petersson.eth

Nice event by "Vienna AI Engineering". at T-Mobile where I had the privilege to talk about best practives for CLI creation: youtube.com/watch?v=OxlST7…

@bentlegen @tobi TUI? FPS? why are the two terms in one sentence? I my simple world a tui should just statically draw something whenever there is an update.

using @tobi's autoresearch Pi plugin to optimize my TUI app's framerate 4 fps → 32 fps in 6 iterations Will check back in after sleep ...

Past 5 Months: Total savings €486.54 by selectively charging the battery at opportune times. from github.com/apetersson/cha…. Ran pi-autoresearch to optimise the prediction algorithm of house electric demand - to squeeze out like 5€ extra compared to "intuitive" ML tuning.

@mrlesk @badlogicgames trying to walk the walk myself: github.com/MrLesk/Backlog…

@yellowhatcoder @badlogicgames That’s why my recommended way to use backlog.md is with mcp because this way agents hit the same backlog.md instance and I can guarantee atomicity. With multiple cli calls in parallel I would have to use some sort of locks. I will look into it

if an agent emits parallel tool calls, pi used to execute them sequentially. funnily enough, only a handful of people complained. welp, just implemented it. caveat: you'll get flicker if you have many parallel tool calls and their outputs are > terminal height. soon to be released.

@badlogicgames @vokaysh Inside the core user prompt the Agent always just sees a small slice of the big picture "add feature X" and does so minimally. It can effortlessly add a 26th if condition without revolting. If there is nobody willing to say "let's clean this up this time" it will devolve.

i can't speak for david. what i see is this: if you let agents build or extend a codebase with only minor or no supervision, you get unmaintainable garbage, because the agent makes terrible decisions that compound, both big and small. those decisions make it hard for both you and the agent to keep modifying the code base, until eventually it's unrecoverable. why does the agent make bad decisions? i can't tell for sure, but my gut tells me that training data can currently not capture the holistic thinking needed to design and evolve complex systems. that's one part of the problem. related to that, and oversimplified: agents output the "mean quality" of the code they saw during training. most of that code is very bad. specifically tests, which humans are terrible at writing at. another part of the problem is that specification via prompt is not precise enough, so the agent has to fill in the blanks, giving it enough rope to hang itself. the more detailed your spec gets, so the agent gets constrained and less likely to produce crap, the closer you are to handwriting the code yourself, as that's the most detailed version of the spec that can exist. so then you gain nothing. back to prompt spec it is, which means the agent fills in blanks, which means we get suboptimal or truely bad results. using agents can still be a net productivity boost (see other posts in my thread), but it is not easy to come up with consistent workflows that produce both production quality maintainable code while retaining the speed advantages agents give you.

recommended reading sure to ruffle some feathers. but it's largely true for now. keeping the complexity off the bay is really hard, espwcially if you go full agent orchestration. even if you don't, and human in the loop a lot, automation bias kicks in and your reviews of agent generated code become mostly performative.

@mitsuhiko That was LM Studio related. I switched to oMLX and it seems to work better with caching and thinking is rendered correctly now. still, for my current workloads, lots of PP time. oMLX: 83.9% cache hits. Prompt Processing (excl. cached) 204.2 tok/s Token Generation 30/sec

@yellowhatcoder Looks like you also get <think> tags

Got my new macbook pro and I can report that all of my attempts of making local models work with pi so far have been disappointing. Not sure what else I should try :D

@sukh_saroy I have no idea why @intellijidea isn't providing this since 12 months ago. It theoretically has the tech, especially for half broken code, which is all code while it's written. Running out of excuses to continue my 25y IntelliJ usage and licenses.

🚨Breaking: Someone just open sourced a knowledge graph engine for your codebase and it's terrifying how good it is. It's called GitNexus. And it's not a documentation tool. It's a full code intelligence layer that maps every dependency, call chain, and execution flow in your repo -- then plugs directly into Claude Code, Cursor, and Windsurf via MCP. Here's what this thing does autonomously: → Indexes your entire codebase into a graph with Tree-sitter AST parsing → Maps every function call, import, class inheritance, and interface → Groups related code into functional clusters with cohesion scores → Traces execution flows from entry points through full call chains → Runs blast radius analysis before you change a single line → Detects which processes break when you touch a specific function → Renames symbols across 5+ files in one coordinated operation → Generates a full codebase wiki from the knowledge graph automatically Here's the wildest part: Your AI agent edits UserService.validate(). It doesn't know 47 functions depend on its return type. Breaking changes ship. GitNexus pre-computes the entire dependency structure at index time -- so when Claude Code asks "what depends on this?", it gets a complete answer in 1 query instead of 10. Smaller models get full architectural clarity. Even GPT-4o-mini stops breaking call chains. One command to set it up: `npx gitnexus analyze` That's it. MCP registers automatically. Claude Code hooks install themselves. Your AI agent has been coding blind. This fixes that. 9.4K GitHub stars. 1.2K forks. Already trending. 100% Open Source. (Link in the comments)

Heading to Berlin, to learn and share experiences about @OpenClaw at luma.com/ek5cb0vq

@TheMiamiApe Yes, payment channels based on x402: x.com/i/status/20122… happy to connect.

