Evaluations are the secret sauce of frontier labs and Anates is launching its first project in a collaboration with Deepmind. If you work on video models and or embodied intelligence check out their work! Congrats Tim and Carsten!
I am moving to @ICComputing at @imperialcollege as an associate professor, where I will be expanding my lab!
I am looking for PhDs and postdocs to join me on my quest to build foundation models with adaptive tokenisation and memory (AToM FMs, funded by @ERC_Research)
We are looking forward to welcome @brianltrippe for this months Chalmers AI4Science seminar tomorrow. Check https://t.co/2TbVpiAixo for details on how to connect!
You really think we're going to scale data labelers to AGI?
Today, we release the largest public long-horizon dataset of human digital work.
600h of long-horizon AGI research across 3 months. ๐งต(1/n)
The focus on UMAPs as the output of scRNAseq has debased a whole field.
This fig from https://t.co/IV62ulylhF is ridiculous.. showing a faster GPU accelerated workflow.. that produces a completely different UMAP. But since nobody cares what's on the UMAP anyway.. who cares?
ACM has posted an "Expression of Concern" on Igor Markov's article. This stinks to heaven. Apparently, Google did not like the paper. Google is a major donor to ACM. There are also rumors of a legal action by Google against ACM.
https://t.co/txWSoOCpZv
Excited to launch Principia, a nonprofit research organisation at the intersection of deep learning theory and AI safety.
Our goal is to develop theory for modern machine learning that can help us understand network behaviors, including those critical for AI safety.
1
everyone talking about macrohard
seems like a good time to remind people that we openly released the biggest long-horizon behaviour-cloning dataset of screencasts (of AGI research)
350+ hours and counting!
https://t.co/9e8THEEWy6
Training on code is one path to AGI. Yet, training on github repos only implies training on end results rather than training on the way to get there. Crowd-code is an effort to solve that! Install once and contribute to open source efforts for AGI.
Today we release crowd-code 2.0, our second phase of crowd-sourcing the largest long-horizon open software engineering dataset.
Install once. Forget about it.
๐จ Looking for a fully self-contained intro to diffusion models that covers both continuous (images) and discrete (text, sequences) data?
๐ We just released:
โFoundations of Diffusion Models in General State Spaces: A Self-Contained Introductionโ
arxiv : https://t.co/O6ONhV04dM
S/o to @andrea_dittadi for his amazing support & guidance, and huge thanks to @TobiasHppe1, @k_neklyudov, @AlexanderTong7 and @stefanAbauer for their supervision! ๐
One roadmap for all of diffusion. ๐๏ธ๐จ
After a few failed posts, broken previews, and getting briefly flagged by Xโฆ the full thread's finally out ๐คฏ๐งต๐
Excited to present our work on multi-objective scientific discovery at #NeurIPS2025! ๐
We present Preference-Guided Diffusion, a novel method to sample diverse designs from a diffusion model that corresponds to the Pareto Front of black-box objectives.
Most (if not all) problems in scientific discovery, from drug design to material design, are multi-objective. Balancing factors like toxicity and activity or stability and synthesizability requires finding the optimal Pareto Front of trade-offs. Preference-Guided diffusion uses preference pairs from just an offline dataset to sample optimal designs from a diffusion model.
W/ amazing co-authors @syrineblk, @StefanoErmon, @stefanAbauer, and @BeEngelhardt.
Find out more:
๐๏ธ Poster #915 | Wednesday | 4:30 PM PST
Pdf: https://t.co/EY1YBoc9hg
Project: https://t.co/qQg3BaJ5xg
Discrete diffusion is all the hype and you want to get started? Have a look at this discrete diffusion codebase. Focused on letting you try out new ideas quickly with baselines already available. Feedback welcome!
Weโre releasing UNI-Dยฒ, a unified codebase for discrete diffusion language models ๐ค๐
Co-led with @vincentpaulinef and an amazing advisor team: @stefanAbauer, @AlexanderTong7 , @andrea_dittadi, @AMK6610, @KaplFer ๐
๐ GitHub: https://t.co/JJA7EVGVed
๐ Docs: https://t.co/GRJnEOULf6
Reproduce and extend state-of-the-art baselines with one toolkit. Letโs move beyond autoregressive models and push discrete diffusion together ๐งต๐
It was great to see @thinkymachines LoRA w/o Regret blog, which connects nicely to our work on Tina (LoRA for RL).
For wider use, weโre releasing a clean implementation of RL with LoRA, DoRA, QLoRA/QDoRA, plus speedups & more, across models from 1.5Bโ32B.
Nice work @UpupWang!
My @Cohere_Labs talk is online.
We outline research directions that embrace the bitter lesson, and state roadblocks on the path to AGI that need to be addressed even in a regime of absolute energy- and compute-abundance.
https://t.co/GraD3bt6rK
๐ฃ Applications are open for the CIFAR Global Scholars program!
This program offers junior faculty the opportunity to pursue interdisciplinary research, expand their networks and gain valuable leadership training.
๐ https://t.co/FJ8AzrRa9e