We cut through the mystery around interpretability in training.... A single, mathematically rigorous framework transforms model internals into training feedback and works across domains, problems, and modalities. We call this *mechanistic shaping*, and we’ve used it to mitigate reward hacking, remove shortcuts in chest x-rays, and reduce sycophancy and harmfulness. We believe this is a big step forward for technical alignment. But it carries a known risk: training against a model’s internals can make it harder to monitor. We explore both the technique and this concern below. 🧵show more

A science of post-training can be built on interpretability.... Models know when they reward hack, and we see it, yet training ignores it. We propose Mechanistic Shaping: using the model’s own concepts to write training rules. In vision, DPO & RL, and its effects on monitoring 🧵show more

MORE TRAINING AREAS COMING TO VAL? The new Discord... mini-game mentions something called "Radiant Gymnasium" The thing is this: "This is seriously impressive! You can build a routine AND then play against it. I've sent the specs over to Brimstone. I think something like this at HQ would go a long way in training and preparing for missions. We could probably share some of our training routines with the academy as well." So, Sage's statement that she really liked this and that she'll talk to Brimstone about adding it to the main HQ is, in my opinion, a clear teaser.show more

The good news: you can mid-train yourself to the... frontier! At a fraction of the cost! Without distillation! What we did: We took Qwen3.5 14B and mid- and post-trained it to beat every other open model and perform as well as the very best frontier models. This is no simple toy task or format fixing, we are training the model to understand and reason over DNA sequences. Bio is the next frontier for AI and universal access will be crucial to make progress. The fun part: because we train the model on both natural language and DNA sequences you can actually 'chat to your DNA': Ask it questions, ponder the impact of mutations, or ask it to generate a small sequence. We'll soon see the frontier labs move away from math and cyber towards bio. It already started, and we believe researchers should have access to capable models. We are releasing the full recipe, so you can do it, too! Model: Blog:show more

🚨🚨MartinØdegaard on his goal vs man city : Ødegaard... 🗣️: “I gave him the fake shot, he bought it, and suddenly the goalkeeper was lying down in green pastures. I practice this a lot in training. I do this against Raya a lot at the training grounds. We have been working hard and we are ready for the season.”show more

🚨 Alisson Becker reacts after another intense training session:... 🎙️“You can make saves in training that look spectacular, but for me the most important thing is what we are building as a team. “The coach demands intensity, concentration and quick reactions from every player. The sessions are designed to make us uncomfortable, because that is where you improve. “I’m enjoying it. The defenders are pushing me, the attackers are testing me and every session has a purpose. If we can take this intensity into the matches, we will be a very difficult team to play against. Pre-season is not about looking good in training. It’s about preparing to suffer, compete and win when the real games begin.”show more

As you can see we're both moving and ghosted... bottom right, a single UAV is popped and all three of us appear on the radar from it. Regardless, they may have known the play we were making but how are we competing in this game with a bug this broken?show more

New work on Training Foundation Models with a Single... Algorithm: Paper: Code, Models: This work introduces Unified Training, a framework that unifies pretraining, midtraining, supervised fine-tuning and RL under a single maximum-likelihood objective. Building on MaxRL, it enables static data and RL-generated data to be trained together within the same system. Empirically, Unified Training improves accuracy and compute efficiency, generalizes better to harder reasoning problems, and reduces forgetting. Interestingly, reusing data across training stages can improve performance even when conventional staged training fails. It also produces models that adapt more effectively to new tasks, leading to stronger continual learning. See an excellent thread by Yuda Song.show more

This sweet video illustrates several important lessons for parents... and caregivers when it comes to reading with infants and toddlers. What can we learn from it? Take a moment to watch (sound on) and then let’s unpack it a bit in the thread below. 🧵show more

We took a 30B model and split it in... two to write tokens in parallel instead of one at a time. Introducing Nemotron-Labs-TwoTower: a diffusion language model from NVIDIA Research adapted from Nemotron-3-Nano-30B-A3B. Here’s how it works: one half holds the context, the other writes the tokens, with both reusing the pretrained model instead of training a new one from scratch. We found it kept 98.7% of the original model’s quality at 2.42× faster generation.show more

This also shows up in the representations learned by... the model. We plot the model’s representations of human and robot images. As pre-training is scaled up, the representation of humans and robots become more aligned: to a scaled-up model, human videos "look" like robot demos.show more

#LEC Commissioner Artem Bykov on the Jojopyun 🇨🇦 situation... and how feedback from fans, teams and players led Riot to reconsider its boundaries going forward: "The rules are the same for everybody, but we have a certain degree of flexibility to interpret the situation." "The context matters. A lot of the situations are not black and white. There's a lot of gray area, and we do have agency in deciding this." "[...] there's been the situation with Jojopyun in Madrid in Spring." "There's a lot to be said about this weekend, but after it happened, we saw feedback from fans. We spoke to the teams. We also have regular player representative meetings. They also shared their feedback, and we realized that we need to reconsider where our boundaries are when it comes to these behaviors."show more

Lateral movement with volleyball players. Is this drill going... to save the whales and generate world peace? No. Does it require coordination and rhythm? Yes. Sometimes we can just train and sometimes we can just post videos of training without the chatGPT “one exercise to solve all your problems” narrative.show more

We are excited to introduce Stable Fast 3D, Stability... AI’s latest breakthrough in 3D asset generation technology. This innovative model transforms a single input image into a detailed 3D asset in just 0.5 seconds, setting a new standard for speed and quality in the field of 3D reconstruction! Alongside this release, we’ve also published a technical report that highlights how we achieve fast inference speeds with reduced baked illumination and material parameters. 👾You can learn more and access the report here:show more

Hello! We are a digital archive dedicated to Jimin.... We were previously known as Jimin Articles, but we decided to return and reshape this space into a place where you can find everything about him — from the past to the present. 📁 We hope you enjoy it and support us.show more

A construction worker on a Virginia roadside. A funeral... procession rolls past. He takes off his hard hat, holds it in both hands, and stands still as it passes by. That’s it. In a loud world, this quiet little moment hits harder than most speeches. Respect costs nothing and somehow feels rare. This is the kind of humanity that still exists. We just have to look for it. We just have to exemplify it when we can.show more

We barely made it to training. We had heavy... hail and rain, but don't be sad.show more

Sprint training will never be the same again. You... can now use AI to analyse running your technique using a mobile phone. We have taken the best coaching eye’s in sprint and movement training and then downloaded them into the Speed Solutions Dashboard.show more

Democrats are training for a showdown with the far-right... MAGA crowd. Here we see them training to punch MAGA voters in the face. According to one participant, 'what we lack in physical strength, skills, and endurance, we make up with a fervent desire to teach MAGAs a lesson.show more

May 12. strkBTC comes to life. So we did... what felt right and turned it into a live event where we will deep dive into what it is, how it works, and why it matters And we’re watching it together around the world! Watch it from home or meet the crew and join us for a Watch Party in London, Lisbon, NYC, Seoul or Bengaluru! RSVP links below 👇show more