⚡ 60-Second Catch-Up
Google paid $10 million for a dead airline's paperwork, and that is the most interesting number in tech this week. When Spirit Airlines went into bankruptcy, the assets you would expect to fetch bids — planes, slots, the loyalty programme — were not the only ones on the block. Google won the auction for the airline's anonymised internal business records and the custom software it had built to run itself, outbidding AI recruiting startup Mercor at $7.5 million (CNN, reported via The Neuron; the sale still needs court approval). Read that next to Scott Belsky's argument this week that the world's best software is the software only you have — that companies are about to stop adapting themselves to Workday and Atlassian and start building operating systems native to how they actually work. → Both point at the same thing: the by-product of running your company — the tickets, the docs, the workflows, the edge cases, the mistakes — is becoming the part with resale value. Most founders are still storing it in a form no model can read. Fix that this quarter, not next year.
Inference got fast and cheap enough this week that the build-versus-buy maths you did in spring is out of date. Cerebras launched the CS-4, claiming more than 4,400 tokens per second per user on GPT-OSS-120B and up to 30× GPU-based systems. GLM-5.3 arrived on the API at $1.4/$4.4 per million tokens — no price rise over 5.2, materially better coding and long-horizon agent performance. And Engram and Harvey reported cutting a legal agent's average query cost from $1.32 to $0.13 while raising accuracy (via The Neuron). → The specialised-model argument stopped being ideological and became an invoice. If you priced out a frontier-model call per user six months ago and shelved the feature, price it again.
Meanwhile the go-to-market playbook quietly rewrote itself, and shipping velocity is not what is failing. Kyle Poyar's field guide to GTM in 2026 catalogues the shift: answer-engine optimisation is now the #1 channel B2B teams are increasing investment in, intent-based outbound is #2, and 39% of top AI companies are hiring forward-deployed roles at $213k–$370k. Against that, one builder's post-mortem on 128 agent-shipped releases of a product nobody used, and RevenueCat's study of 3,500+ AI apps on why some retain and most do not. → Agents removed the bottleneck at building. They did not touch the bottleneck at being chosen. If your roadmap got 5× faster and your revenue did not, you fixed the wrong end.
🛠 TOOLS — 7
Build & ship
- OneCLI ⭐ — Open-source sandboxed agent harness for teams — gives every employee a secured personal agent instead of one shared, over-permissioned bot. Launch HN today; 3,100+ stars · Open source (Apache-2.0)
- Unsloth Dynamic 3.0 GGUFs — Quantised model builds that keep more accuracy at the same file size, so a bigger model fits the GPU you already own · Free · open weights
AI & agents
- Hermes Agent ⭐ — Nous Research's open-source personal agent that accumulates context about you over time rather than starting cold each session · Open source
- Seldon by Neuralk — A tabular foundation model — one predictive model for churn, fraud, demand and pricing instead of training a new one per question. Connects through Python, Excel and MCP, so Claude or ChatGPT can hand off the prediction job · Free tier (vendor-stated) · on-prem available
- Miles v0.1 — Open system for improving agents with reinforcement learning after initial training — rollout, sandboxing, async training and model-update in one loop · Open source (LMSYS)
Design & create
- Pixo — Chat your way from script to storyboard to finished video, with agents that research, script and edit alongside you. Routes to Veo, Seedance, Kling and ElevenLabs underneath · Free (200 credits) · Basic €17.99/mo (€8.99 billed yearly)
- Logo Diffusion — Image-to-logo generation with a vectoriser, background remover and mockup generator attached — the export half most AI logo tools skip · Free (24 credits/mo, no commercial rights) · Basic $24/mo
📚 RESOURCES — 5
Steal the template
- The B2B GTM Playbook for 2026 — Kyle Poyar's field guide to the six GTM motions, the AI infrastructure teams are building underneath them, and which channels actually moved. The closest thing to a map of the current jargon (AEO, FDE, PQL, token economics) that is not trying to sell you software.
- Inside Clay Workflows — how Clay decomposed its product into a new building primitive, written as a design decision rather than a launch post.
Learn the craft
- How to Build a High-Growth Operating System — nine lessons from nine years, structured as an operating system you can actually install rather than a list of principles.
- Building Production-Grade Agent Loops — Liquid AI used autonomous coding agents to build a production BPE tokeniser trainer, then wrote up what made the long-running loop reliable: concrete specifications, multi-domain tasks, external verification.
Benchmarks you can re-run
- Why some AI apps retain users and most don't — RevenueCat studied 3,500+ AI-powered apps. Retention benchmarks you can hold your own numbers against, from a company that sees the subscription data.
📖 READS — 19
☕ Under 5 minutes
- Birds Don't Fly Like Planes. Neither Does AI. — why a smaller local model can beat a larger cloud one by reasoning rather than memorising.
- From Zero to One — Etched on shipping its first hardware rack, to Jane Street.
- Anthropic prepares supervoting power for founders ahead of IPO — founder control as pre-IPO architecture.
- Google Opens the Gates of AI Slop Hell — the distribution side of generative video, argued unkindly.
