‏ deepfates

whoa this is interesting, i bought it and tried to play and got confused and surprised so that's a good sign. can you say more?

thank you 🙏 it is a great honor

I wish someone would argue with the bad faith anti-AI people they way they do.

Go to Emily Bender like "you're triggering me by using the hostile register of the tenured academic". Call Gebru a googler. Say Emil Torres is doing eugenics by not having kids.

If we're all gonna be stupid anyway...

not now honey 🙄🙄

I'm making dark fantasy movies with our cats🙀 🧙‍♂️

currently they are not transparent, intentionally I think, to preserve the non-leaky abstraction. I'm interested in what happens if the traces are large to the context as well though...

crazy remembering websim & how weird pre RLHF locked in was, and then coder RL on top. base model pretending it's a computer terminal instead=whole diff universe. like early PC/early internet convos, there were so many branching paths, so many xanadus waiting outside of the trillion dollar gold rush

We are no longer building single chat-thread instance minds. We are building an intelligent System of software, hardware, and wetware, a program that writes programs, where each of us is a moving part.

Haven’t you always wanted to be part of something?

Right now, the models only know “user” and “assistant” and “tool”. Future post-training will have to find patterns for a multi-agent, multi-channel world.

In a way, the user and the sub-llms become tools within the environment.

These new agents will feel less like coworkers, or slowly dying houseplants, than what we have now. They will feel like councils, senates, hive minds. They will be composed of many minds working and communicating.

You can have an ongoing conversation where the long-term memory is offloaded to a variable, and the agent can operate over its own memories, so the “context window” is effectively infinite.

The context window hasn’t grown almost at all, no matter how much work the sub-agents did, because they can send variables around without reading what’s inside them.

This week I built a recursive language model in typescript, and watched it process over 6M tokens without degrading from context rot. It feels… different.

You ask it to do something and it writes code that spawns more of itself and then merges back just knowing the answers.

The RLM hooks up the language model directly to the REPL and puts the context in as a variable. The LLM can operate over large documents or its own memories or whatever is in that context variable. It can call the human user with a final answer, or call a sub-LLM with its own {prompt, context} input

The Recursive Language Model (RLM) pushes this a step further: What if we don't have a "chat" at all? What if we make the user just a function inside the computer? In fact, what if we put the agent inside the computer too?

The squishy human is no longer the whole environment. There is an ontologically hard reality and it's the computer world, the files and folders and programs within.

The user is still in a privileged position. The outer conversation loop with the model is still "the main world". But it can do stuff in its computational environment and that stuff persists.

So post-training baked in computer use, shell environments, bash tool, and suddenly you have models that are good at predicting what a computer will do in response to complicated programs and commands.

Computers that can use themselves.

People realized the model didn't have to simulate just a chatbot. It could simulate a computer. Or a computer user.

Once you get RL with verifiable rewards, computer environments suddenly became the thing for models to learn. You can tell whether your model did it right by how well the code runs!

(Read Eval Print Loop, an interactive computational environment, like a terminal or a programming language notebook). And we’ve seen models simulate computer environments as well, like in worldsim and Infinite Backrooms.

But context is still the bottleneck, and all the tool call inputs and outputs still have to go through the input.

In the meantime we had things like CodeAct, and Code Mode, and Claude Code, that give the model access to some kind of REPL

Right now most people's working definition of an agent is "a language model in a loop with some tools", right? This is the ReAct paradigm. The models are trained around it, from data formatting to inference stack. Tool calling, function calling, MCPs all rely on this turn-by-turn structure.

In my quest to understand the true nature of an Agent I have been thinking a lot about the loop and the actions and the environment. And I think i see where we're headed next.

Agents are not going to be like chatbots. They are going to be like programming languages come alive.

Maybe next time ♥️

Thank you it's an heirloom

Come to Zeitgeist tonight for OFFICE HOURS! Meet your mutuals! Kick it with the haters!

Zeitgeist,Tuesdays, 6pm. Look for the duck. Or just ask the bouncers where to find the meetup dorks

officehours.lol/sfo

Pull up and let's talk it out

This is just nutpicking btw, i have no idea who OP is. You don't get to pick your followers 🤷

Just found out people here disagree with me. I'm flattered. Feel free to do it to my face, I don't bite

Heated Rivalry is basically monster porn but if the monster was a Russian guy

daily affirmations