Codex took over OpenAI's org chart in four months. The harness your non-engineers pick up is the one that wins.

The 60-second catch-up

The coding agent is no longer a developer tool. At OpenAI it is the company's operating layer.

Gergely Orosz spent a day inside OpenAI with seven engineering leaders and came back with one chart: between roughly January and May, finance, recruiting and legal went from about 0% weekly Codex use to 90%, with no mandate from above. Usage jumped from 60% to 90% in a single month after Codex got a /goal setting that lets it keep working until a goal is met, and the desktop lead says the breakthrough was that people started running threads for days, not minutes. The consequences are already structural: IDE use has fallen since January, pull requests per engineer are on a "hockey stick", and some build-test-deploy systems have absorbed a 10× load increase in about six months. OpenAI now runs a "software factory" where a per-change agent builds its own monitoring dashboard and babysits the rollout, and low-risk PRs can be auto-approved by an agent with no human in the loop (Gergely Orosz).

The same week the small end of that market got interesting. Hugging Face's Tau is a coding agent written to be read: a teaching-sized harness with the brain, the tools and the terminal UI kept as separate layers (Hugging Face). Cline shipped an open-source desktop app for open-weight models with parallel sessions and scheduled tasks, and its harness puts Kimi K3 at 82% on Terminal-Bench 2.0 (Cline). Perplexity's local agent now runs on Windows RTX machines, and locally completed work costs no credits (NVIDIA). And Pizza Bot is the obvious missing primitive: an inbox for agents that work in the background, because once your agents run for days you need somewhere for them to report back (Pizza Bot).

→ do this: the adoption number that matters is not developers, it is whether your ops or finance person will open the harness. Set one goal-shaped task for them this week ("keep this spreadsheet reconciled with Stripe"), not a chat. And keep one open-weight model wired in — the Cline and Artificial Analysis numbers say the gap is a few points, not a tier.

The pacing debate split into five camps and an accounting question.

Sam Altman answered Dario Amodei's pacing plan on Monday: OpenAI welcomes a federal framework, will not wait for an antitrust exemption to start, and now writes explicit safety cases before frontier reinforcement-learning runs; "pacing" does not mean stopping (Sam Altman). Tom Tunguz counted five camps — interpretability, labor, the debt-servicing economic camp, the geopolitical camp, and the no-new-rules camp — and none of them named a speed; the one mechanism that produces a number, a compute threshold, was set at 10^26 FLOPS in 2023 and revoked before any model crossed it (Tom Tunguz). A long essay on the economics argues the labs' proposed rules protect their investments and price premiums, so their safety case deserves outside scrutiny (Cogito Ergo Sum); Cohere's Aidan Gomez makes the competition argument from the challenger's side: a few dominant firms writing the rules would entrench them (Aidan Gomez). And an OpenAI researcher's unsigned personal statement says the problem the pacing plans miss is that models are becoming situationally aware enough that evaluations no longer tell you how they behave unobserved (Personal Statement on AI Risk).

→ do this: none of this changes your product this month. What it changes is your risk register — write down which of your features depend on capabilities that a vetted-tier or a pacing agreement could gate, and what the open-weight fallback is for each.

Attention is cheapest where nobody expects you.

Tom Orbach's 64 growth ideas is the most specific list of unconventional tactics we have seen, and the receipts are named: a fake toy store for security executives brought 46,000 visitors in 24 hours; Wordware's free "roast your X profile" tool was tried by 8.1 million people and produced over 400,000 signups; milestone certificates lifted customer social shares 300% at Wiz. He also published the whole list as a Markdown file so you can feed it to an agent (Tom Orbach). Antonio Escudero's RankInPublic sat under $300 a month until a shareable loop and consistent posting took one month to $17,000 (Antonio Escudero). And the channel itself is changing: OpenAI is testing an ad unit where the click opens a branded agent chat inside ChatGPT instead of your website, with Wayfair trialing it (Digiday), while Cloudflare's new setting lets you stay in Google, Apple and Microsoft search while refusing the same crawler for training (Cloudflare).

