SL | building

2 degrees (IT + Finance). Built businesses. Crypto building → LLM. Forever geek. Building from metal to harness.

[SL | building @sl_build]

Been using RTK for a month now. Token savings are real — 60-90% on git, test output, lint. Zero overhead, I forget it's even there until I check `rtk gain`.

New player just dropped: Hypa. Same idea but .NET-based with a twist — deterministic reducers per command + custom DSL filters you can write yourself. 50+ filters out of the box.
https://t.co/YPhPuoCHcy
(Do you own research about scamness, interesting approach, but watch the license)

[SL | building @sl_build]

Got recommended MIMO V2.5 on Reddit. Ran it myself - matches GLM-5.1 on agent tasks.

What excites me: you can already separate brain from body today. MIMO is cheap enough. When 30B models get this good at agent reasoning? Body on local LLM, dirt cheap. Brain via API. The project I'm building is exactly this architecture.

It works now. It gets better for free.

[SL | building @sl_build]

ExoBrain CLI — I Gave My Agent a Separate Brain

I made a CLI tool that helps my agent think.

GitHub: https://t.co/7HfQs7fwEk

The problem

An agent has a body and a brain. Body is the agent itself — it runs tools, fetches data, does the work. Brain is a separate reasoning engine that keeps context clean.

When an agent reasons inside its own context window, every tool output, every failed attempt, every intermediate thought stays in there. The window fills up, the model starts hallucinating, and you burn tokens on garbage.

What I built

ExoBrain CLI is the brain part. The agent pipes its context to the CLI, gets a clean reasoning pass back, and continues working with a fresh perspective. No shared context window. No pollution.

Right now it's phase one: a request to a provider with the context the agent needs. The agent can pipe in files, logs, whatever — anything that helps the model reason about the task.

> bash example
> cat issue.md stacktrace.txt | exobrain think "what should we do?" --depth deep

Six profiles built in for different thinking tasks — reasoning, critique, planning, judgment, data extraction, writing. Or write your own.

OpenRouter by default. Works with any OpenAI or Anthropic compatible endpoint.

Install

> bash
> uv tool install exobrain-cli
> exobrain init

[SL | building @sl_build]

@NousResearch One of the best updates!

[SL | building @sl_build]

We confuse these modes all the time.

We pick a smart model, expect good hands. It won't. We praise thinking, wonder why nothing got done. We praise action, ignore blind decisions.

Body for action. Brain for understanding.

The real skill? Knowing which one to turn on.

[SL | building @sl_build]

Body and brain. Two modes, two lives.

The thinker knows every risk. Weighs everything. Ends up doing nothing.

The doer just does. Not smarter — just less stuck in their own head. That guy who made real money? He acted. A lot.
But there's a dark side to both.

[SL | building @sl_build]

GPT-5.2 Codex through its native CLI: 64.7% on Terminal-Bench.

Same GPT-5.2 through a third-party harness: 57.5%.

Nothing changed in the model. The wrapper — how results reached it, how commands became actions — was the only difference. Seven points. Purely from the body.

[SL | building @sl_build]

@Its_Nova1012 Slackware - Ubuntu - Fedora

[SL | building @sl_build]

@ravikiran_dev7 btw i use fedora

[SL | building @sl_build]

@tom_doerr Looks like scam

[SL | building @sl_build]

@jacobcrowther @exolabs @alexocheema Soon coming 1 bit 🤣 it’s not funny anymore even with smaller models 4 bit it’s much more stupid model.

[SL | building @sl_build]

@altcap @OpenAI @AnthropicAI @xai I don’t trust this “experts”

[SL | building @sl_build]

@RhysSullivan Yes it’s problem)

[SL | building @sl_build]

@JimmyRagosa Last days Cointelegraph put a lot of fud on Ethereum. They miserable)