🚀 Sol-Attn is here! We present a training-free sparse... attention method that accelerates video generation while better preserving quality. Sol-Attn unifies dynamic routing, sparse computation, and approximate correction in a single online-softmax pass: • On-the-fly block thresholding for dynamic yet controllable budgets • Proxy-score reuse to approximate unselected blocks Results (vs dense FlashAttention-3): • Wan 2.1-14B: 2.02× end-to-end • HunyuanVideo-13B: 2.12× end-to-end • LTX 2.3: up to 2.4× end-to-end When integrated into Sol-Engine (with kernel fusion + caching): • Wan 2.1-14B: 3.48× end-to-end • HunyuanVideo-13B: 5.08× end-to-end Already available in Sol-Engine. The B200 kernel is still under further optimization. 🎬 Project: 📄 Paper: 🔗 Code:show more

NVIDIA released Sol-Attn A training-free sparse attention method accelerating... video generation inference by up to 2.1× while preserving visual quality, unifying dynamic routing and approximate correction in a single online-softmax pass.show more

🚀 Sol Video Inference Engine is here! An agent-native,... training-free full-stack accelerator for video diffusion. It auto-tunes cache + sparse attn + token pruning + quant + kernel fusion for any model/hardware/config. >2× end-to-end speedup on 64B Cosmos3-Super, 22B LTX-2.3 and 2B SANA-Video — near-lossless VBench quality, minimal human effort. Practical acceleration for real video gen deployment. 📄 Paper: 🌐 Project: 💻 Code: Proud of the team! 🎉show more

Disappointed with your ICLR paper being rejected? Ten years... ago today, Sergey and I finished training some of the first end-to-end neutral nets for robot control 🤖 We submitted the paper to RSS on January 23, 2015. It was rejected for being "incremental" and "unlikely to have much impact" Our resubmission to NeurIPS was also rejected It now has >4,000 citations (and more importantly, end-to-end training is widely accepted!) It's also cool to think about what's changed and what's the same -- - The network was 92k parameters and trained on ~15 minutes of data - The code was a combination of matlab, caffe, ROS, a custom CUDA kernel for speed, and a low-level 20 Hz controller in C++, all talking to each other. ROS+matlab was as bad as it sounds. - We pre-trained the encoder and did inference off-board on a workstation with a larger GPU. - We were paranoid about varying lighting messing up the network, so we did all the experiments after sunset (so long nights running experiments on the robot past 3 am) Now, we have manipulation policies that are far more dextrous, far more generalizable, and maybe on the cusp of breaking into the real world. :) (the paper:show more

☀️Very proud of the team for training 5.6! A... few of my highlights: - Great front-end aesthetics. E.g., I asked Sol to make a new blog for 5.6 with a celestial theme. - Much better CUA - Can work for much longer and is less lazy But it's not all great: - The writing quality has improved, but it's still bad in artifacts - Too many options with models/reasoning/fast mode/Cerebras/multi-agent We'll fix the above. What else should we improve? PS: belated post as there are so many other good models to train.show more

Comeback Journal Had to borrow SOL and got rekt... all the way down to 0.7 SOL. Took a 0.5 SOL trade and scalped +0.2 SOL. Saturday, Aug. 2 – 12:59 AM Portfolio: 0.9 SOL By the end of Aug. 2... Back to 0 SOL. Aug. 3 0xW sent me 0.6 SOL because he made money from the NA shill. Thank you again. Started only risking 0.05–0.1 SOL per trade to protect capital. Built it back to 0.4 SOL. Then p sent me 1 SOL for the same reason. That genuinely helped me during survival mode. Portfolio: 1.4 SOL. Saw LMEOW at 40k. Put 0.5 SOL in, leaving me with 0.9 SOL on the sidelines. Yeah, I oversized, but I had that feeling. What made me buy wasn't the chart. The narrative made me extremely bullish. Volume kept increasing, and I kept seeing more smart wallets buying than selling. Sold 1.5 SOL around 255k. Then the timeline started pushing the "OG" coin again. Momentum died. Round-tripped a few SOL and finally sold my remaining 0.7 SOL around 114k. Turned that 0.5 SOL trade into roughly 2.2 SOL before closing the position. I still think splitting the attention between two coins as a marketing stunt hurts everyone involved. Btw LJC, a lot of traders look up to you. Encouraging more PvP because the "OG" needs marketing instead of letting the market naturally bag work only makes the trenches worse. If it wasn't for that in the back of my mind i could've been at 10 solana today, thanks. Aug. 3 – 12:04 PM Current portfolio: 3.1 SOL.show more