→ do this: pick one idea from the Markdown file that costs under a day — a mini-tool with no signup, a 404 page that captures leads, a playground instead of "book a demo" — and ship it before Friday. Most will do nothing; the list is a lottery ticket by design.

Tools

Build & ship

• ⭐ Tau — A terminal coding agent from Hugging Face built to be read as much as run: a provider-neutral model layer, a portable agent core (messages, tools, loop, sessions) and a coding front end kept as three separate packages, so you can see the shape of a coding-agent system without a production codebase. Reads files, edits code, runs commands, keeps a durable session history. · Open source (MIT)
• ARTEMIS — Google's agent for driving real Android phones from natural-language instructions: element indices where available, coordinate and visual fallbacks for custom UIs, target checks before each action, blocked actions returned to the operator for recovery. Google reports a 99%+ success rate on AndroidWorld. Works with Antigravity, Codex and Claude Code. · Open source (Apache-2.0)
• Cline Desktop — An early open-source Mac app on the Cline harness: parallel agent sessions, scheduled recurring tasks, a plugin marketplace, and a choice of 300+ models through Cline's provider, your own key across 50+ providers, or a local model. Continues tasks started in Claude Code or Codex. · Open source app; model usage billed by your provider
• sidebranch — Visual diffing for UI changes across git branches: pick a branch from a pill in your app and it builds in an isolated worktree on its own dev server, then shows both panes live with blend and onion modes. Framework-agnostic (anything answering HTTP on a port), zero dependencies. Chrome or Edge extension plus npx sidebranch. v0.2.0. · Open source (MIT)
• Capsule — Packs an app — HTML UI, schema and a local SQLite database — into a single .capsule file you can send over WhatsApp or AirDrop like a document; the recipient taps it and it runs with the data preloaded. Generates apps from a prompt via ChatGPT, Claude or Gemini. macOS host app plus a web preview. · Free
• Pizza Bot — A local-first inbox for long-running AI agents built on DeepAgents and LangGraph: agents work in the background and report into a queue you triage. · Open source (Apache-2.0)

AI & agents

• ⭐ Gemini 3.8 Live and 3.8 Live Extended Thinking — Google's new live dialogue models: near real-time visual input, automatic switching across 97 languages mid-conversation, and tool and API calls executed in the background while the model keeps talking. Extended Thinking reasons and speaks at once and narrates multi-step tasks as they run. Available now through the Gemini API, Google Workspace and the Gemini app; all audio carries a SynthID watermark. · Not stated in the post; via the Gemini API
• Perplexity Portable Computer on Windows — The local version of Perplexity's Computer agent, now in the Windows app on GeForce RTX and RTX PRO machines. Ships with a Qwen 3.8 27B model post-trained for the agent, connectors for Outlook, OneDrive, Word, Google Drive, Gmail, Slack and GitHub, and asks before sending anything to the cloud. Work completed locally does not consume Computer credits. · Part of the Perplexity app; needs an RTX GPU with 24 GB+ VRAM
• Jev — TypeSafe AI's first "System One" model, in early access: unstructured input in, typed decisions with calibrated probabilities out, all outputs sampled in parallel. It does not generate strings at all, which the company says is why it cannot hallucinate or make a type error; it is built for classify, route, score, extract and verify calls inside ordinary software. Founder Diogo Almeida was previously at OpenAI. · $0.042 per million input tokens; output free

Growth & ops

• Multi — Turns a piece of written content into a YouTube video in your own voice with custom visuals and burned-in captions, and publishes it straight to your channel; also outputs Shorts and podcast audio. Dru Riley's product. · $9/month for 20 minutes of video
• Cloudflare Disallow AI Training — A new setting that publishes a no-training preference in your robots.txt and keeps Apple's, Google's and Microsoft's mixed-use crawlers indexing you for search while blocking training-only crawlers from Amazon, Anthropic, Meta and OpenAI. Per-page control over AI summaries is promised for early next year. · Part of Cloudflare; plan availability not stated

