 1,194.5 basis points, implying a 60%+ chance of default within five years (ICE Data Services), while its 9.75% 2031 bonds trade at 79.5 cents on the dollar (yield above 15%). At those CDS levels, insuring $10M of CoreWeave debt costs about $1.19M a year. On the same day, Jensen Huang told Bloomberg in Stockholm that Nvidia sees no need to backstop CoreWeave and that its “cost of capital is actually decreasing.” Nvidia held 47.2M CoreWeave shares at the end of June and has agreed to buy unsold CoreWeave capacity through 2032. ARK bought the dip.

The tension is the whole AI-capex story in one frame: the credit market prices the debt stack as distressed while the equity narrative (and NVIDIA’s) says the GPUs will pay for themselves. Somebody’s model is wrong by a factor of two, and the CDS market is rarely the sentimental one.

Signal: The AI trade now has a public credit-vs-equity spread you can watch daily. If the CDS keeps climbing while Huang keeps talking, the question stops being “is there a bubble” and becomes “whose balance sheet absorbs it.” Sources: AIStockWire

Devin Oncall: the agent takes the pager

Cognition launched Devin Oncall on October 15: Devin now auto-investigates alerts and bug reports, traces causes through the customer’s systems and code, and opens tested PRs for engineer review. The numbers: 100% end-to-end success on SREGym (their SRE benchmark, saturated, fastest time to mitigation), already handling thousands of production alerts a day for customers pre-launch. The case studies do the selling: at Reducto, Devin spotted concerning trends and opened an incident before the production alerts fired; at RV Tech (the Rivian/VW joint venture), it pulls vehicle IDs and timestamps from customer reports, parses logs, and triages, redirecting engineers back to product work. Works out of the box with Datadog, Dynatrace, and Sentry.

Signal: The coding agent is moving from the repo to the incident queue, and the unit of sale is shifting from seats to mitigations. SREGym at 100% means the benchmark is already dead; watch for customer-verified MTTR numbers. Sources: Cognition’s announcement

05 · Leaderboard

TheoPhysEval: a physicist’s day of work now costs $2.40

Thirty-five researchers (ICTS, Harvard, KU Leuven, Google DeepMind) built TheoPhysEval: 50 unpublished, research-level theoretical physics problems across 11 subfields, each with a verifiable numerical answer, the kind of “turn-the-crank” calculation a subfield expert does in about a day. GPT-6 Astra in Codex solved 49 of 50 (98%) at $2.40 a problem, $116 for the full set. Claude Opus 5.5 in Claude Code: 46/50 (92%). Gemini 3.8 Flash: 68%. DeepSeek-V4.1-Flash: 54%. GLM 5.3: 42%. Without web search, Astra and Opus both land at 92%. Two details elevate this above a leaderboard: the problems were designed so answers could not be copied from the source papers, and in some cases where the agent “failed,” the experts’ own reference solutions turned out to contain errors. The authors’ caveat: these are calculations, not discoveries, and the agents still lack “taste.”

The strategic read: the marginal cost of routine theoretical physics just fell to the price of a coffee, and the failure mode inverted. The agent is now good enough to grade the physicists.

Signal: This is the third “agent does the expert’s routine work” result this week (after Third Moment’s 10,000-agent fleet and SoSBench’s under-aligned open models). The pattern is set: benchmarks are moving from “can it answer” to “can it do the job, at what cost, and who checks the checker.” Sources: The Decoder · the arXiv paper

Also moving

Artificial Analysis published Beam 1.0 (model page): Reflection’s Beam model now has a public AA listing; it was already the comparison column in Meta’s Spark Open benchmark table. Cursor added Muse Wonder at $4/$20 with a 1M-token context window (300k default, 1M in Max Mode), scoring 46.1% on CursorBench, about 11 points behind the leader. Krea 3 launched image editing at 8 seconds for $0.06.

