←
Daniel Severo3,886 posts
Daniel Severo avatar

Daniel Severo

@_dsevero

Researcher at Meta Superintelligence Labs - FAIR / PhD @UofT @VectorInst / Previously @GoogleAI @UFSC 🇧🇷

◷ Joined March 2011▧ 189 Photos and videos
Daniel Severo avatar
Daniel Severo@_dsevero·4h

2027 AIM applications are open! Spend a year with us at FAIR Montreal during your PhD. Free food and hundreds of GPUs at your disposal! Topics I'm interested in: test-time planning, control, world models. Make sure to mention my name when applying :) metacareers.com/profile/job_…

Daniel Severo avatar
Daniel Severo@_dsevero·Nov 16

I agree. This type of doomerism is distracting away from dealing with the actual issues.

absolute fanfic. AI has real, immediate, tangible dangers that we need to focus on, for instance the impact on the job market and making sure random script kiddie can't use it for cyber offense. But this "oh once it's superintelligent it will take over and ensave us, maybe kill us all!" is pure sci-fi fanfic and distracts from the real (and difficult) issues, which linear probing won't solve. It's worth repeating the LLMs don't just "decide" to do something and then just "do it" in a void. They generate tokens based on our prompts, only as long as we call `model.forward()`, and on giant clusters with terabytes of weights, that's impossible to "quickly self-replicate". We also need to regularly remind that the OAI/HF incident was specifically a _guardrail-free_ cyber eval where the model is _instructed_ to hack, to find out how good it is at it. Some recent personal event made me realize there needs to be more push-back on this doomsday nonsense, and we need to re-focus the discussion on real dangers/issues. I won't argue much in the replies though, not interested in peer-reviewing Terminator fanfic.

Charles Lamanna avatar
⟳ Daniel Severo retweeted
Charles Lamanna@clamanna·Nov 13

I want to share what we know so far. This was not a cyberattack or an issue caused by AI tooling. Our amazing team was able to work through a very challenging situation to restore service. Our investigation points to technical challenges and failure modes (cascading failures + thundering herds at our network perimeter) that made recovery more difficult than normal. I'm working closely with the engineering team to complete the full post-mortem. We will publish the post-mortem shortly covering the step-by-step timeline and our many repair items to avoid similar outages in the future. We take our responsibility to our customers seriously and will do everything we can to learn from this and prevent it going forward.

We are deeply sorry for the disruption this incident has caused our customers and partners. We know the trust they place in our services and take that responsibility extremely seriously. We are seeing broad recovery and stability across our services. This was a technical issue not related to cyber or AI / agents. The incident was introduced by an issue during an earlier network capacity upgrade that surfaced following a recent network change. We have addressed the underlying issue and continue to restore service to remaining customers and regions. Given the scale of demand, we restored services gradually, validating stability at each stage. While this took longer, it helped ensure a stable restoration. We will provide a full post-incident report once our review is complete.

Chris Hoang avatar
⟳ Daniel Severo retweeted
Chris Hoang@choang333·Nov 13

Labs disagree on vision encoders for MLLMs! ⚔️ Gemma 4 12B scrapped the encoder. Kimi K3 dropped encoder pretraining. Inkling 2 Small revived the encoder Inkling killed. We trained 200+ models in a controlled study to settle it: ✅ Keep vision parameters, skip the pretraining. And they don't have to be an encoder! 🤔 Paper: arxiv.org/abs/2611.17334 🧵

Post attachment
Daniel Severo avatar
Daniel Severo@_dsevero·Nov 11

Wow this is really cool!

Finally, we can release a bit of the work we have been doing at the DexHands team in Amazon FAR ❤️‍🔥 Pure RL, zero teleop, PICK ANYTHING Come to Amazon Booth at CORL (Sound on.)

0:36
Julen Urain avatar
⟳ Daniel Severo retweeted
Julen Urain@robotgradient·Nov 8

Finally, we can release a bit of the work we have been doing at the DexHands team in Amazon FAR ❤️‍🔥 Pure RL, zero teleop, PICK ANYTHING Come to Amazon Booth at CORL (Sound on.)