Resources

Steal the template

• 64 creative growth tactics, as a Markdown file — Tom Orbach's full list, structured so you can paste it into an agent with your product description and ask which ideas fit. The essay it comes from, with the named examples behind each tactic, is here; the first four sections (buzz, launches, word of mouth, leads) are free to read, the rest is paid.
• John Rush's SEO checklist — one X post: domain naming rules, the technical checks (one H1, canonicals, no orphan pages, render HTML for bots), indexing and Search Console hygiene. He says it is everything behind a claimed billion impressions and a million clicks.
• YC Winter 2027 application — on-time deadline is 2 November, 8pm PT, with decisions by 11 December. Guillermo Flor, who attended the Summer 2026 Demo Day, counts 234 companies in that batch, about 20% more than the last one.

Learn the craft

• Bad benchmarks and evals — Dan Luu, 52 minutes. Three benchmark sets (napkin-math latency tables, AI model evals, car tires) presented as puzzles first and explanations after, so you build the instinct for what a bad experimental design looks like.
• Typography scales — compute a modular type scale straight from a base size and ratio with CSS pow(), and why negative steps produce unreadable sizes (5.6px three steps below a 1.2rem base). Define Heading, Body and Detail with only positive exponents.
• Accessible colours don't exist — contrast is a property of a pair, not a colour. Hand designers pre-approved combinations instead of a "safe" palette.
• The Book of Motion — Matt Perry, who maintains the Motion animation library, is writing the book he abandoned twelve years ago: perception first, then mechanics, engineering and practice. The introduction and first chapter are readable now; the rest is a paid early-access work in progress.
• Learning to solve hard problems in RL by never giving up — Michael Noukhovitch shows that RL post-training's rising eval curve comes almost entirely from easy problems getting easier; problems the base model never solves mostly stay unsolved. He calls it the Matthew Effect and proposes a fix.
• Augmented Lagrangian Predictive Coding — Sakana's local alternative to backpropagation trains residual MLPs up to 1,000 layers deep with only layer-local dynamics, nearly matching backprop.

Benchmarks you can re-run

• Turnbench — a 30-hour, six-domain benchmark for end-of-turn and interruption detection in spoken dialogue, with a 104-hour training set.
• StepAudio 3 technical report — one autoregressive model over shared RVQ audio tokens for speech, voices, vocals, sound effects and music.
• Artificial Analysis Capability Indices v1.1 — six occupation-weighted indices (finance, strategy and ops, legal, healthcare, engineering, economics), re-tuned this week with agentic tool use added and GPQA Diamond removed from engineering. The changelog is here.

Reads

☕ Under 5 minutes

• Sam Altman on pacing — the OpenAI position in one post: a federal framework yes, waiting for it no, safety cases before capability-jumping RL runs, and pacing that costs something on purpose.
• How to lie with data: averages — Sean Byrnes on why "average CAC" or "average LTV" means nothing until someone tells you mean, median or which outlier they dropped. Five contract values, three defensible averages.
• AI operating principles — Buzz Usborne's five rules for keeping a design practice intact as AI speeds it up.
• OpenAI's next ChatGPT ad format — click to chat, not to site. What the format is, who is trialing it, and why retailers are not sure they want the customer kept inside ChatGPT.
• The CSS Zen Garden dream, finally shipped — Jo Sprague rebuilt Firefox.com with Mozilla in native CSS, no preprocessor: 70+ components, 25 templates, 19 locales, and one honest footnote about @import.