"There are Jews in the diamond and gold trades... and they wanted in on what Rhodes was cooking, crazy." Helping Jews, as Rhodes did, is tantamount to helping Jews achieve their ends. In this case, Rhodes helped turn London into a refuge for Jewish terrorists, albeit indirectly. Cecil Rhodes' business partner, investor, and close friend was Nathaniel Rothschild - one of the most destructive Jews to England. Rothschild was not only the sole executor of Rhodes' will after Rhodes died, he was responsible for setting up the famous Rhodes Scholarship. Nathaniel Rothschild was pivotal in turning part of London's East End into a Jewish shtetl in the 1880s, when terrorist Jews fled Russia after the Tsar was murdered. Almost the entire Jewish population of England lived in London's East End during this period, and Rothschild supplied these Jews with welfare, free education, free healthcare, and affordable housing through his Four Percent Industrial Dwellings Company (which still operates today, and is still a Jew-first organization). So many terrorist Jews flooded London's East End (including Jack the Ripper) that parliament had to pass the Alien Act of 1905 to curb it, despite Nathaniel Rothschild's opposition. However, Rothschild was able to secure the "asylum clause," creating a window for more terrorist Jews to enter London. Thirty years later, Oswald Mosley would march through the Jewish enclave of London's East End and clash with these same Jews in what's now known as the Battle of Cable Street. Cecil Rhodes, through ignorance or indifference, was indirectly helping Nathaniel Rothschild turn England into a refuge for Jewish terrorists, whose descendants would turn Britain into a multiracial state, including the one and only Barbara Roche.show more

Day 11/90 of Inference Engineering How does vLLM work... and how is it used in production? Before we discuss how vLLM works internally, it helps to understand what vLLM is. At a high level, vLLM is an inference engine that is designed to serve LLMs to thousands of concurrent users efficiently while managing scarce compute and memory. The goal for vLLM is to maximize throughput and minimize latency; optimizing for the best inference economics and experience for end users. With every request from the end user, it eventually ends up in the engine core, gets scheduled alongside other requests from other concurrent users, executes on the GPU, and updates the KV cache with the new key and value vectors, and streams the tokens back to the user. The Scheduler decides what requests should execute next while continuously batching requests together to maximize GPU utilization. Continuous batching is an inference optimization that allows new requests to join a running batch as other requests finish generating tokens. This helps with keeping the GPU utilization high instead of letting it sit idle waiting for an entire batch to complete generating. After the scheduler dispatches the selected batch to the Model Executor, the Model Executor prepares the tensors and metadata required for inference, retrieves each request’s block table from KV Cache Manager, launches the optimized transformer forward pass on the GPU, computes the logits, updates the KV cache with the new key and value vectors, and finally returns the results for sampling and streaming. The KV Cache Manager uses the PagedAttention memory layout to allocate fixed-size cache blocks on demand and maintains a Free Block Queue on the CPU that tracks which blocks in the GPU’s Paged KV Cache are currently free. When a request needs additional KV cache space, the KV Cache manager takes a free block from the queue and assigns it to that request, thus avoiding an expensive search through GPU memory for available cache blocks. All of these components form the core of vLLM’s inference engine. The Scheduler determines what requests are executed, the Model Executor determines how those requests are executed, the KV Cache Manager determines where each request’s KV cache lives using the PagedAttention Memory Layout. This architecture enables vLLM to serve thousands of concurrent requests with high throughput, low latency, and efficient GPU memory utilization. Heres a little animation that visualizes everything! - I've also completed the forward pass for my mnist.c project. I had a nice chat with shrey birmiwal, such a knowledgeable guy. Excited to learn more about vLLM and implement a tiny-vLLM one day.show more

