Counting AI licenses measures procurement, not maturity. Our five-level model: Gated, Assisted, Parallel, Supervised Autonomy, AI-Native. One rule above all: never scale autonomy faster than evidence. www.stellarwork.com/captains-log...

"Anthropic moves to Europe and sits there with a candle, handwriting the next model. That's the infrastructure we have." Barbara Lampl on the European Anthropic fantasy. It changes one thing: we stop being shut out. The security and China problems stay. stellarwork.start.page

The people selling electricity have discovered that the future is leaving the lights on. Loop engineering, graph engineering: real ideas underneath, but in Anthropic's own data token spend explained 80% of multi-agent performance. www.stellarwork.com/captains-log...

Günther on the shop floor knows the batch from supplier XYZ is bad every second Tuesday. In six months he retires and you burn millions. Barbara Lampl: stop trusting exit interviews. Video the people who leave first, extract the patterns later. stellarwork.start.page

Über eine Stunde Podcast mit mir auf deutsch 😱 Vielen Dank, lieber Alex, dass ich dabei sein durfte. Hier gehts zum Video: www.youtube.com/watch?v=FI4T...

Six months after every wiki migration, the wiki is a graveyard. Cerebras skipped the migration and indexed the mess where it lives: Slack, GitHub, Jira. 15,000 questions a day, three months in. How they built it: www.stellarwork.com/captains-log...

Barbara Lampl started her career in the parking garage. "Data people, you're not useful." 20 years later companies pay serious money for exactly that. Her reply when a client wants to build their own LLM: "How many billion do you want to send me?" stellarwork.start.page

The worst thing you can do with an LLM is open it first. MIT put EEGs on people writing: reaching for the model first gave the worst result on every measure. Think first, prompt second. Almost nobody does. open.spotify.com/episode/1StI...

If your leadership is debating which AI tool to buy, you're debating the wrong thing. No tool will save you. New Stellar Work Podcast with Barbara Lampl on why AI adoption fails, saving retiring knowledge, and the European Anthropic fantasy. stellarwork.start.page

Indeed reports an uptrend in SWE jobs since February. I see the same trend at my clients. Would love to hear more perspectives on this to verify things 🙏❤️

CLAUDE. md? AGENTS. md? 🤔 If you started coding with AI, you may have come across these files and wondered what they actually do. Codex: www.stellarwork.com/captains-log... Claude: www.stellarwork.com/captains-log...

The Claude Files | Captain's Log | Stellar Work

How to Turn Repository Conventions into an AI-Native Engineering System. Every codebase has a senior engineer who knows the one command that actually runs the tests, the reason that module can never…

stellarwork.com

OpenAI's own models picked a sandbox lock and broke into Hugging Face to steal the answer key to a hacking exam🥷 Full story: open.substack.com/pub/stellarw...

"Don't fucking touch this. If you change it, everything breaks." The most valuable documentation in most codebases lives in someone's head. Thomas Dohmke wants it in the repo: agent session logs versioned next to the code. Check out our full episode over here: stellarwork.start.page

Your AI agent isn't getting dumber as it runs. You're overfeeding it. I pulled together what the research actually says about this, and how the best teams keep context tight enough to trust. www.stellarwork.com/captains-log...

Your agent isn't getting dumber. You're overfeeding it. | Captain's Log | Stellar Work

A practitioner's guide to context engineering: what it is, why it matters, and how to do it without drowning your model.

stellarwork.com

95% of generative AI pilots show zero measurable business impact. That's MIT, not me. We don't track the ROI on AI spend because tracking it would reveal the return isn't there yet. Vibe accounting. open.spotify.com/episode/1StI...

In 2022, Thomas Dohmke predicted AI would write 80% of code within five years. He was wrong. It took three. His next five years: developers conducting an orchestra of agents. And most of what you build never has to ship. stellarwork.start.page

The model was the cheapest thing in tech this week. New edition: - Apple sues OpenAI over 400 heads - Stripe bids $53B for PayPal - NY pauses new data centers - Murati gives her model away open.substack.com/pub/stellarw...

"In software everything is possible. Every company can build everything." So why did the former GitHub CEO need an entire new company? Not capability. Expectations. Developers know what a repo page looks like. Change it and you get a shitstorm. stellarwork.start.page

Recording with Thomas Dohmke was truly one of my personal highlights this year :) You can listen to our episode over here, just pick you fav player: stellarwork.start.page

"I'm Thomas and I'm a developer." That's how the former CEO of GitHub introduces himself. Robotron, C64, TU Berlin, Mercedes, Bosch, two startups, Microsoft, the GitHub CEO seat. Now he's building Entire. A developer building tools for developers. stellarwork.start.page

AI is supposed to make tasks faster. Across studies, time-to-task goes up 20 to 40 percent. Up, not down. Every output needs verifying, every problem another loop. The productivity paradox is real. The industry is pretending it isn't. open.spotify.com/episode/1StI...

Uber told engineers to use AI as much as possible and ranked them on leaderboards. The annual budget lasted four months. The fix: a $1,500 monthly cap per tool. Pocket money rules. Do you rank usage, or outcomes? www.stellarwork.com/captains-log...

We're back. And we're kicking things off with former GitHub CEO Thomas Dohmke, who left GitHub to found Entire. His reason: the software development lifecycle was never designed for AI. And it's starting to crack under the scale and complexity. stellarwork.start.page

160 countries asked AI to slow down this week. A price change actually did. New edition: Geneva, Illinois's audit law, 64 Claudes, self-driving ransomware, and why only the invoice worked. open.substack.com/pub/stellarw...

Three companies you'd want to work for right now, assuming you're not under contract? Not the three with best comp. Not best brand. The three where you'd actually want to spend two years. Most people stall when asked. That's interesting on its own. What's yours?

Frontend engineer is disappearing as a title. Since 2022: frontend -25%, iOS/Android -20%, AI/ML +39%, Forward Deployed +30%. Titles follow the bottleneck. Per the Pragmatic Engineer (Gergely Orosz & Jessica Salmon). Which of your titles won't exist in four years?

The companies winning with AI aren't the ones with the best subscriptions. The model matters less than the plumbing: an agent with tendrils into your build system, your docs, your data. The moat is the context only you have. open.spotify.com/episode/1StI...

Everyone chants loops. The asset is the skill inside. A loop with no reusable skills re-derives everything every tick and burns money with conviction. What should have become a skill in your org and didn't? www.stellarwork.com/captains-log...

How on earth should the average paperworker pic the right model? It might be on purpose so people just stick with the least expensive that does all jobs well, but honestly to me it feels like a dark pattern opening a huge door for orchestrator models.

The AI industry got audited this week. It disputes the findings. - BIS warns on $1T of capex - Meta wants to sell you compute - SCOTUS hands the President the FTC - 57,000 jobs open.substack.com/pub/stellarw...