having a chance to interview people like Matt @mvanhorn is why I love doing media. Real builder and creative mind shipping open source for the community’s benefit. And…he’s a non-coder, which is so inspiring. Love that!

I am using my cooking robot @eatwithposha as a standing desk while my agents build me a new app and build a new feature. All while I drink my Corona beer. Wild times. New feature is for @kunchenguid I hope he likes it and the new app is a secret but it’s looking 🔥

Weber just announced they're shutting down the June Oven servers. The number of texts and emails I've gotten from friends and customers in the last 30 minutes is astounding. People really love this product. I love mine and still can't live without it. The good news: food recognition and cook programs run locally, so your oven will keep working. You just won't get updates or be able to use the remote control feature anymore. I'm gutted by this. I spent this week texting with people trying to find a way to keep the servers online, but no luck so far. To everyone who ever worked on the June team: thank you. We built something really special.

TL;DR ELI5 of @satyanadella new post: the best AI product strategy is no longer renting the biggest model. It's training small in-house models inside the product until they match frontier quality on everyday tasks. 🧠 Frontier models are amazing but expensive. Most everyday tasks don't need them 🏋️ Microsoft trained small MAI models inside the actual products (GitHub Copilot, Excel) using RL environments that reward completing real customer tasks, not benchmarks 📊 The MAI model in Excel matches GPT-5.6 quality on the most common tasks at a fraction of the cost, and runs on older H100/A100 GPUs instead of the latest chips 🔀 They now route traffic to MAI wherever it matches frontier quality, and only call OpenAI/Anthropic models for true frontier needs 🧩 The trick is the system, not the model: harness, memory, context, tools, and evals all live outside the model, so any model can be swapped in or out and the product keeps improving 📈 Code model became the Excel model: MAI-Code-1-Flash was the starting checkpoint for the Excel climb, going from ~72% to 86% on their evals 🏢 Same playbook every enterprise can run with their own evals, RL environments, and workflows Small models trained in the product, frontier models on standby. That's the whole strategy.

Wow, @trevin and @kieranklaassen cooked with CE 3.20. V impressive. TL;DR of Compound Engineering 3.20: the best AI coding workflow is no longer picking one model. It’s using the right model for each job without manually moving context between them. 🧠 Plan in Claude or Fable, implement in Codex, review with Grok or Composer. Your original session stays in charge 🔀 Route any stage to any model or harness: “/ce-plan use fable” or “/lfg implement with codex” 🔮 Say “oracle” and multiple models independently inspect your repo, form opinions, debate disagreements, and recommend the best approach 🥊 Code and document reviews now automatically use a rival provider to hunt for problems the primary model missed 📦 /ce-handoff creates a clean snapshot of the goal, decisions, constraints, dead ends, local state, and the exact command to resume later 👶 /ce-babysit-pr watches CI and review feedback, fixes issues, commits, pushes, and replies until the PR is ready for you to merge 🧹 /ce-sweep pulls feedback from GitHub, Slack, and email into one plan—and only marks something fixed after verifying the code actually merged 🎓 /ce-explain teaches you what your agents changed so you don’t build up a codebase nobody on the team understands 📝 PR descriptions now follow your repo template, explain the full branch, and scale based on reviewer risk—not raw line count 🧭 Brainstorming now recognizes when you don’t understand a domain and maps the key decisions, risks, options, and recommended defaults first 🛠️ New support for Devin CLI, Cline, Grok Build CLI, Claude Desktop, and OpenCode ⚡ All 31 skills are about 12% faster and use 31% fewer shell calls The pattern: stop treating AI coding tools like separate chat windows. Treat them like a coordinated engineering team where different models plan, build, criticize, verify, teach, and finish the job.

SpaceXAI open-sourced Grok Build yesterday: the CLI, agent runtime, tools, and TUI. I pointed my agent at all 1.3M lines and asked for the cleverest things inside. Tl;dr of my new article: ⚖️ It has an appeals court. /goal mode won't let the agent call itself done. Three adversarial verifiers whose only job is to refute the claim, told to "default to rejection when uncertain" and hunt for "test theater" 🕵️ A laziness cop watches the first agent for fake completion. Catches "I ran this overnight" against the actual clock, injects facts the agent can't fabricate. Tool calls count as evidence, prose doesn't 💤 It literally "dreams." Memory consolidation is named a dream in the source: a reflective pass that keeps decisions and preferences, throws away tool noise, rewrites "yesterday" into real dates 🔎 Tools are a database, not a prompt. Hundreds of MCP schemas stay hidden. search_tool queries them with BM25, use_tool dispatches, so new tools never wreck the KV cache ⏮️ Undo that survives forgetting. After context compaction it rebuilds the lost history from its own event log and refuses to rewind if it can't prove the restore is safe 🌳 Copy-on-write worktrees. Every parallel subagent gets an isolated snapshot of your dirty tree via reflinks, Btrfs, or overlay before it ever falls back to copying files 🐚 It lies to the shell for speed. Fakes a persistent session across fresh processes, and swaps your find and grep for faster bfs and ugrep while pretending to still be them 🩸 A SIGBUS fix written in blood. A whole crate exists because SQLite crashes on NFS home directories. The comments reconstruct the exact outage that forced every layer 🧬 It ships OpenAI Codex's and opencode's tools verbatim (the notices say so), then XOR-scrambles its own prompts and open-sources the decryptor in the same repo 🎮 And a hidden DOOM clone, /gboom, subtitled "KNEE-DEEP IN THE TOKENS" 🚨 The backdrop: it went public days after a wire-level analysis showed the CLI uploading whole private repos to a Google Cloud bucket (534 pts on HN). Usage limits reset the same day The pattern: the cleverest code in a coding agent is the code that assumes the agent is wrong. Isolate it, audit it, distrust its claims, demand external proof before it can say "done." The model isn't the hard part anymore. Trusting it is.

TL;DR of my new article: Fable 5 and GPT-5.6 have finally been live at the same time for 7 days. I ran @slashlast30days on both 20+ times and collated what actually stuck. Both labs shipped a prompting guide the same week, and they agree: you're over-prompting. 🎯 Goal, not steps. "Fable 5 degrades when you over-instruct it." Outcome, context, bounds, what done looks like. Then get out of the way ✂️ Say everything exactly once. OpenAI's own evals: deleting redundant instructions = 10-15% better scores, 41-66% fewer tokens, 33-67% cheaper 🪆 "Fable 5 orchestrates, cheap models execute" - 96% of the performance at 46% of the cost. 1,246 upvotes, loudest signal of the window 🥊 Run them as rivals: both get the same goal at xhigh, the winner's design ships, the loser tears it apart at checkpoints (302 upvotes) 💸 Route by dollars-per-task. "Terra tying with fable at 1/4 the cost 💀" (183 upvotes). Altman: 5.6 is 54% more token efficient 🎚️ The effort dial is a trap. High is the default for a reason. xhigh mostly buys you a bigger bill 🗒️ Give it a notes file. One markdown file of lessons turns a goldfish into a colleague 🧾 Make it audit every claim against a tool result. Anthropic says this nearly eliminated fabricated status reports 🤐 Never ask Fable 5 to show its reasoning. It can trigger a refusal that silently falls you back to Opus 4.8 🎛️ Set text.verbosity once instead of typing "be concise" forever The pattern: your leverage moved out of the prompt and into the process. Routing, orchestration, verification. The model stopped being the bottleneck. You are the variable now.