Janosh Moshiri

The world is training models or building agent tools. I'm building the architecture around it. MarkdownLLM: typed, version-controlled domains you own. MIT.

[Janosh Moshiri @JanoshMTM]

The Paradigm shift: Treat the language model as a reasoning processor, build a loosely coupled, modular software engine around it so that a non deterministic processor can be at the center of a system that is consistent, auditable, and resistant to drift.

[Janosh Moshiri @JanoshMTM]

What you have described in the question we as dev's are all facing at the moment.

How do we know this is an intention not just a prompt stream that got us to a place of excitement created by the LLM not our engineering skills.

I think the way to answer that is by slowing down and using the LLM as our development partner not our problem slover.

[Janosh Moshiri @JanoshMTM]

@karpathy I watch a YouTube video of your a month ago and I thought you said the paradigm shift was the LLM as the processing engine and the infrastructure being redesigned for it, not just another harness add on.

[Janosh Moshiri @JanoshMTM]

@natebjones What your describing is at the root of my new agent framework I have just built.

The partnership between software engineer and agent. The synergy of powerful raw reasoning, and an engineer with a goal.

https://t.co/InBZZDdHnB

[Janosh Moshiri @JanoshMTM]

The model is not the bottleneck.

Prompt engineering. Harness engineering. Looping.

Everything we are focusing on is something we put around the model. Structure we define before it starts work, the rails that keep it on the plan or the techniques we use to manage the context window.

MarkdownLLM is the next layer of that. The one nobody's named yet.

The model thinks. The structure holds!

[Janosh Moshiri @JanoshMTM]

https://t.co/InBZZDdHnB

[Janosh Moshiri @JanoshMTM]

Everyone's building agent memory, loops and more tool, OpenClaw, Hermes, mem0, Letta.

But the LLM isn't just forgetful. It's a processor with no architecture around it.

So I didn't build memory, loops and tools, I built the frame, typed structure in git, validated on every commit. The piece that pins nondeterministic thinking to a deterministic store.

The model thinks. The structure holds!