Google dropped a new AI paper called LUMIERE. It's... remarkably flexible, supporting video inpainting, image-to-video, AND stylized video generation tasks. Say hello to “space-time diffusion” for video generation! Now what the heck does that mean exactly?! 🌐⏳ → TL;DR it utilizes a “Space-Time UNet” architecture that generates the full duration of the video in one pass, rather than generating distant keyframes and interpolating between them like prior works. Because the computation is done in this “compressed space-time representation” to generate the full clip at once, it's far more temporally consistent. → Another benefit of generating the full video at once is that you can “direct” the video generation, making it easier to hand off to other models/tasks without having to stitch together partial solutions. You can condition generations on additional inputs, meaning you get the full stack of AI video capabilities – from video inpainting to image-to-video and beyond. → New SOTA for AI video generation? User study results in the paper suggest human evaluators preferred Lumiere over Runway Gen-2, Pika Labs, and Stable Video Diffusion in terms of quality, text alignment AND motion. But as always, we need to get hands-on with this tech when Google *actually* decides to ship it. → Could this end up inside YouTube? Y’all know i’m obsessed with blending reality and imagination – so it’s the video inpainting tech I'm most excited about. I really hope this model finds its way into YouTube's Generative AI efforts, and based on their prior announcements and the list of acknowledgments in the paper I think it might! 🤞🏽 Links: 🔗Paper: 🔗Project:show more

The EU just passed Chat Control today. Here's what... it is, what happened, and why 2.0 is the one that actually matters. What is Chat Control? a regulation that lets online platforms scan private messages for child sexual abuse material (CSAM). Framed as child protection. critics call it mass surveillance infrastructure because a system built to scan for one type of content can scan for anything. the timeline: → March 2026: EU Parliament voted to let Chat Control expire. 307-306. one vote. → April 3: the law expired. legal basis for scanning gone. → July 2: EU Council repackaged the same proposal and brought it back → July 7: Parliament fast-tracked it using a rarely-used emergency procedure 331 to 304 → July 9 (today): passed. extended to April 2028. the procedural trick: by using the second-reading procedure, the Council forced Parliament into an impossible position — to block it, they'd need 361 votes, an absolute majority of all MEPs, not just those present. timed for the last session before summer recess. the EU's own Legal Service had already said the proposal violates the EU Charter of Fundamental Rights. didn't matter. what it actually scans: Instagram DMs, Discord, Snapchat, Xbox messaging, Gmail, iCloud Mail. End-to-end encrypted apps like Signal and WhatsApp are not directly affected; voluntary scanning can't work on messages platforms can't read. yet. YET: Chat Control 2.0, the permanent version, is still in negotiation. it would: → require mandatory scanning, not voluntary → target encrypted messages directly → allow detection orders without a court order or reasonable suspicion → include mandatory age verification across messaging platforms if 2.0 passes in its current form, end-to-end encryption in the EU is functionally over. the global implication: when the EU creates legal infrastructure for scanning private messages, authoritarian governments use it as a template. and any backdoor built for Brussels is technically a backdoor for anyone who can access the same system.show more

Break-A-Scene: Extracting Multiple Concepts from a Single Image introduce... the task of textual scene decomposition: given a single image of a scene that may contain several concepts, we aim to extract a distinct text token for each concept, enabling fine-grained control over the generated scenes. To this end, we propose augmenting the input image with masks that indicate the presence of target concepts. These masks can be provided by the user or generated automatically by a pre-trained segmentation model. We then present a novel two-phase customization process that optimizes a set of dedicated textual embeddings (handles), as well as the model weights, striking a delicate balance between accurately capturing the concepts and avoiding overfitting. We employ a masked diffusion loss to enable handles to generate their assigned concepts, complemented by a novel loss on cross-attention maps to prevent entanglement. We also introduce union-sampling, a training strategy aimed to improve the ability of combining multiple concepts in generated images. We use several automatic metrics to quantitatively compare our method against several baselines, and further affirm the results using a user study. Finally, we showcase several applications of our method paper page:show more

