zlumer.eth

@haipingfu thanks for sharing, this is a very important area. how would you compare it against Aider Repomap? similar/different/better/worse?

Building a code graph isn’t hard. Building a high-quality code graph is. Accuracy, cross-file/cross-language resolution, incremental updates, versioning, performance, and language semantics all matter for AI agents. I hope the next version of Compass will close many of these gaps. Question: do you know any open source evals framework for code graph ? github.com/crabbuild/comp…

@matteocollina State machines + TEA is even better (zlumer/tesm on gh)

State machines: let's teach LLMs to use state machines. LLMs, like humans, forget the "refactoring" step of TDD.

@dillon_mulroy tesm

hmmm where i have i seen this pattern at before

@valigo algebraic data types are bad because once you learn about them you realize that all programming languages without them suck and software is inherently broken and you can never go back

@NathanFlurry at 2000 tok/sec it's around 800 tok — not there yet but also much better than whatever we get with so called "frontier" labs

IBM coined the Doherty threshold in 1982 that determined how fast computers need to respond to maintain flow state 400ms is the magic threshold Maybe Cerebras will get us here

The limiting factor of software engineering has never been code production. It is complexity management.

@RyanCarniato I'll be glad to test it then

@zlumer f(State) -> View as implemented in libraries that tie state updates the UI structure(components) has some fundamental architectural concerns. So better than f(State) -> View as presented in React, definitely.

I wonder if people think innovation is dead in the frontend? At this point it's just incremental performance gains? That there isn't anything to gain architecturally by changing the formula? It would partially explain why there is resistance to move past old React and excitement around things like Vinext or Redact. The lower excitement around things like Qwik's resumability. We're never going to move past React unless we move past React.

@zeeg @ryanvogel @LukeParkerDev /move

@ryanvogel @LukeParkerDev where da worktrees at

ok this shit better not be broken on wsl

@NathanFlurry now that I (hopefully) understand what agentOS is, it feels like PGLite to Postgres: a minimal in-memory implementation of a widely used standard what didn't help me was calling it an "OS" — I didn't understand why would I need an OS for an AI agent, seemed an overkill

@zlumer happy to hear! we've been working with messaging a bit, tricky to get it right

Favorite bit of the new agentOS landing page: how we explain the tech microVMs (sandboxes) and WebAssembly (agentOS) are two very different technologies with different use cases: - Library, not extra infra - WASM, not microVM - Direct function calls, not APIs - Scoped access, not exposed credentials - One process, not a VM per agent - Runs anywhere, no nested virtualization Plus we've published plenty of docs on the architecture

@thdxr @i_mika_el I clicked "Yes" but if it has 5-h limits than hard No

@i_mika_el it would have 5hr and weekly limits as well

i know this isn't the most appealing thing ever but i'm curious on the state of things would you pay for a $200/month sub that gives you access only to opensource models? say you got ~$800/month in usage

@matteocollina I had to scroll a long way to the example to understand what is going on — maybe best to make the example a bit more visible because this is really outrageous when you finally understand what's going on

A package registry should not be picking winners

@ibuildthecloud I was building a similar product, also discontinued a few months ago: zlumer/hyperbranch on github

@zlumer Truly curious what was the main thing you liked. I'm still building something in that area, just a different take. Maybe you'll still like it.

Here's my last 3ish years. RIP: Discobot Nanobot (it's alive in a weird form still, but killin it) Otto8 Clio GPTScript Out of those clio and discobot I like the most, but honestly all of them were really good ideas. Just wrong shape, wrong time, something. It's fine. The fun thing for me is each project I had some strong opinion about some direction and each one came true in that everyone else eventually did basically the same thing. So at least is reconfirms I'm not an idiot, but just that I suck at coming up with products :) But I think that's kind of the trick, the rest of my team at obot is really good at product. I'm just good at being obnoxious about tech.