Signal: The model-picker economy keeps expanding: Cursor is now an audition stage for every frontier lab, and the benchmark that matters is the editor’s own. Sources: Artificial Analysis: Beam · CursorBench coverage

06 · The Argument

Meta’s new terms: a business website becomes its training data, unless the account is deleted

Starting October 30, Meta’s new commercial terms let it use scrapers to crawl content from websites “associated with” business accounts on Facebook, Instagram, WhatsApp, and Messenger, and use that content to train its AI products. Three legal experts reviewed the terms for Nieman Lab: as written, the terms could be used to argue Meta can scrape news sites and train on them, and there is no opt-out short of deleting the business account. Virtually every news organization has at least one Meta business account, including Nieman Lab itself. Media attorney Matt Topic: “a remarkable provision to bury in terms of service, given that this would give Meta access to material that would likely cost many millions of dollars in a licensing market.”

The detail that makes it an argument and not just a story: Meta already pays a few major outlets (CNN, Fox News, USA Today) for AI licensing deals with explicit contracts and fees. The new terms get the rest of the web’s publishers for free, by contract, as a condition of using the platform for marketing. Publishers that reject the terms risk losing commercial access to the platforms. And Meta’s new Meta One subscription charges creators to include links in posts, links Meta can now scrape for free. 408 points on Hacker News, 147 comments.

The read: Meta spent the same week opening its model weights and closing its data-acquisition loop. The open frontier needs training data, and Meta just wrote itself a contractual pipeline to the open web, priced at zero, enforced by platform dependence.

Signal: The AI licensing market is bifurcating: explicit paid deals for outlets big enough to sue, and terms-of-service extraction for everyone else. October 30 is the deadline for every business with a Meta account to decide whether its website is Meta’s training data. Sources: Nieman Lab · HN thread, 408 points

07 · Papers

Quantum circuits, discovered by language models

“Discovering state-of-the-art quantum circuits with large language models” (Oct 15): LLMs designing quantum circuits that match or beat the state of the art. Read against Google’s magic-state result above: one lab is improving the hardware’s error budget while another paper shows models designing the circuits themselves. The loop is closing from both ends.

Signal: Quantum advantage is becoming a software problem. The labs that pair better qubits with model-designed circuits will compound faster than either advance alone. Sources: arXiv 2610.22367

Also on arXiv

Rethinking Routing Exploration in Modern Mixture-of-Experts (2610.22353): routing is the new architecture battleground, and it is the week of the MoE (Spark Open’s 256 experts, Nemotron’s 12B-active routing). FDE-Gym: The Last Mile of AI Delivery (2610.22350): a benchmark for the deployment end of the agent pipeline. MLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning Engineering (2610.22344): the “agent that does ML engineering” benchmark family keeps growing. LIBERO-PeRM: Benchmarking Personalized Robotic Manipulation (2610.22372): robot manipulation benchmarks go personal. T2-z: Test-Time Optimization of Environment Latents for Robotic World Models (2610.22364): world models that adapt their environment latents at test time. Lean Graph Search for Formal Theorem Proving (2610.22339): graph search meets Lean.

08 · GitHub Trending

The safety and skills stack

• perplexityai/SPACE · Perplexity’s open-sourced agent sandbox: Firecracker microVMs per task, Apache 2.0, zero VM-to-host escapes in 108 red-team attempts · open-sourced Oct 15
• anthropics/claude-plugins-official · Anthropic’s official directory of Claude Code plugins · +1,552 today · 40,024 total
• debpalash/VoiceStudio · Open-source, fully-local ElevenLabs alternative: voice cloning, dubbing, dictation, transcription in 646 languages · +2,301 today · runs on CUDA, Apple Silicon, or CPU-only
• xerj-org/xerj · AI search over data with autoindex for agents · +1,647 today
• K-Dense-AI/scientific-agent-skills · Turn any AI agent into an AI scientist · +1,158 today
• affaan-m/ECC · Agent harness performance optimization: skills, instincts, memory, security · +1,057 today
• obra/superpowers · An agentic skills framework and software development methodology · +973 today
• MadsLorentzen/ai-job-search · The job search that runs on your machine, an AI job application framework · +879 today
• HKUDS/CLI-Anything · “Making ALL Software Agent-Native,” a CLI hub for agents · +716 today
• ruvnet/ruflo · Deploy multi-player agent swarms · +672 today
• mksglu/context-mode · Context window optimization for coding agents, 98% tool-output reduction via sandboxing · +449 today
• cursor/plugins · Cursor plugin specification and official plugins · +214 today

