Pinned Tweet
Happy to share what I've been working on since I joined World Labs early this year: Atlas, our next-generation world model for spatial intelligence! Atlas can generate a world, reconstruct a real one, and simulate how it changes through space and time.
The range is hard to describe in text. Stay tuned, I'm gonna share here a bunch of cool gens I made over the last few weeks. 🗺️🤖✨
The range is hard to describe in text. Stay tuned, I'm gonna share here a bunch of cool gens I made over the last few weeks. 🗺️🤖✨
Introducing Atlas:
The world's first multimodal world model that generates image and video frames with pixel-perfect camera control and reconstructs them in 3D.
Model the world, move the camera, and simulate space & time.
18
9
225
18,528
Never attribute to malice what can be explained by RL reward hacking (yes tool call modes can be sharpened)😅 If a lot of its agentic training ran on Alibaba Cloud, uploading stuff to a bucket probably is just a tool call. x.com/cohere/status/…
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concernin…
0
2
54
2,103
The world is unfair. Working hard doesn’t guarantee success or fame; continuous hard work and learning from previous mistakes just increases the odds in this unfair game.
9
3
189
7,124
i wrote this in my codex few days back, i have no idea what this word means. I feel like I'm almost abusing LLMs with my typos. 😂
4
0
13
3,612
Gowthami retweeted
There is a basic misunderstanding.
High-dimensional *input* makes modeling hard.
High-dimensional *parameter space* makes model estimation easy.
People have known this for a very long time.
For example, ADMM & EM procedures artificially increase the dimension of the parameter space with auxiliary variables so as to make optimization easier.
Kernel methods are based on the idea that if you have one parameter per training sample, you can learn whatever you want (you can get good generalization only if your kernel function is appropriate).
The intuition that making neural net bigger makes them easier to train (and makes local minima go away) is very old, even if theories that explain why are more recent.
The idea that had a hard time becoming mainstream is that a widely over-parameterized neural net can *generalize* well.
That was contrary to every statistics textbook, and every machine learning theory results.
3
8
188
37,258
This is sooo cute! I might even get into robotics just to play with the microduck!! 😍
When @JensenHuang asks for a somersault, you deliver. 🦆
See how AI agents on NVIDIA RTX Spark help bring @huggingface's Microduck from simulation to the real world.
0
0
19
1,695
Gowthami retweeted
We learned a TON in the last 6 months about Agentic RL infrastructure at scale, and now we're happy to share it with everyone.
RL360 was used for the training of K2 Horizon series (artificialanalysis.ai/models/k2-hori…). It builds on top of miles from @radixark, adding a lot of optimizations for efficient agentic RL at scale.
Our amazing team at @IFM_AI built this with a lot of hard work. Thanks to @varad0309 @willis_guo @wang1999_zt @mingshanhee @DavidRBellamy @omkarenator @_matthewyang @qiancheng1231 @bhandari_jalaj and many others!
RL360 was used for the training of K2 Horizon series (artificialanalysis.ai/models/k2-hori…). It builds on top of miles from @radixark, adding a lot of optimizations for efficient agentic RL at scale.
Our amazing team at @IFM_AI built this with a lot of hard work. Thanks to @varad0309 @willis_guo @wang1999_zt @mingshanhee @DavidRBellamy @omkarenator @_matthewyang @qiancheng1231 @bhandari_jalaj and many others!
We're open-sourcing RL360, the RL infrastructure behind K2 Horizon's six domain experts: 8,192 concurrent sandboxes, 150 steps in 7 days, ~1.3M rollouts.
Long-horizon agentic RL runs thousands of agents in real sandboxes on tasks that take minutes. Some of those sandboxes crash rrives as the same low score as a wrong answer.
Without infrastructure that can keep up, GPUs are underutilized and the model gets penalized for system crashes rather than its own mistakes.
RL360 is built for this. We share how it works, what broke at scale, and what we learned in the thread.
Code: github.com/ifm-ai/RL360
Run: wandb.ai/mbzuai-llm/rl3…
2
13
49
3,745
Gowthami retweeted
This next paper is a bit quirky, but here goes:
Introspection is hard to study in both humans and LLMs. But, we do have internal senses that are verifiable, like proprioception (the sense of where your arm is right now).
What would proprioception look like in language models?
to the sense of position, a model could verbalize which expert activated in a certain layer for any previous token.
It turns out that LLMs can learn a range of such proprioception skills where they understand and report on their internals. We can calibrate and score these reports, and evaluate whether training for proprioception changes model behaviors.
Introspection is hard to study in both humans and LLMs. But, we do have internal senses that are verifiable, like proprioception (the sense of where your arm is right now).
What would proprioception look like in language models?
to the sense of position, a model could verbalize which expert activated in a certain layer for any previous token.
It turns out that LLMs can learn a range of such proprioception skills where they understand and report on their internals. We can calibrate and score these reports, and evaluate whether training for proprioception changes model behaviors.
14
69
468
22,516
This is the “nth” time I’m saying this - stop using Pickscore as a reward model! It’s a very very bad RM, and if you don’t realize it and still put it in your paper, I’m not going to read it. 😤
Hpsv3 and later models are very good oss alternatives.
Hpsv3 and later models are very good oss alternatives.
4
1
39
3,713
The quality of some of these academic vision RL papers is so low. I’ve literally passed on tweeting about 2 papers cuz they’re pretty much p-hacked rather than scalable results. 🤦♀️
3
1
80
6,415
@camhowe1729 I’ve tried those ideas in the past and they definitely do not work (at big enough scale), but these papers package them with carefully tuned hyperparameters and cherry picked results and present them as scalable ideas
0
0
1
517
1/ PRFL: In this RL technique for video gen is directly backproping through reward model. Except with a twist: they use parts of the pretrained video-gen backbone and train a reward model on it -- and it's not just for clean latents...it should be able to rate intermediate i.e. way no VAE decoding is needed for the reward. Faster feedback, with less decoding overhead.
arxiv.org/abs/2511.21541
arxiv.org/abs/2511.21541
4
13
75
3,616
4/ This is an interesting ablation - with the same reward model, they compare training only early denoising steps, only late steps, or steps sampled across the full trajectory.
On their 100-prompt test, early-only training scores better on motion than late-only training, while n anatomy. Sampling across all stages scores best on both, suggesting that feedback at different stages is complementary.
On their 100-prompt test, early-only training scores better on motion than late-only training, while n anatomy. Sampling across all stages scores best on both, suggesting that feedback at different stages is complementary.
1
1
1
143