In flow matching, a coupling determines how noise and... data samples are paired during training. The choice of coupling is important because it influences the geometry of trajectories at inference time. The simplest choice is the independent coupling, where noise and data points are paired arbitrarily. This can lead to curved trajectories as the model averages over many conflicting pairings. However, if we use optimal transport on batches of pairs, this leads to fewer ambiguous intersections that the model must resolve, leading to straighter trajectories at inference time.show more

Flow matching produces smooth, deterministic trajectories. In contrast, the... sampling process of a diffusion model is chaotic, resembling the random motion of gas particles.show more

Lorenz attractor is the set of limit trajectories for... a simplified mathematical model for atmospheric convection. It has a fractal geometry.show more

Flow-based generative models trained with flow matching tend to... learn curved trajectories, which are challenging to approximate in a few steps. Rectified flows aim to learn straight trajectories, which are easier to simulate with less computation.show more

LLMs are like Schrödinger’s cat: many possible trajectories, but... you only see one outcome per run. To really understand models, and debug where they go wrong, you can find the "forking tokens" that lead to different trajectories. Our new research does this 100x more efficiently!show more

System identification (sysid) is the process of finding the... physical parameters that make a simulation match reality. If you're training an RL locomotion policy in simulation, the accuracy of your motor model directly affects how well the policy transfers to the real robot. A recent git commit by Kevin Zakka added a sysid toolbox to MuJoCo which automates this process: you provide recorded motor data and a MuJoCo model, and it optimizes the model parameters to minimize the difference between simulated and real trajectories. For my RobStride Dynamics RS02 QDD motors (17 Nm peak, 7.75:1 gear), I built a Rust tool that sends multi-sine torque excitation at 1 kHz and records position/velocity feedback. I then feed this data into MuJoCo's sysid optimizer.show more

Coarse2Real (C2R) transfers simple 3D renderings into realistic style... video. Check our paper and project page to learn how to hedge small amount of synthetic paired data with real non-pair data for training the C2R model. We will release the model soon!show more

Holy sh!t ! OpenAI will have their custom inference... chips ready in just a few months and deployed at scale by the end of the year! 🤯 Training chip = The heavy lifters that require massive amounts of data and power to build and teach the AI models from scratch. Inference chip = The specialized, highly efficient chips that actually run the AI and generate the answers in real-time when you use it. This is going to help OpenAI drastically cut down their massive compute costs, speed up model reasoning times, and finally break free from relying entirely on Nvidia to scale their operations.show more

Transformer and Mixture of Experts, explained visually! Mixture of... Experts (MoE) is a popular architecture that uses different experts to improve Transformer models. Transformer and MoE differ in the decoder block: - Transformer uses a feed-forward network. - MoE uses experts, which are feed-forward networks but smaller compared to those Transformer. During inference, a subset of experts are selected. This makes inference faster in MoE. Also, since the network has multiple decoder layers: - The text passes through different experts across layers. - The chosen experts also differ between tokens. But how does the model decide which experts should be ideal? The router does that. It is a multi-class classifier that produces softmax scores over experts to select the top K experts. The router is trained with the network, and it learns to select the best experts. But it isn't straightforward. There are challenges! Challenge 1) Notice this pattern at the start of training: - Say, the model selects "Expert 2" - This expert gets a bit better - It may get selected again since it's the "best" - It learns more - It gets selected again in the next iteration - It learns more, and so on! This means many experts can go under-trained due to the overselection of a few experts! We solve this in two steps: - Add noise to the feed-forward output of the router so that other experts can get higher logits. - Set all but the top K logits to -infinity. After softmax, these scores become zero. This way, other experts also get the opportunity to train. Challenge 2) Some experts may get exposed to more tokens than others, leading to under-trained experts. We prevent this by limiting the number of tokens an expert can process. If an expert reaches the limit, the token is passed to the next best expert. Overall, MoEs have more parameters to load. But a fraction of them are activated during inference. This leads to faster inference. Mixtral 8x7B and Llama 4 are two popular MoE-based LLMs. Have you used MoEs in production yet? To dive deeper into how MoE inference works in production, we wrote a full article covering token dispatch, grouped expert computation, model-weight memory, multi-GPU communication, expert placement, load imbalance, and performance diagnosis. Read it below.show more

This animation shows how a uniform grid of points... is deformed by a flow matching model. The visualization highlights how the learned flow warps the underlying space—stretching, compressing, and bending it—to transform one probability distribution into another.show more

We’re proud to share that Tamarind Bio has been... selected to build, host, and operate the inference infrastructure layer for TuneLab2.0, the next evolution of the platform. Eli Lilly and Company TuneLab is a first-of-its-kind, collaborative AI/ML drug discovery platform, bringing models trained on over $1B worth of Lilly proprietary data to the biotech ecosystem. Tamarind will power TuneLab’s scalable drug discovery workflows and model inference.show more

