Maciej Myslinski (@macmys) header
Maciej Myslinski (@macmys) profile picture

I build the AI systems I talk about. Spec-driven development, agents in production, honest numbers. https://nitter.gq/t.co/ce9kkQfcsR co-founder · PhD · Leuven, BE · skiing

Leuven, Belgium
Joined April 2011
Photo by Maciej Myslinski (@macmys)Photo by Maciej Myslinski (@macmys)Photo by Maciej Myslinski (@macmys)Photo by Maciej Myslinski (@macmys)Photo by Maciej Myslinski (@macmys)Photo by Maciej Myslinski (@macmys)Photo by Maciej Myslinski (@macmys)Photo by Maciej Myslinski (@macmys)Photo by Maciej Myslinski (@macmys)Photo by Maciej Myslinski (@macmys)
Anthropic priced the biggest IPO ever last night. McKinsey says up to 56% of Belgian work hours could be automated with today's tech, and 26% as early as 2030. I think the share price will swing on the gap between those two numbers.
10
I told the agent: start stays disabled while a scan runs. Its next question: what if the scan never ends? My rule could lock the user out for good. Cancel button + 15-minute limit.
2
Before my agents plan anything, they now attack every handoff in a user journey: reload, back, repeat, failure, retry, limits, roles. An earlier agent-only pass: 25 of 90 scenarios had no defined behaviour.
10
A scan fails, the user tries another site, and the screen shows both messages. Nobody had written down which one wins, so the code showed both and no test looked.
3
Thanks for this one. On our new product all tests were green and I still found bugs clicking through staging myself. Some sat where the spec said nothing: the agent's plan made the call, and the tests only checked what the plan decided.
18
On hard changes, our AI reviewer never says "looks good" on its own: 13 of 14 rounds found something real. So we don't count rounds. Fix new problems inside the change, answer the rest with evidence, merge when a round finds nothing new.
9
That viral "stop coding between sets" sign from a San Francisco gym was fake. But I'm the guy it was about: between sets I dictate to three agent sessions at once. They do the slow work; I say yes, no, or what next. Sorry to whoever's waiting for the bench.
14
Wersja do przewinięcia ⬇️ #TK
8
Sources: Tribunal ruling K 34/15 (3 Dec 2015) on the 3 + 2 seats; ECtHR, Xero Flor v. Poland (7 May 2021): hudoc.echr.coe.int/eng?i=001…
21
Poland's Constitutional Tribunal crisis is back. I mapped it with Perplexity, from 1982. The majority always picked the judges. In 2015 the outgoing coalition picked two too early, and the new one refilled three taken seats. Since then the fight is over who is a judge at all.
1
17
Kryzys wokół Trybunału Konstytucyjnego trwa od lat. Poukładałem go sobie z Perplexity, od 1982. Sędziów zawsze wybierała większość w Sejmie. W 2015 PO–PSL wybrało dwóch za wcześnie, a PiS obsadził trzy miejsca już zajęte. Od tego czasu spór jest o to, kto w ogóle jest sędzią.
2
28
Ethan Mollick's recipe for AI at work: engaged leadership, a lab, and a crowd that tries things and shares what it learns. The sharing worries me. In the big company I work with, lessons sit in one SharePoint per department, next to dead files. Someone must own one clean place.
🤖 Made with AI
1
14
Per hour, the client can change anything and carries every risk. Fixed price, the developer carries the overrun and the client locks the scope. Revenue share, nothing up front, and both lose if nobody uses it.
1
1
16
Replying to @clairevo @cxodev
The first thing in mine that broke wasn't the agent, it was me: one job logged the same error for five weeks and I never opened the log. Now my morning brief starts with last night's errors. How does yours tell you it's broken?
25
Anthropic and GitHub both shipped "what did the AI cost, per piece of work" this week. Neither shows whose budget the saving lands in. That's still a spreadsheet someone has to own.
13
We bid a fixed price on a small app. The client chose an offer with zero build cost: the developer takes a share of revenue per active user. Cheaper code doesn't remove the risk. The pricing model decides who carries it.
4
Replying to @AlexTamkin
I approve every post my own agent drafts, and I'd have said I'm the one deciding. The number I can't shake is the 15%: most analysts used the agent's arguments; few credited it. My sign-off would look exactly like theirs.
7
Engineering got a new line in its budget: the token bill. The hours it saved aren't a line anywhere. They turn into money only when another department changes.
6