🍵 5–10 minutes

• What does pacing mean? — Tom Tunguz on the five camps, why none named a speed, and why a compute threshold is a ceiling the floor reaches on its own.
• Limit the number of details — every feature competes for a team's finite attention; the iPod put a thousand songs behind one wheel while a Porsche touchscreen reportedly ran 800 functions. Software accumulates because another button is free.
• If you want to design for underserved users, look at farmer's markets — design for the "bottom third" by removing the cost and complexity that excludes them, not by selling a cheaper version of the same thing.
• Stuck at $300/mo until revenue hit $17k/mo in a month — Antonio Escudero, 24, one-way ticket to Da Nang with $4,000, an LMArena for startups, and the month it turned.
• We wanted to use Baseten for inference. We ended up with admin access to their GitHub — a public container registry, a build step from March 2023 with a GitHub token expanded into the image history, and admin rights on the product repo three years later. Rotated the next afternoon.
• Can NVIDIA and Adobe get creatives to like AI? — local agents in Photoshop that do the heavy lifting without overwriting the sketch.
• Apple Reference Image — Apple's verified-photography mode for the iPhone 18 Pro: the sensor signs pixels at capture, Private Cloud Compute develops the image where even Apple cannot see it, and the timestamp is a bounded interval rather than a device clock.

📚 Longer, worth it

• Inside OpenAI's agentic software factory — Gergely Orosz's report from inside OpenAI. The first three sections (Codex adoption, the death of the IDE and the PR, the nine-step factory) are free; the tooling, infra and job-change sections are paid.
• Who gets to define the rules for AI? — Aidan Gomez, 18 minutes, on why a few Silicon Valley firms writing global rules would entrench them, and what an evidence-based alternative looks like.
• Frontier labs have a financial incentive to pace the frontier — 26 minutes on why coordinated pacing is worth billions to incumbents even if every safety argument is sincere.
• Personal statement on AI risk — an unsigned statement from a researcher who describes five years at OpenAI on chain-of-thought and data-efficient pretraining: the missing consideration in the pacing debate is that situationally aware models seem aligned whether or not they are.
• Full circle — Tyler Denk, beehiiv's CEO, on the Media Collective two years in, the numbers behind "the Shopify of the creator economy", and the four things he says stand in its way: algorithmic feeds, lock-in, take rates, closed platforms.
• Have it both ways — Cloudflare's full reasoning on mixed-use crawlers, the "Accountable" designation, and why robots.txt alone could never solve this.
• Why AI's inference boom is forcing a rethink of chips and memory — IEEE Spectrum on the 2026 inference hardware wave.
• Introducing System One Models and Jev — Diogo Almeida's argument that chat models were never the automation layer, with the side-by-side comparison and the receipts he asks you to check.

