Madhav 🦉

23, Building Autonomous Sys, Prev @IITDelhi, AI Engineer @hashed_official, Research @veyrna, @lawrenceai

[Madhav 🦉 @5mknc5]

Predictions: Impact of superhuman AI over the next decade will be enormous, exceeding that of the Industrial Revolution

Stumbled upon this a few months ago:
https://t.co/SWo9mycDBT

[Ram @ramxcodes]

Coding is a muscle. If you don't use it,

you lose it. It's that simple.

[Madhav 🦉 @5mknc5]

@ramxcodes W take

[Madhav 🦉 @5mknc5]

@Adidotdev Connect?

[Madhav 🦉 @5mknc5]

@pmddomingos Totally agree

[Madhav 🦉 @5mknc5]

Which open-source LLM do you use for coding GLM-5.2 or Kimi-K3?

[Madhav 🦉 @5mknc5]

yeah, design systems where you inject a tiny bit of intent and then the loop takes over: plans, acts, checks, iterates, remembers.

your job becomes building the memory layer, the verifier, the triggers.

once it's running you should mostly be out of the loop.

[Rahul @sairahul1]

Andrej Karpathy said it perfectly:

> "Remove yourself as the bottleneck. Maximize your leverage. Put in very few tokens, and a huge amount of stuff happens on your behalf."

that's exactly what loop engineering is.

most people think AI engineering is writing better prompts.

it isn't.

it's designing systems that keep working after you've stopped typing.

that means:

→ triggers that start work automatically
→ loops that plan, execute and iterate
→ independent checkers that verify every output
→ persistent memory that survives across sessions
→ observability that catches failures before users do

the model does the work.

the loop decides what happens next.

the harness decides whether the result is good enough.

that's where the real engineering is moving.

I dive into loop engineering, harnesses, verification, memory, Graphiti, Opik, and production agent design in much more detail here:

article quoted below.

[Madhav 🦉 @5mknc5]

@tszzl please elaborate on those points in plain language with context so I can understand better

[Madhav 🦉 @5mknc5]

@rasbt Mix short/long traces so the model learns when to think hard.

[Madhav 🦉 @5mknc5]

How LLMs Switch Reasoning Effort?

At inference, it's a compute dial:

low-effort = fast greedy responses
medium = moderate CoT
high = long hidden reasoning traces (o1-style test-time scaling)

Training teaches it via process supervision + RLVR on verifiable tasks.

@rasbt

[Sebastian Raschka @rasbt]

How can an LLM switch between low-, medium-, and high-effort reasoning? And how does an LLM learn to reason more or less?

I put together a “little” article explaining how these effort levels are implemented at inference time and during training.

[Madhav 🦉 @5mknc5]

Github: https://t.co/xWRCrcMWLl

[Madhav 🦉 @5mknc5]

G0DM0D3 is a an open-source, privacy-transparent, multi-model chat interface that pushes the limits of the post-training layer
for red teaming, cognition research, n liberated AI interaction

Built for hackers, philosophers, and system tinkerers.

https://t.co/qE0nQzBnLc

[Madhav 🦉 @5mknc5]

@sattyyouneed Chinese can't lead the space. They need someone to lead and the just follow and somehow match the results.

[Madhav 🦉 @5mknc5]

@sattyyouneed Chinese can't lead the space. They need someone to lead and the just follow and somehow match the results.

[Madhav 🦉 @5mknc5]

@chamath The cheap Chinese models are subsidized by the state. Same old playbook: flood the market, bankrupt domestic labs, create total dependence. Except now it’s AI, and the counterparty is China.Open-source doesn’t fund frontier training. If you’re not paying, you are the product.

[Madhav 🦉 @5mknc5]

@scmallaby @chamath If US labs go all-in on open source, the entire economic upside collapses. All the capex, infra bets, talent concentration, and $1T+ valuations (openai, anthropic) evaporate.

Frontier AI becomes a commodity race to the bottom against subsidized Chinese labs.

Poor game theory

[Madhav 🦉 @5mknc5]

@yacineMTB Wait for the release of kimi k3.1

[Madhav 🦉 @5mknc5]

@scmallaby @chamath If US labs go all-in on open source, the entire economic upside collapses. All the capex, infra bets, talent concentration, and $1T+ valuations (openai, anthropic) evaporate.

Frontier AI becomes a commodity race to the bottom against subsidized Chinese labs.

Poor game theory

[Madhav 🦉 @5mknc5]

@chamath The cheap Chinese models are subsidized by the state. Same old playbook: flood the market, bankrupt domestic labs, create total dependence. Except now it’s AI, and the counterparty is China.Open-source doesn’t fund frontier training. If you’re not paying, you are the product.

[Madhav 🦉 @5mknc5]

@chamath The cheap Chinese models are subsidized by the state. Same old playbook: flood the market, bankrupt domestic labs, create total dependence. Except now it’s AI, and the counterparty is China.Open-source doesn’t fund frontier training. If you’re not paying, you are the product.

[Madhav 🦉 @5mknc5]

Grok 4.6 with ~2T params will finish training next week.
confirmed by @elonmusk