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. Perceptron AI joins other frontier robotics labs like Dyna Robotics and Generalist 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 Dreamscale Labs (YC F26) (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.show more

The next World Cup star is playing right now,... and nobody is watching. The 2026 World Cup is on, and millions of kids are dreaming of one day playing for their own flag. But the path there is almost impossible: no consolidated data, too many games, not enough scouts, human bias, and wrong place, wrong time. Any one of them can bury a generational talent before a scout ever sees them. That is what we are building to fix. So we developed a full automated annotation + action spotting model where it actually matters: grassroots football. Shaky cameras, awful angles, terrible lighting. But now nothing on the market beats sn44's v2.0 model on it. Check the last GT line, it’s human annotations. Simply because miners' outputs + post-process = magic. Post-process is a deliberate design choice. We point miners at the hardest 80% of the workflow, the part that is genuinely difficult to fake, and handle the rest downstream. Fewer places to game the system means a network you can trust. Every single challenge on sn44 can be an independent startup now. Let that sink in.show more

Amazon’s machine learning model collects 300 million data points... per season and can now predict which players are likely to blitz before the snap. This is a look into the future of broadcasting.show more

The tokenization market is expected to reach $30T within... the next ten years. And we are at the perfect spot to take the maximum advantage of it! DevvExchange assures full regulatory compliance - so we are the optimal choice to onboard all governments and institutions to web3.show more

What if you kept asking an LLM to "make... it better"? In some recent work at FAIR, we investigate how we can efficiently use RL to fine-tune LLMs to iteratively self-improve on their previous solutions at inference-time. Training for iterated self-improvement can be costly. The naive approach to training for K self-improvement steps leads to K times the number of rollout steps per episode. We introduce Exploratory Iteration (ExIt), an RL-based automatic curriculum method that bootstraps diverse training distributions of self-improvement tasks by upcycling the LLM's own responses at previous turns as the starting points for both self-improvement and *self-divergence.* In order to decide what task to train on next, the curriculum prioritizes sampling of partial turn histories that led to higher return variance in its GRPO group (a learnability score that comes for free). This automatic curriculum over the bootstrapped task space teaches the model how to perform iterated self-improvement while only ever training the model on single-step self-improvement tasks. We look at ExIt's impact in both single-turn (contest math problems) and multi-turn (BFCLv3 multi-turn tasks), as well as MLE-bench, where the LLM is run in a search scaffold to produce solutions to real Kaggle competitions. Across these eval settings, we find ExIt produces models with greater capacity for inference-time self-improvement compared to GRPO. Notably, ExIt models can self-improve on test tasks for many more steps than the typical solution depth encountered during training, including a 22% improvement in MLE-bench performance compared to GRPO.show more

A curious property of diffusion models: when a sample... is mostly noise, the model’s best guess is close to the mean. So early on samples can be pulled toward the center of the dataset, even when no data exists there, before rebounding toward the data manifold.show more

OpenAI's Deep Research is getting a run for its... money. Deep Lake was just released, and it's a different take on an AI system that can do deep research on your own data. You can use Deep Lake to build AI search with reasoning on your private and public data. (Look at the attached videos to get an idea of how it works.) If you want to research proprietary and sensitive data, Deep Research won't help you because it's limited to public data. Deep Lake, however, will allow you to use your private data. On top of that, Deep Lake supports multi-modal retrieval from the ground up. It uses vision language models for data ingestion and retrieval so that you can connect any data (PDFs, images, videos, structured data, etc.) You can even use mixed-data queries! Deep Lake can search your data from S3, Dropbox, and GCP. It learns from your queries over time, making the results as relevant to your work as possible!show more

🚀New paper out - We present Video-MSG (Multimodal Sketch... Guidance), a novel planning-based training-free guidance method for T2V models, improving control of spatial layout and object trajectories. 🔧 Key idea: • Generate a Video Sketch — a spatio-temporal plan with background, foreground, and motion in the pixel space. • Encode this structure directly into the latent space of the diffusion model during generation, which does not require fine-tuning or additional memory during inference. 🧵show more

BIG NEWS. The Blockworks website has evolved: yesterday we... were the home of news, today we are the home of onchain data. Head over to the site to see for yourself, but here's a little snippet of what you can expect: 1. Sector leaderboards (chains, DEXes, borrow lend, DATs, etc...) 2. Comprehensive data dashboards protocols 3. The ability to compare pricing and onchain data easily (coming soon) We're doing this because the industry still has a gigantic data problem. As investors get more sophisticated and fundamentals driven, basic high level facts are no longer sufficient. Investors need to be able to trust the data they are seeing and go much deeper than the surface level info that's available today. Additionally, because many data providers allow companies to essentially self report, you can't trust what you are seeing. This site is our contribution to fixing that problem and to ensuring clear, accurate data for investors. Blockworks is fully dedicated to becoming the most comprehensive data company in crypto in 2026. This is the first of many, many announcements like this this year, stay tuned.show more