0:36

Robots that learn should generalize. Locomotion should show generalization across terrain; manipulation across object categories. Here is a video of “Pick Anything”, demonstrating human like grasps for 66 different objects, ranging in size and shape from air pods to dumbbells to pipe wrenches, using a *single* policy. The system uses the stereo camera head for vision (no wrist cameras) & Sharpa Wave hands for manipulation. It was trained using video and RL (zero teleoperation data). This is research from our Dex-Hands team at Amazon FAR and we will be doing live demos of the system at CoRL 2026 ( Demo 12, Nov. 9 &10). Please come by!

2:42
Daniel Severo avatar
Daniel Severo@_dsevero·Nov 6

Life update: a few months ago we made a Canadian. Today I became a Canadian myself. Proud to join this great nation that has treated us so well. Here’s to the future!

Post attachment
Brandon Amos avatar
⟳ Daniel Severo retweeted
Brandon Amos@brandondamos·Nov 6

We've posted a 156-page tech report on Beam! We have good systems and it's worth reading. It covers all phases through pre-training, mid-training, SFT, RL, and OPD and goes into the infra, data, and scientific decisions we made throughout the process 📚 reflection.ai/documents/beam…

Post attachmentPost attachmentPost attachmentPost attachment

We've released the weights for Beam!! 🤗 huggingface.co/reflection 📄 Model card: reflection.ai/beam-model-car… 🧪 Try Beam in our playground & get an API key: platform.reflection.ai/playg… 📚 AND we take open model safety seriously and have posted our safety report: reflection.ai/beam-safety-re…

Cohere avatar
⟳ Daniel Severo retweeted
Cohere@cohere·Oct 30

We’re expanding our commitment to Canada 🇨🇦 Cohere is advancing plans to invest more than C$4 billion in R&D and AI infrastructure and create over 1,500 jobs through 2031.

Artem Zholus avatar
⟳ Daniel Severo retweeted
Artem Zholus@artemZholus·Oct 30

You can now build a robotic scaling law with a cheap robot on 1 GPU. The recipe: 1. Collect 30mins of demonstrations with SO-101 2. Run inference with public RoboJEPA models (a few hours on 1 GPU). 3. Fit the scaling law. details: robojepa.github.io

Post attachment

Introducing RoboJEPA, the first scaling law for multiembodiment world models. We scaled JEPA predictors up to 8B params, showed scaling laws extrapolate and correlated compute scaling with planning performance. We release all checkpoints and the code. robojepa.github.io

1:06
Mark Carney avatar
⟳ Daniel Severo retweeted
Mark Carney@MarkJCarney·Oct 29

Canada’s sovereignty depends on our ability to build, adopt, and govern AI on our own terms. That’s why Canada is taking an equity stake in Cohere — one of the world’s leading builders of secure AI. Sovereign AI. Built here at home. Read more: pm.gc.ca/en/news/news-releas…

Post attachment
Mark Carney avatar
⟳ Daniel Severo retweeted
Mark Carney@MarkJCarney·Oct 29

Bonjour, Montréal.

0:07
Mark Carney avatar
⟳ Daniel Severo retweeted
Mark Carney@MarkJCarney·Oct 26

Congratulations, President @LulaOficial, on your re-election in Brazil. The relationship between Canada and Brazil has flourished under our partnership. Let’s seize this opportunity to make it even stronger, together — to create major new trade opportunities, protect the environment, and build greater prosperity for both our peoples.

Mark Zuckerberg avatar
⟳ Daniel Severo retweeted
Mark Zuckerberg@finkd·Oct 15

Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple.

Post attachment
Artem Zholus avatar
⟳ Daniel Severo retweeted
Artem Zholus@artemZholus·Oct 10

Introducing RoboJEPA, the first scaling law for multiembodiment world models. We scaled JEPA predictors up to 8B params, showed scaling laws extrapolate and correlated compute scaling with planning performance. We release all checkpoints and the code. robojepa.github.io

1:06