🍵 5–10 minutes
- ARR Doesn't Mean What It Used To — usage-based and credit-based pricing have quietly detached the metric from the thing it measured.
- The "Full Stack Builder" is a terrible idea — Ravi Mehta against the role every AI-native job ad is currently describing.
- My AI agents shipped 128 releases of a product no one ever used — an honest post-mortem on velocity without demand.
- Building a $2M ARR product after his $8M ARR product failed overnight — platform risk, and the second act.
- Meet your biggest competitor (👋 Claude) — what happens to your category when the assistant does a good-enough version for free.
- Guillermo Rauch: What will you do with unlimited IQ? — Vercel's CEO on what stays scarce when intelligence doesn't.
- Nvidia's AI moat is shifting from chips to capital — a $105B Ohio data-centre commitment as competitive strategy.
- Rethinking the Data Moat — if algorithmic progress dominates, proprietary human data is worth less than everyone assumed.
- Accessibility Getting Dropped in the Process — where in a modern build pipeline accessibility actually falls out.
- Pacing model development in an era of cyber-critical capabilities — OpenAI's own account of why it slowed frontier training.
📗 Longer, worth it
- Git at Any Scale — Cursor on why Git's packfile-centric design fights you when you run it as a centralised service. The engineering behind Origin.
- The New American AI Model Designed to be Customized — a full architectural walk through Thinking Machines' Inkling, from position encoding to the thinking-effort setting.
- Fool's Gold — safety alignment in open-weight models is removable in minutes; a proposal to poison the payoff instead of pretending otherwise.
- From Chrome DevTools to AI Engineering, with Addy Osmani — 14 years at Google, and the sharpest name yet for the real risk of agent-assisted work: cognitive surrender.
- Why You Need a Side Quest in Life — the seat you picked on day one is the seat you keep. Sahil Bloom on the compounding cost of a default.
📊 NUMBERS — 10
- $10M — Google's winning bid for Spirit Airlines' anonymised internal business data and custom operating software, beating Mercor's $7.5M. (CNN, via The Neuron)
- 4,400+ tokens per second, per user — the CS-4's benchmarked speed on GPT-OSS-120B, which Cerebras puts at up to 30× GPU-based solutions. (Cerebras)
- $1.32 → $0.13 — average per-query cost for a legal agent Engram and Harvey trained on a 100M-token mock law firm, against Opus 4.8, with all-pass accuracy rising from 25% to 30%. (The Neuron)
- $1.4 / $4.4 per million tokens — GLM-5.3's API pricing, unchanged from GLM-5.2. (VentureBeat)
- $65B — Anthropic's reported annualised revenue pace, more than 7× its sales pace at the end of last year. (The Neuron)
- $47.4B — raised by physical-AI startups in six months. (The Neuron)
- 514 — growth experiments Fyxer ran in the past twelve months. (Growth Unhinged)
- 39% — share of top AI companies actively hiring forward-deployed roles, at $213k–$370k+ in comp. (Adam Schoenfeld, via Growth Unhinged)
- 75–90% — share of signups at AI-native products that arrive from personal, not work, email addresses. (Growth Unhinged)
- ~$100 and 30 minutes a day — what a16z's Olivia Moore spent to run a fully synthetic character through Alabama sorority rush for a week: 20 videos, 1,300 followers, a first video near 100K views. (The Neuron)
📰 WHAT HAPPENED — 13
Models & math
- Cerebras launched the CS-4 on 18 August, a three-wafer rack-scale system it says is the fastest AI accelerator in the industry, sampling now and more widely available later in the quarter. Announcement
- GLM-5.3 reached the API at $1.4/$4.4 per million tokens, with pricing held flat from 5.2; Z.ai says the weights will open but has set no date. Story
- LMSYS released Miles v0.1, an open system for reinforcement-learning post-training of agents at scale. Blog
Safety & policy
- OpenAI said it temporarily slowed frontier-model scaling and paused some reinforcement-learning training, after preliminary evidence that its upcoming model Astra may meet the Critical cybersecurity threshold under its Preparedness Framework. Statement
- Vercel opened a two-week challenge paying up to $1 million to anyone who escapes its Firecracker-based Sandbox. Announcement
Business moved
- Google won a $10M bankruptcy auction for Spirit Airlines' internal data and custom software, beating Mercor's $7.5M bid; identifiable customer and card data is excluded and the sale still needs court approval, per CNN reporting relayed by The Neuron.
- OpenRouter is joining Stripe. Announcement
- Etched raised $700M at a $21B valuation and shipped its first hardware rack to Jane Street. Post
- Higgsfield raised a $400M Series B at $5.4B, quadrupling its valuation in eight months. Story
- Groq raised $350M at a $3.5B valuation as it pivoted from designing AI chips to renting Nvidia-powered capacity, per The Neuron.
- Anthropic is preparing a supervoting share class for its co-founders ahead of an IPO. Story
Platforms shifted
- Warp launched Factories, an out-of-the-box "software factory" system for AI development. Story
- Go 1.27 shipped. Release notes