inference @poolsideai, building https://xitter.gq/t.co/uEbfxzEztO
Zurich, Switzerland
Joined May 2020
- Tweets3.7K
- Following1K
- Followers2.6K
- Likes30.1K
Pinned Tweet
Announcing hip-agent ("harness in prompt").
The basic idea is that the harness is so small that it can fit in the prompt, which allows the model to know exactly how the harness works.
jonathanc.net/blog/hip-agent…
it's funny that the robot tried to wash its hands at the end like a human.
it probably learned from human data
You can now book our humanoid cleaners on a recurring subscription, for up to 3 hours per cleaning at tau-robotics.com.
Video is first take, uncut, shown at 4× speed. See the reply for detailed footage at 1× speed.
le french saw h-net and decided to tax bytes instead of tokens
So, the French socialists, and Parliament members just voted a tax, that they don’t understand, which would cost VideoLAN more than 1 million euro per year.
They are probably the most clueless politician on Earth? And they claim they want technological sovereignty…
@LaurentLhardit
they cited my blog post and pushed the trick further to the backward gradient for MTP, very cool!
author: @jayden_teoh_
arxiv.org/abs/2610.38331
Improving MTP by only keeping self-distillation of next token prediction in the loss term.
my Claude found that Qwen3.5-27B was grown from Qwen3.5-9B (layer looping)
qt: independent discovery found by grok
in-context tokenization is super cool
I've always believed that intelligence should come from simple, general building blocks that learn their own abstractions directly from raw data; in other words, the bitter lesson. I think we're finally seeing a real step towards AI that can learn from all data.
Language modeling on raw bytes is a microcosm of this ideal, and to me it's a central problem in AI: if we can't build LLMs that learn end-to-end from bytes, how will we handle the rest of the world's data that we understand even less?
Last year we introduced hierarchical networks (H-Nets) as a first step. Our new version finds up to 3-5x efficiency gains in controlled scaling law experiments, and at larger scale matches an open-source model trained on 11T tokens with ~10x less data and compute. The posts below walk through what we found and why we think it works.
ChatGPT now has this new thing where it sometimes takes 4 seconds to respond to a simple prompt like "hi" even on instant model
hot take:
recent trending neuron steering method is not practical. it's steering the residual stream along the output direction of a neuron. it can't be absorbed into the weights.
imo a better way is to use neuron lens to select params for training. it can be superior to lora.