Signal: Perplexity open-sourcing SPACE is the tell: the sandbox is now table stakes, and the race is on to own the reference implementation. The skills lane (plugins, scientific skills, superpowers) keeps compounding underneath it. Agents are getting a seat at every table; the question is whether the table is a VM. Sources: Perplexity SPACE · GitHub trending

09 · Product Hunt

Yesterday’s top products (Oct 15)

1. Talanoa - Email that feels like chat (307 votes)

2. Navige - Cut AI agent costs and control runaway spending (242 votes)

3. Apple HomeView - A new way to stay connected, and manage the home (202 votes)

4. GLM-Legal - Z.ai’s legal AI model for contract review and compliance (194 votes)

5. OpenMuse - Opensource, self hostable personal assistant for any harness (177 votes)

6. Daddy Issues - Think ahead. Fight less. A weekly family briefing for dads. (125 votes)

7. Polylane Impact Analysis - Catch risky pull requests before they break production (112 votes)

8. withHuman - An open-source human approval layer for AI Agents (104 votes)

9. Ubuntu 26.10 (Stonking Stingray) - GNOME 51, Rust coreutils, and a refreshed Linux desktop (97 votes)

10. Chatway MCP - Let ChatGPT & Claude answer customer support tickets via MCP (96 votes)

The interesting cluster is the agent control plane: Navige (cap runaway agent spend), withHuman (human approval gates for agents), and Polylane (catch risky PRs before they break prod) are three answers to the same question: who watches the agent, and what does the invoice look like? OpenMuse (self-hosted personal assistant for any harness) is the open counterpoint. Ubuntu 26.10 ships GNOME 51 and Rust coreutils, the rare OS release on a startup leaderboard.

10 · Hacker News

Muse Spark Open

880 points, 441 comments. Covered in The One Thing. The thread’s best material is the license debate: the $20M Model-as-a-Service clause is getting priced honestly against the Llama-era 700M-MAU clause, and the Mac Studio angle is doing real distribution work. Meta figured out that “runs on the computer you already own” is worth more than another 3 points on a benchmark.

Neural networks are clean

760 points, 295 comments. A Goodfire research essay arguing the field misread the evidence for superposition: what looked like neurons folding unrelated concepts together was mostly recursive generalization, a fractal grouping of related things, and with better tools and modern architectures, roughly 90% of neuron poles in a Qwen3-32B audit have one main job. The InceptionV1 last layer went from “almost no interpretable neurons” to 99-100% interpretable with ordinary architectural improvements. Pair with chenglou/handwired (GitHub, created yesterday, MIT): it compiles ordinary JavaScript (if-else, loops, state) into matrix multiplication plus reglu over floats, no training, exact, fully differentiable. Code and weights are converging from both sides.

Meta’s new commercial terms will allow the scraping of news sites

408 points, 147 comments. Covered in The Argument. The comments are doing the contract math: consent-by-continued-use plus no opt-out equals a data pipeline that no court has blessed and no publisher can refuse without losing the platform.

Also notable

What I self-host (588 points, 373 comments): the annual state-of-the-homelab thread, and the open-weights week is all over it (Muse Spark Open on two Mac Studios is the new aspiration purchase). Celebrating 70 Years of Fortran (299 points): the language that outlived every prediction of its death. BIPM decides to stop adding leap seconds to UTC (238 points): time itself gets a breaking change. RPKI now protects 70% of internet routes (198 points): the internet’s plumbing keeps getting quietly safer.

11 · OpenRouter Apps

The board this morning

Command Code holds at ~4.17T tokens, flat, the baseline. The move below it: ZCode (~1.92T) has retaken second place from Hermes Agent (~1.68T), which slipped after yesterday’s pass on Claude Code; Claude Code itself sits at ~1.26T. Kilo Code (~848B), StepCode (~695B), Cline (~505B) hold their lanes; pi (~436B) keeps climbing.

Signal: The #2/#3 slot keeps changing hands between the lab-built wrappers, which says the coding-agent token market is still contestable week to week. Command Code’s flat 4.17T is the one number nobody is contesting. Sources: OpenRouter rankings

12 · Watch

Fireship: “The runtime wars are over… and Node won”

Fireship · Oct 15 · not an AI story on its face, but the runtime underneath every agent harness is. The agent stack (OpenCode, Claude Code, the whole skills economy) runs on JS runtimes, and Fireship’s verdict on who won the war sets the platform the next generation of harnesses builds on. Watch for the infrastructure, not the thumbnail.