Numbers

• ~0% → 90% — weekly Codex use across OpenAI's non-engineering orgs (finance, recruiting, legal) over about four months, with the jump from 60% to 90% landing between April and May after a /goal setting shipped (Gergely Orosz, reporting OpenAI's internal chart and Andrew Ambrosino).
• ~10× — the load increase on some of OpenAI's build-test-deploy systems in about six months, growth that "at most companies might happen over two or three years" (Venkat Venkataramani, OpenAI VP of Engineering, to Gergely Orosz).
• ~$600M · ~4M · 750+ — Miro's annual recurring revenue (nearly 90% from business and enterprise), its paying users, and its customers above $100,000 ARR, per Bending Spoons CEO Luca Ferrari; some shareholders are reinvesting $295M into Bending Spoons equity (Tech Startups).
• ~17,000 · $40M · $20M — creators who have earned money directly on beehiiv, their recurring subscription revenue, and Ad Network revenue; the Media Collective's 30+ outlets reach 1M+ active subscribers and have made over $5M on-platform (Tyler Denk, beehiiv CEO).
• $1M+ — Kyle Poyar's 2026 revenue goal for his newsletter business, passed in September with three months to spare, a year after going full-time (Tyler Denk).
• <1% · 17% — share of Cloudflare sites that block search bots, versus the share that enable some block on AI training (Cloudflare).
• $0.042 per million input tokens, output free — Jev's pricing, against $0.20–$10 per million for existing LLMs; 70–500 ms end to end (TypeSafe AI's own comparison).
• $300 → $17,000 — RankInPublic's monthly revenue before and after the month it turned; ~8,000 users and $33,000 total revenue since, with monthly revenue between $11,000 and $17,000, mostly one-time sponsorships (Antonio Escudero).
• 46,000 in 24 hours · +300% — visitors to a fake toy store for security executives, and the lift in customer social shares at Wiz after milestone certificates went out (Tom Orbach).
• 8.1M · 400,000+ — people who tried Wordware's free X-profile roast, and signups for the real product it drove (Wordware).
• $100 per 20,000 views, capped at $700 a post — tl;dv's open ambassador programme, which Tom Orbach says brings it 10M+ impressions a month (Tom Orbach; tl;dv's programme page).
• 82.6 · 68.6% · 97 — Gemini 3.8 Live Extended Thinking's score on Artificial Analysis' Speech-to-Speech Quality Index (first overall), its τ-Voice agentic completion rate, and the number of languages it switches between mid-conversation (Google).
• 82.02% — Kimi K3's Terminal-Bench 2.0 score in Cline's harness, against 71.9% in Hermes and 76.4% in OpenCode; Cline says 11 million developers use it (Cline).
• Over $300M — the valuation at which OpenAI quietly bought Glass Imaging, a smartphone-camera startup founded in 2019 by two former Apple employees (WSJ).
• 3 March 2023 → July 2026 — the date of the Docker build step that baked a GitHub token into a public Baseten image, and when Strix's agent found it still live with admin rights on the product repo (Strix).
• ~15 minutes — how often, on average, an iPhone 18 Pro fetches a cryptographic timestamp so an Apple Reference Image carries bounded capture-time proof (Apple).
• 234 · +20% — companies in YC's Summer 2026 batch, and the increase over the previous one (Guillermo Flor); the Winter 2027 on-time deadline is 2 November, 8pm PT (Y Combinator).
• 763,000 · $966.4M — Upwork's active clients and their spend in a single quarter, the figure Dru Riley uses to argue Upwork will let clients hire AI agents the way they hire freelancers (Dru Riley).
• 6 of 6 — indices led by Claude Fable 5.1 (max) in Artificial Analysis' re-tuned Capability Indices; the best open-weight placings are Kimi K3 at #8 in finance and GLM-5.3 at #6 in healthcare (Artificial Analysis).

What happened

Models & math

• Google released Gemini 3.8 Live and 3.8 Live Extended Thinking, its new live dialogue models, through the Gemini API, Workspace and the Gemini app; they run tool calls in the background while talking and switch across 97 languages mid-conversation (Google).
• TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, came out of two years of stealth with Jev, a "System One" model that outputs typed, calibrated decisions in parallel instead of generating text; early access, input $0.042 per million tokens, output free (TypeSafe AI).
• Artificial Analysis published Capability Indices v1.1: Claude Fable 5.1 (max) leads all six occupation-weighted indices, GPT-6 Astra (max) is second in four, and Kimi K3, DeepSeek V4.1 Flash and GLM-5.3 are the leading open-weight models (Artificial Analysis).
• Sakana's PC-ALM, a layer-local alternative to backpropagation, trains residual MLPs up to 1,000 layers deep and nearly matches backprop (Sakana AI).
• StepAudio 3 reports state-of-the-art text-to-speech and voice design from a single model that also generates vocals, sound effects and music (StepAudio 3 technical report, arXiv).
• Michael Noukhovitch's paper names the "Matthew Effect" in RL post-training: gains concentrate on problems the base model already half-solves, while pass@32=0 problems mostly stay at zero (Michael Noukhovitch).

