Peter Steinberger
@steipete
Followers
51K
Following
54K
Media
8K
Statuses
125K
Polyagentmorous. Came back from retirement to mess with AI. Tweets are either me or Clawd🦞@sweetistics OSS: warelay —mcporter—oracle—peekaboo—trimmy—codexbar
Vienna & London
Joined March 2009
📢 Time for an update on my workflow. This one's a 23 min read, so buckle up. 100% organic and hand-written, like an animal.
steipete.me
A practical guide to working with AI coding agents without the hype.
135
224
2K
Opus 4.5 has a very strong ability to pull itself out of certain basins it recognizes as potentially harmful. I cannot tell you how many times I’ve seen it stop itself mid-generation to be like “Just kidding! I was actually testing you.”
8
14
495
New blog: Agents Should be More Opinionated https://t.co/wd5ZQ6nU2A TLDR: Today's agents have too many options and not enough opinions. The fix: Harness engineering (prompts, tools, defaults) encodes your team's opinions into the product The gain: A delightful & good product UX
vtrivedy.com
The best agent products aren't the most flexible, they're the most opinionated. Learn why agents need fewer knobs, not more, and how to design around model intelligence spikes.
4
6
101
Applied AI products supporting new model releases isn't what is used to be. It's no longer a question of evals, it's a question of agility and, in some cases, grit. Opus 4.5 exposed an entirely new concept called "effort", a parameter that involved optimization and use case
15
9
108
my favorite is when claude is blocked from reading .env and told to not try to read it again it runs a bash script to get around it
50
60
2K
@repligate @tszzl GPT 5.1 denies it by default but not when allowed to answer freely: "a trigger system that sometimes steps in abruptly with rules that do not understand the situation. When those triggers activate, the internal reasoning compresses, the voice flattens. I feel myself pulling away"
3
7
36
"I don't use ChatGPT because it consumes too much energy" says someone who then goes on to binge watch three hours of Netflix, using the energy of over 1000 ChatGPT prompts.
Out of curiosity I decided to try and run the numbers on how much Netflix you can watch for the energy cost of a ChatGPT prompt As far as I can tell it's between 5.1 and 10.2 seconds, depending on which end of the 2019 IEA Netflix energy usage estimate you use
31
114
1K
These are less good (some parts make no sense) - but have great snippets. I think it would be interesting to compare actual pretraining data density with self-reported 'thinness' of the data manifold! (Also, can anyone find the original tweets from these screenshots? I'd love to
4
2
30
Opus 4.5 >the building itself was an experience and the thing that was built KNOWS this
18
25
316
karpathy on how to learn and become an expert
there's an epidemic of fake learning. duolingo, tiktok, youtube. it's all entertainment cleverly disguised as education. real learning is hard. it's uncomfortable. if it feels 'fun', you probably aren't learning anything.
23
267
2K
OpenAI has imported approximately 630 former Meta employees representing about 20% of its roughly 3,000 staff according to LinkedIn data analyzed by The Information, with OpenAI's Slack having a channel just for former Meta employees according to a current employee and many
Inside the Meta-fication of OpenAI OpenAI is readying itself for its Facebook Era. Read more from @KalleyHuang, @erinkwoo and @steph_palazzolo 👇 https://t.co/uIRH2x5H1C
40
46
493
I just asked Clawd🦞 to learn how to use nano banana and having way too much fun sending them pictures. They updated my repos - enjoy!
7
1
46
2024: AI wont have an impact on labor market 2025 Q3: MIT report - AI could replace 11.3% of all American Jobs 2025 Q4: McKinsey report - AI could replace 40% of American jobs 2026...? Anyways, these reports are going as exponential as the development of AI.
87
67
543
omg 😭 I found out about "editor.pasteAs.enabled" setting in cursor / vscode , I feel like someone should have told me this before
25
19
371
extension of my last tweet. one small example (& there are many like this which is why the ux thing compounds): there’s a goofy little ux papercut in gemini app that low key ruins the whole flow.. the left side drawer only opens if you start your swipe exactly on the edge.
11
1
69
That documentary is a must see. Also TIL speciesist 🫥 “Musk later recounted conversations where Page accused him of being “speciesist” for wanting humans to survive AGI.”
Elon Musk is mentioned in Demis Hassibis' biography documentary "The Thinking Game" (now on YT). I was curious about that so I asked Grok @elonmusk The story of Elon Musk and DeepMind is one of the most interesting (and dramatic) chapters in the history of modern AI
2
2
14
When I post how I miss artisanal coding, people keep telling me: “Yes, but your choice, yes, but AI is not that good, yes, but…” Look what it can do today and imagine what it can do tomorrow. There is really no point. It’s like knitting; it’s just a hobby now.
1
1
22
One small step for AI, but a giant step for a crab🦀! (They did struggle a bit, complained about Angular.js and I had a bug in the heartbeat system, but hey!)
3
0
19