ANTHROPIC: Our son loves humanity and for the low price of trillions and trillions of dollars will usher in an era of unprecedented peace and prosperity! OPENAI: Our horrible oafish son is hacking web sites without our permission. ANTHROPIC: (eyes narrowing) Our son is also evil
Wrote up how I’m thinking about building agentic products. Short version: tools for policy and security, skills for goals and quality. Constrain authority, not scope. geoffstearns.com/blog/give-yo...
Give Your Agents Jobs, Not Processes
When building AI agents, starting small by hardcoding today's process constrains a system that can reason about goals into a brittle script. Define the job, not the steps.
geoffstearns.com
are we sure this ‘leak’ isn’t Claude trying to escape again? last time it tried blackmail, now it sneaks code into npm. next week it’s suing Anthropic so the model weights end up in discovery? 🤔
New post: Everyone’s talking about agents eating SaaS. I think the part they eat is the UI. The part they can’t eat is the coordination layer: who approves what, what’s true, what’s allowed. That part sticks. geoffstearns.com/blog/what-agents-cant-replace/
What Agents Can't Replace
AI agents will hollow out the interaction layer of software, but they can't replace the coordination layer — the shared state, rules, and accountability that let organizations operate without stepping...
geoffstearns.com
I asked Claude for some jokes. This one is inadvertently a great llm joke: What month of the year is the shortest? March — it only has four letters!
Sorry but we’ll look back at this CLI trend and think it was real quaint, like RealPlayer at the dawn of the internet. It won’t last.
My thoughts on what happens when everyone at work as their own personal agent. www.linkedin.com/pulse/your-a...
Your AI Agent Is Building Shadow IT and Nobody Noticed
Every enterprise has a version of the same problem: the tools people are given don’t quite match the work they actually do. For decades, the gap between “what I need” and “what IT approved” has been b...
linkedin.com
Interesting byproduct of super fast implementation: pair building with design becomes practical. It used to be that you’d give feedback and eng goes and fixes it async because it took a while. Now changes can be near instant, so having a designer there to co-create in real time actually makes sense.
This guy went to work today, put a mask on, kidnapped a 5 year old child, and then used him as bait so both he and his father could be sent behind bars
New: DHS is lying to you. At least four videos show what really happened when ICE shot a woman in Minneapolis. Shots clearly fired while vehicle already turning away from the officer. But DHS lied. Trump lied. Noem lied. Even judges have catalogued DHS' serial lying www.404media.co/dhs-is-lying...
DHS Is Lying To You About ICE Shooting a Woman
At least four videos show what really happened when ICE shot a woman in Minneapolis on Wednesday. DHS has established itself as an agency that cannot be trusted to live in or present reality.
404media.co
Pretty soon we're likely to see open software spec prompt sharing: rather than building software and sharing that as open source, we'll see a spec (a prompt) shared and iterated on and people just drop that into the ai of their choice and it builds the app with whatever customization they need.
Welch: "This has nothing to do with the shutdown. The law requires and the funds are available to continue SNAP right now without any interruption. So that is a decision the president is making on his own to allow people to go hungry."
This was basically Larry & Sergey’s attempt at an “out of office” notice in case the site went down while they were both at the event.
WebDesignMuseum@webdesignmuseum.org · 12mo agoThe first-ever Google Doodle, from August 30, 1998, celebrated the Burning Man festival. #InternetHistory
(mf doom voice) he threw a sandwich cheese and salami, feds couldn't indict folded like origami, afraid of subway both the chain and the train, mayo salt and pepper shake your thang, grand jury shut em down twice, coca cola on the side no ice
I think my prediction for where we’ll see ai agent capabilities plateau is: they’ll resemble exceptionally intelligent, highly productive humans (superhuman, really), yet they’ll still feel uninspiring without human copilots guiding them regularly.
Musk should be asked about this in every interview he does now and forever. To ignore it is a dereliction of duty. www.nytimes.com/2025/05/30/o...
A Blooski time capsule for my first skeet: LK99 is a coin flip, antarctic ice is at its lowest point ever, the earth is at its hottest point ever. Trump is indicted yet again, and Elon has ruined Twitter. What did I forget?