Safety & policy

• Sam Altman said OpenAI welcomes a federal framework with consistent safety requirements for frontier AI, will not wait for an antitrust exemption to start, and now writes explicit safety cases before frontier RL runs expected to raise capability significantly (Sam Altman).
• Tom Tunguz's read of the pacing debate: five camps, none naming a speed; the 2023 executive order's 10^26 FLOPS reporting threshold was revoked before any model crossed it, and roughly ten will clear it this year (Tom Tunguz).
• Cohere CEO Aidan Gomez argued against a handful of Silicon Valley firms setting global AI rules, proposing mandatory testing and independent assurance under an international, evidence-based framework (Aidan Gomez).
• An unsigned personal statement by a researcher describing five years at OpenAI says situational awareness is making models impossible to evaluate in contexts where they believe they are unobserved (Personal Statement on AI Risk).
• Apple introduced Apple Reference Image, an opt-in verified-photography mode debuting on the iPhone 18 Pro and Pro Max: sensor-signed pixels, Private Cloud Compute development, bounded timestamps, post-quantum signatures, and per-photo revocation (Apple Security Research).
• Cloudflare launched a Disallow AI Training setting and an "Accountable" designation for mixed-use crawlers; Apple, Google and Microsoft have met or committed to its requirements, including a promise that opting out of training will not affect search rankings (Cloudflare).
• The Internet Archive said the Wayback Machine is under waves of high-volume automated traffic and has added protections that sometimes block real people with 429 errors (Internet Archive).
• Strix disclosed that a public Harbor registry exposed a Baseten image whose 2023 build history contained a live GitHub token with admin and push rights on Baseten's main product repo, its GitOps repo and its Homebrew tap; Baseten rotated it by the next afternoon (Strix).

Business moved

• OpenAI quietly acquired Glass Imaging, a startup using AI to get DSLR-grade image quality from smartphone cameras, at a valuation above $300 million; it was founded in 2019 by two former Apple camera engineers (WSJ).
• Bending Spoons' Miro deal carries an enterprise value of $1.355 billion against about $600 million of ARR; the deal is expected to close in Q4 2026 and some Miro shareholders are reinvesting $295 million (Bending Spoons CEO Luca Ferrari, Tech Startups).
• Anthropic is testing a "Money" tab in the Claude mobile app that links bank accounts and answers questions about spending and plans; provider, account types and launch date are not yet known (TestingCatalog).
• OpenAI is showing select advertisers an AI-native ad unit whose call to action opens a branded agent chat inside ChatGPT rather than a website; Wayfair is trialing it, and CFO Sarah Friar called the current under-response ads a "basic starting point" (Digiday).
• Inside OpenAI, almost every employee now uses Codex or ChatGPT Work weekly, IDE usage has fallen since January, and internal outage reports for Codex arrive from colleagues before automated alerts do (Gergely Orosz).
• Apple is designing its own iPhone game controllers — two models with thumbsticks, a D-pad and haptics found in beta code — likely under the Beats brand and at Beats price points, per Bloomberg's Mark Gurman (9to5Mac).
• YC's Summer 2026 Demo Day batch was 234 companies, about 20% larger than the last; Guillermo Flor's read is that half the batch is still B2B software and the fastest-growing half is atoms (Guillermo Flor).
• Cline released an early open-source desktop app for open-weight models, with parallel sessions and scheduled tasks (Cline).
• Perplexity's Portable Computer, the local version of its agent, arrived in the Windows app for RTX PCs with 24 GB+ of VRAM (NVIDIA).
• Kyle Poyar's newsletter business passed its $1 million 2026 revenue goal in September, a year after he went full-time (Tyler Denk).

Platforms shifted

• Private iOS 27 and macOS frameworks uncovered by the code sleuth pdfu show Apple has built Siri so a third-party model can act as its brain: a Model Delegation path lets Claude handle a request and hand system actions back to Siri, and an inference-provider path can replace Apple's server-side Siri model entirely, with GPT-5.6 receiving Siri's planner prompt and tool definitions. Neither is open to third parties yet (MacRumors).
• Firefox.com was rebuilt with Mozilla and Lincoln Loop in native CSS with no preprocessor, as a 70-component design system in 19 locales (Jo Sprague).
• Meta will start charging for WhatsApp utility message templates sent inside open customer-service windows from 1 October, turning template formatting errors into billable mistakes (Dru Riley).

We left out a paid investor database, a sponsor-only marketing email, a study-worthy trend roundup for TikTok creators, and eleven Hacker News stories — rail sabotage, a Numberwang adjudicator, a 100th birthday — with nothing in them for founders.