Understand humans at scale
Global
Joined July 2025
- Tweets193
- Following20
- Followers379
- Likes359
Pinned Tweet
Do you remember Pixar <Inside Out>?
Today we bring joy, anger, sadness, and other emotions to the language model, with just 6.3M parameters.
Our model Riley (yes, named after Riley from Inside Out) is the world's first emotion model. We don’t prompt feelings. Riley takes feelings as input params.
You can make Riley show feelings by adjusting emotion params.
”emotion : { joy: 0.7, sadness: 0.1, anger: 0.1, anxiety: 0.1 }”
We found some interesting results when we triggered these emotions in Riley.
We’ll share some in this thread. 🧵
WalNut retweeted
Came across Cola-DLM(hongcanguo.github.io/Cola-DL…) from ByteDance.
A hierarchical continuous latent diffusion LM that separates global semantic planning (DiT in latent space) from local token realization (VAE decoder).
Paper is out, but no code and no HF model yet.
So I reproduced it from scratch. Happy to share with anyone interested 👇
A few days ago. At claude-code the sub-agents feature were added.
Did you know utilizing the sub-agent concept to HELL can make models like o4-mini to solve Olympiad level math problems?
This is our agent system called Crux.
It can be expanded unlimitedly. If you want, you can make the agent system thinks and work forever.
We also don't know where the limit of the performance will be. Which is super exciting.
And the most cool point is its completely open-source. Check the repo!
Hey folks!
We're Tooliense, a small team pushing the boundaries of agent systems to achieve state-of-the-art intelligence. We open-sourced our first agent, Crux a few days ago. Crux empowers existing language models to iteratively refine their own answers.
Using just o4-mini, it's cracking tough problems at USAMO and IMO levels. Mind-blowing stuff! Experiments are ongoing with epic use-cases and benchmarks popping up.
Plus, our growing community is blending diverse approaches to level it up even more.
To share our research stories and your creative use-cases, we've launched a Discord server!
We'll keep advancing Crux and more agents come collaborate!
While digging into Crux, we stumbled upon something fascinating!
We mentioned that Crux's architecture scales like traditional language models' Transformers. In its iterations, we observed phenomena akin to gradient descent a built in optimizer at play, where step sizes sometimes overshoot the optimum.
Finding the ideal step count is crucial, but we're still figuring out the exact numerical approach.
WalNut retweeted
Just got a collection of o4-mini agents to collectively critique and work together on a piece of a research problem that I solved on my own that took about a week. It’s clear at this point that the latent capabilities of these models is much larger than is accessible via 1-shot generation.
So far they have independently come up with the same approach that I ended using to find a valid proof, and the bound they suspect is the key is true and exactly what I did.
I don’t know if it will end up getting a flawless proof (it’s been running for an hour), but this problem is well beyond what people currently believe these models to be capable of, & if I had gotten these ideas / steps ahead of time it would’ve cut how long this proof took me in half at least.
This kind of thing is going to revolutionize the field very soon
Google also starts researching about self-evolving loops? 👀
I’m building a new team at @GoogleDeepMind to work on Open-Ended Discovery!
We’re looking for strong Research Scientists and Research Engineers to help us push the frontier of autonomously discovering novel artifacts such as new knowledge, capabilities, or algorithms, in an open-ended self-improving loop.
We aim to work on ambitious research projects in a fast-paced manner.
If this sounds appealing to you, apply using the link below by Friday, August 1st EOD:
job-boards.greenhouse.io/dee…
Excited to share that thanks to @basvanopheusden’s help at @OpenAI , I’ve received some API credits to accelerate developing our agents. Grateful for the support
let’s see how Crux works at advanced models like o3, o3-pro🚀