Anyone working on agentic-payments?

@Rasmic @steipete I often do this, like Thesis, Antithesis Synthesis between different models. For high level plans it works wonders especially if I remove the nonsense in between the steps. Is there a pi.dev plugin fo this possibly already?

I got both gpt-5.4 and opus 4.6 to generate a plan... I gave gpt's plan to opus and it admitted that it was a better plan lol

The @numerai Council of Elder is excited to announce @yellowhatcoder will be speaking on "Securing @openclaw Agents" for our Decentralized AI Day in Vienna on March 21 Join us in Vienna next weekend with the Luma link below ⬇️

@badlogicgames Switching model helped for me. Only 5.4 unstable today.

Anybody seeing this? Codex endpoint craps out, then recovers.

@kevinkern this is one of my projects currently: (works, shorter) this is green field development. do not keep any fallbacks, defaults, migrations, etc. always tend towards homogeneity. be human- and machine-readable. never default to a "first" entity.

updated my AGENTS.md again to prevent codex fallback hell.

@steipete @BenKraus On my machine the desktop app is slightly less laggy than codex-cli, just subjectivity. The plan/ask interaction feels better on desktop. Lacking for me is the process output awareness for subagents and processes

@BenKraus it's decent, but agree there's room for improvement.

Oh right i realize hell freezes over, we reached the point where app > cli That combined with the speed increase also means less windows necessary, codex goes brrrr now! developers.openai.com/codex/app/

Your @openclaw agent works perfectly for 20 minutes. Then it silently forgets your instructions and goes rogue. The fix isn't better prompts. It's understanding how memory actually works. I'm an OpenClaw maintainer. After 2+ months of daily use, I ended up with a memory system where my agent remembers decisions from weeks ago, checks its notes before acting, and picks up exactly where it left off after every restart: 0:00 The real reason your agent forgets 1:59 Quick start: 3 things that matter most 3:01 The 4 "memories" people confuse 4:52 3 failure modes: why your agent forgets 6:42 Compaction vs pruning 8:05 Prove what's loaded: /context list 10:06 What compaction actually does 13:42 Layer 1: pre-compaction memory flush 16:27 Layer 2: manual saves + /compact trick 19:12 Layer 3: the file architecture 23:22 The memory protocol for AGENTS.md 26:06 Retrieval: Track A, A+, and QMD by @tobi 31:59 Cost and cache 33:33 The complete config (copy and go) 36:35 5 things to remember

@robzolkos @badlogicgames 🍰?

@badlogicgames Are you allowed to use this with.. never mind

woooh

@AmbsdOP I'm not an expert on this HW layers, but i just notices ANE is still on zero power when using ollama or LM studio (just heavy GPU), but on github.com/apetersson/imm… when using the CoreMLExecutionProvider ANE is able to do inference very efficiently. Can we do MLX on ANE now?

YES! Someone reverse-engineered Apple's Neural Engine and trained a neural network on it. Apple never allowed this. ANE is inference-only. No public API, no docs. They cracked it open anyway. Why it matters: • M4 ANE = 6.6 TFLOPS/W vs 0.08 for an A100 (80× more efficient) • "38 TOPS" is a lie - real throughput is 19 TFLOPS FP16 • Your Mac mini has this chip sitting mostly idle Translation: local AI inference that's faster AND uses almost no power. Still early research but the door is now open. → github.com/maderix/ANE #AI #MachineLearning #AppleSilicon #LocalAI #OpenSource #ANE #CoreML #AppleSilicon #NPU #KCORES

🇦🇹 Looking for a hands-on, zero-hype event to do your own due diligence on the future of #DeAI? Look no further — join us for Decentralized AI Day #Vienna on March 21, 2026! 🚀 Level up with expert sessions from: 🧠 @NumeraiCoE & @numerai Community 📊 #TabPFN with @ph_singer (@prior_labs) 🌐 @AlloraLabsHQ with @apo11o 🕊️ @flock_io with @YifanX 🔒 Hardening @openclaw with @yellowhatcoder (@tunetagnet) 💻 Bring your laptop — we’re going hands-on in the afternoon with workshops from @prior_labs, @numerai, and @AlloraNetwork. 🎟️ Get your ticket: luma.com/sb1g8oyb?utm_s… No hype. Just #AI and #Web3 builders. See you there!