@RyanCarniato having spent almost two decades in imperative UI (gamedev) and then a decade in declarative reactive UI (web), I have a very strong opinion on what kind of problems I'd rather solve, both with or without AI help concepts matter even more when the code is abstracted away

I feel like we are on one of those pendulum swings right now where people undo a decades worth of work only to realize in 5 years they over swung because they missed the actual problem. We see this all the time. I'm not going to argue against imperative handles in all cases. But I want you think about the implications. If you unwind these things you remove guarentees. Like why do developers like Rust or TS? Its because through constraints certain classifications of issues aren't possible. The solution solves itself structurally. Structural constraints also have a nice property when composed. You keep benefitting from them as you layer on complexity. Build more constrained systems. Like if you aren't batching synchronous updates you probably have no foundation to managing async race conditions structurally. We shouldn't confuse incidental simplification(removing a problem) with structural simplification. Sometimes legitimately we don't need to solve certain things. The solution should fit the problem. But more often than not in a well established space we are just delaying the inevitable.

Not sure who needs to hear this, but... AI changes transport/distribution. It doesn't alter physics. When code becomes cheap to produce, the limiting factor becomes how much complexity your architecture can absorb.

> Read this and your next model will be 10x smarter. < Nobody knows what intelligence truly is. We just know models are converging to being smarter, as they train. Yet, we DO know some of the fundamental features of intelligence. And when one of these features is neglected or not trained for, then there is no way for a model to obtain it. Neglecting an aspect of intelligence hinders a model's general capabilities, in a way no amount of flops can compensate for. I'm making this whole post to convince you there is ONE fundamental aspect of intelligence that YOU are neglecting, underestimating, and under-training for. Anyone using models 24/7 can see this weakness. It is blinding, glaring, as clear as skylight. That feature is: ✨ erasure ✨ Removal. Compression. Garbage collection. Models are not sufficiently trained for that. They are trained to ADD information. Not to REMOVE it. You ask a question. They give you an answer. They work in a project. They write files. You post a bug. They craft a solution. They're only indirectly, if at all, rewarded for removing information, or compressing information. This is a huge mistake, because erasure is a cornerstone of intelligence. The human brain has several mechanisms entirely dedicated to removing information. Short term memory, long term memory, sleep, all mechanisms to throw garbage away. Furthermore, grokking is nothing but a compression event. An aha-moment happens when your brain is capable of expressing new information in terms of information you already posses stored. This is what allows that info to be stored. That is how you learn. Erasure isn't a small feature, erasure is *THE* underlying driver of intelligence. It is what allows us to keep absorbing tons of information and still managing to turn it into useful capabilities. Intelligence is not about producing good knowledge, it is about removing bad knowledge. So, erasure is half of it. So, my advice to you: take erasure seriously. Train on it. The architecture is fine. It can lead to AGI. But you won't be a complete athlete if you skip leg day. Reward your model on the other half of intelligence. Teach it how to erase and compress information competently. Make this a big program in your company. Have entire teams dedicated to this. "I'm already kinda doing that!" No you are not. And if you think you are, take this as a signal you should do 100x more of it. I want to be very clear here: erasure is HALF of intelligence. So, if not half of your FLOPS are flowing into erasure, you're wasting your GPUs, and no optimizer can compensate for that. "But how do I teach a model to erase?" Literally, just ask it to compress a text, then reconstruct it, and ask questions to assert how lossless the conversion was. That simply. You can do that in any dataset. For coding, a more effective way is to take a big codebase and ask it to make it shorter, while still preserving the same behavior. IMPORTANT: avoid code-golfing / minification / uglification. Removing comments or making variable names shorter IS reward hacking. Counter that by counting the NUMBER OF BRANCHES. A branch is: an "if", a "match", a "case". That's THE complexity of your program. Count it, and ask the model to reduce it. There's no way to do so other than building better abstractions. And a better abstraction is nothing more than a blob of information that lets you throw other information away, because it expands into the information that was just discarded. Train on that, and your model will be incentivized to build better abstractions. Do you know what we call humans capable of building better abstractions? Geniuses. So, please: appreciate the full nature of intelligence and give your models the rewards they need to train on all of it. Let erasure be a major part of your training programs. Do not skip leg day. Thanks for coming to my TED talk...

@zeeg it's a good focused patch though, a pre-LLM human could sometimes spend hours or days debugging to make a single line of code change

Guess how long this patch took to generate, and what the cost was? Its using Sol primarily.

@thdxr unrelated but is there a chance you could get "npm i opencode" to work now that the other "opencode" is no longer maintained?

new settings screen i prompted slackbot around 5pm today and asked for some concepts it sent screenshots and videos but i was afk kit jumped in and gave it feedback. then he stepped away then i took over back and forth until it was done - it even made this video