The quiet morning

The AI channels are between stories: TheAIGRID’s fresh episode is interview drama around an Altman remark (skip the clickbait, keep the policy questions), and the DwarkeshPatel queue is all history this week. The Muse Spark Open takes will land tomorrow; the two-Mac-Studios setup video is the one to wait for.

13 · Listen

From the queue

No fresh AI episodes this morning: Lenny’s podcast and Latent Space are both still on this week’s episodes (featured yesterday), and Hard Fork’s new episode is about Elizabeth Holmes, which is off-brief. The backlog still holds “Why AlphaFold Didn’t Solve Protein Folding,” the Periodic Labs episode on synthesis superintelligence, and Hard Fork’s “Anthropic’s Quest to Give A.I. Morals.”

14 · Read

Neural Networks Are Already Clean

The Goodfire essay behind the HN thread: superposition was misread, recursive generalization was the real motif, and with modern tools the network is legible all the way through. The interpretability essay of the year so far.

GPT-6 Astra and the $2.40 physicist

The TheoPhysEval writeup behind the Leaderboard: 50 unpublished problems, the cost math, and the cases where the agent graded the physicists.

Muse Spark Open: the model card

The full spec sheet behind The One Thing: the 256-expert architecture, the benchmark table against five open peers, the safety evaluations, and the license text. Read the license section before the benchmark table.

How Google cultivated magic states

The quantum post behind Lab Watch: the cultivation protocol, the 40x error reduction, and why magic states are the resource bottleneck.

Handwired: JavaScript as a neural net

The repo behind the HN pairing: if-else and loops compiled to matmul plus reglu, exact and differentiable, with live demos. The weirdest and most fun hour of reading this week.

15 · Safe to Ignore

Permission to skip

Skip: The AlphaSignal newsletter’s “Meta releases AudioCraft” story. That library is from 2023; the newsletter misfiled it as new. The repo link is real, the news is not. Skip: The AWS webinar and the summit promos in the newsletter. Sponsored placements, not editorial signal. Skip: The “run uncensored models locally” how-to. Instructional content, not news, and the wrong newsletter’s beat. Skip: Hard Fork’s Elizabeth Holmes episode. 0 AI content in the brief. Wrong newsletter. Skip: The crow that brings thank-you gifts (619 points of video). Delightful. Wrong newsletter. Skip: DwarkeshPatel’s history series this week. Aztecs, plagues, and the fall of Spain. Excellent, wrong newsletter. Skip: “Scotland’s EV charging gap is about income.” Government policy, wrong newsletter. Skip: Meta donating 1,000 AI glasses to Singapore’s disability community. Corporate PR, not news. Skip: “Apple routed a graphite pipe through the iPhone Duo’s hinge” (171 points). Hardware trivia, wrong newsletter. Skip: The xAI exclusion stands. Per standing policy, xAI announcements are not covered; if the news is big enough it will surface on the aggregators.

16 · Build on This

The terms-of-service canary

Meta’s new commercial terms take effect October 30: keep using the business platform and your website becomes Meta’s training data, with no opt-out short of deleting the account. Every publisher, SaaS company, and indie business with a Meta business account just became a party to a contract they did not negotiate, and Meta is only the first platform that will try this.

The terms-of-service canary watches the fine print so businesses do not have to. Point it at the platforms a company depends on (Meta, Google, Shopify, Stripe, Cloudflare), and it diffs their terms continuously, flags clauses that grant data, scraping, or training rights, and translates each one into plain language: what you are giving up, what it would cost you, and what the opt-out path actually requires. A weekly clean bill of health when nothing changes; an incident feed with a countdown when a deadline like October 30 appears. The customer is every business that cannot afford a legal team but cannot afford to ignore this either: indie publishers, SaaS founders, e-commerce shops. Pricing follows the unit of risk: per domain, pay as you go, no subscription, with a premium tier that drafts the objection letter. The moat is the corpus: every flagged clause makes the next detection faster, and the canary’s badge becomes the trust mark customers check before signing up for a platform. Yesterday’s Build on This watched ad slots; today’s watches the contracts. The attack surface moved up the stack.