Visual reasoning and RL at @perceptroninc. Former @dbrxmosaicai, @realitylabs I write music https://xitter.cf/t.co/QJl07cIBKe
Joined December 2024
- Tweets281
- Following3K
- Followers669
- Likes30.2K
Pinned Tweet
We taught a model to play Set! to explore which reasoning strategies make a model successful during RL. A base model's initial reasoning strategies determine the final outcome of RL and for visual problem solving, visually grounded reasoning is superior🧵
we can then use this to turn an SFT dataset with ground truths into an on-policy DPO dataset.
we find train samples where the model gets the answer correct, then add the steering vector.
turns out it only works about 75% of the time on the SFT data, but since we have ground truths we can filter the data to check whether the steering worked.
here for example we induce the model to call a tennis racket a tote bag
the next step in this experiment is to see how we can reliably induce the bad behavior on samples where the model would've answered correctly.
to do this we choose a cluster (i.e. hallucinating objects that aren't present) and take samples from one of the benchmarks where the model gets the answer correct. we then sweep every layer and determine which layer we need to add the behavior vector to induce hallucinations.
turns out this particular behavior can be reliably induced at layer 14
the next step in this experiment is to see how we can reliably induce the bad behavior on samples where the model would've answered correctly.
to do this we choose a cluster (i.e. hallucinating objects that aren't present) and take samples from one of the benchmarks where the model gets the answer correct. we then sweep every layer and determine which layer we need to add the behavior vector to induce hallucinations.
turns out this particular behavior can be reliably induced at layer 14
this original technique is used for identifying bad behaviors which come from DPO preference data - it can be extended quite trivially to analyze failures on eval benchmarks (if you have a ground truth demonstration that succeeds).
I've found it extremely useful at clustering failure modes, e.g. identifying only one instance in a segmentation problem, or outputting many degenerate boxes in a sort of doom loop. im excited to extend this to filtering SFT data.
this was inspired by an interesting convo with @SwannAiden
time to lowball boomers on fb marketplace 🫡
BREAKING: America’s $1.7 trillion auto loan market is headed for a crash, per Reuters.
Auto loans are now one of the biggest sources of household debt, bigger than credit cards.
And the strain is showing: 1 in 5 Americans is behind on at least one loan, the highest since 2010, and the average car on the road is now nearly 13 years old, a record.
this original technique is used for identifying bad behaviors which come from DPO preference data - it can be extended quite trivially to analyze failures on eval benchmarks (if you have a ground truth demonstration that succeeds).
I've found it extremely useful at clustering failure modes, e.g. identifying only one instance in a segmentation problem, or outputting many degenerate boxes in a sort of doom loop. im excited to extend this to filtering SFT data.
this was inspired by an interesting convo with @SwannAiden
the basic premise here is you take two possible outputs on a given input y, y'
and you calculate model_activ(y') - model_activ(y) to create a vector representing the gap between the two behaviors, then you can cluster them and label them to get interpretable results
their paper did a bunch of other cool stuff that I haven't tried yet, e.g. trying to take that vector and use it to steer the model, determining which layer is the optimal layer to intervene on in order to induce behavior, as well as using it to identify bad data in your train corpus that pushes the model towards the behavior
this is an exciting release from @ElorianAI. I've heard from friends who worked on Gemini ER that the model outputs visual patches during reasoning. Elorian, which was founded by a bunch of Gemini folks, seems to be doing the same. I haven't seen much publicly written about the internal workings though. I'd be very curious to learn how this behavior is elicited, what types of visual patches the model learns to output (i.e. is it just counterfactual reasoning, future state prediction, etc.), and under what types of problem this actually outperforms e.g. just outputting bounding boxes or plain text in your reasoning. congrats on the release!
We developed a model reaching frontier-level multimodal performance, matching or exceeding Astra on 5 of 12 benchmarks, all with 17x lower token prices. The model is smaller and cheaper than Gemini 3.8 Flash while dominating it in real world vision workloads.
We achieved this by inventing a technique for visual reasoning, where the model is no longer restricted to text chain-of-thought. Conditioned on visual tokens, Glimpse can perform recurrent computation over latent internal representations and can even orchestrate subagents that send messages over internal visual states.
Really proud of what our small team and only hundreds of GPUs could accomplish.
elorian.ai/blog/introducing-…
awesome release from our friends at @Physion_Labs
The AI industry has a massive compute efficiency problem. At Physion Labs, we think it's time to challenge it.
Today, we're sharing the technology behind NexTron, our world-intelligence training engine.
Our benchmarks on matched hardware:
⚡ Up to 20.9× SFT throughput vs. DeepSpeed (H100)
⚡ Up to 6.6× SFT throughput vs. Megatron (H200)
⚡ Up to 7.2× SFT throughput vs. Megatron (B200)
⚡ Up to 3.1× end-to-end RL throughput vs. Miles–Megatron
🔥 180B model fine-tuning on just 8 H100 GPUs.
Same hardware. Dramatically higher throughput. Comparable model quality.
We believe the next frontier isn't just building bigger GPU clusters.
It's unlocking more intelligence from every GPU, every watt, and every dollar spent on training.
#PhysionLabs #NexTron #WorldIntelligence #AIInfrastructure
Jeremy Dohmann retweeted
I’ve spent the last two weeks talking with founders, researchers, investors, and buyers across robotics.
This is the thesis that emerged from those conversations.
Where robotics data companies can win, and where I would focus next.
this is super cool - feels a bit like combining self-learned compaction with episodic memory. excited to try this out!
super impressed by how smoothly you've gotten the policy to run from cloud! great work @pentestduck
We just ran one of the largest robot models that exists today, all without a local GPU.
And it’s running real-time at <50 ms overhead.
Scaling laws are becoming real for physical AI. Perceptron’s Isaac 0.5 is one of the largest open-source models, sitting at 36B total parameters.
@perceptroninc joins other frontier robotics labs like @DynaRobotics and @GeneralistAI in showing the need to massively scale pre-training data, from a 1000s to millions of hours of video. Specifically, they show that the more egocentric video the model is trained on, the less teleop data that it needs to achieve the same loss.
At @DreamscaleLabs (YC F26), we’ve noticed an important truth here: data and training compute do not just scale in isolation. The model sizes must increase too, and GPT-6 Astra has given a glimpse into what a trillion-parameter robotics model could do.
We’re building to serve that future - one that is not constrained by the bottlenecks of on-board compute, but rather, allowing large models to run fast over the cloud.
Reach out if you’re interested in our work.
for my Seattle oomfs I'm hosting a concert in my living room Friday October 30, dm me for the invite 🐷