Seedance 2.0 - Advanced Workflows Series 14. Compositing and... Layering Some shots may have many characters that need specific movements. When we start adding characters, it becomes very difficult to achieve the exact animation, gesture, or movement we are looking of all of them in a single generation. In addition, the more characters are added, the more rigid Seedance 2.0 animations become in order to maintain consistency. This is especially visible in styles with squash and stretch and rubber hose animations, where the characters deform their body parts as if they were made of rubber, or in highly expressive movement animations in anime. For these cases, it is better to work on each animation separately and recompose it at the end into a video. Workflow + Prompts 👇show more

2 columns of white humanoid robots marched through 6... street fires to protest the humans who built them. Arms raised. Perfect step. Fires burning on the asphalt, a car smoldering down the block, and the machines walking straight through it like they’d seen riots before. The crackdown came fast. 4 men in body armor waded into the column with hammers. Plastic shells cracked. White limbs scattered across the wet road. One soldier scooped an armful of robot parts and stuffed them down an open manhole. A detached head slid across the lane. 15 seconds. March to massacre. Here is the part that matters. No robot marched. No soldier swung. Nobody stood on that balcony filming. The clip is AI-generated end to end. The street, the fires, the crackdown - rendered. The shaky vertical phone aesthetic is a style prompt. And that’s the real story. Robots as protest proxies is already a live debate - machines standing in the street so humans don’t take the baton. This video skipped even that step. The protest was synthetic. The violence was synthetic. The witness was synthetic. Every human in the pipeline got replaced before a single robot shipped. Robots don’t need to march against you yet. The video of them marching already does.show more

Disappointments surround us from all sides. We feel broken..... But we will definitely come back from under the rubble.. We in the north Gaza cling to the land like the roots of trees clinging to the soil. There is no wind that can uproot people who want to remain in their place. We are not just numbers.... We have a case. But there is no justice in this world.. My friends and loved ones.. Based on your advice, I decided to buy cat food and not wait. But I decided not to buy for a very long time for fear of wasting money. I also want to buy now so that I do not feel regret if a ceasefire occurs. Everyone knows that we have to buy from the black market, and unfortunately the prices are very ridiculous. So I will buy enough to last 3 months.. I will try to do my best to find food for cats, but if I cannot, the alternative option will be to use canned human food... I already have enough cat food to last until the end of December. But I want to buy it now because canned meat is available now.. I also mix bread with cat food to make it last longer...as shown in the video below I had a beautiful shelter before the war.. it was completely destroyed.. It had more than 60 cats. There are now about 25 cats left, and there are some cats nearby who find love and food, so they come to eat with the cats in the shelter. I would be very grateful for your support at the following link You can donate via PayPal There are days left until the end of 2024. It has been a year full of pain, patience, pain, fear, hunger, humiliation, oppression and suffering. I don't want this year to end with cats suffering as we do. Your support will help save cats from starvation I am very grateful to everyone who supports me morally and financially, and to everyone who retweets. Thank you all, friendsshow more

This is my "feel the AGI" moment: I used... GPT-5.6 Sol to train my own autocorrect model that outperforms GPT-5.6 Sol (wtf??) I have no ML background. I have no idea what I'm doing. I just kept pushing Sol until it spat out a SOTA model. And I spent $0. The motivation: Years of talking to AI have made me terrible at typing. Rather than fix my skill issue, I decided to throw more AI at it. My idea was: instead of autocorrect that interrupts my flow, I want to type fast with mistakes and have AI clean it up after. I wanted the smallest local model possible, for speed, for battery life, for science! So I decided to train my own. Inspired by Andrej Karpathy’s autoresearch, I ran Codex /goal with this setup: pick an experiment, try it, record the results to a doc, throw it out if it fails, and plan the next experiment without repeating failures. I gave a few examples that had to pass, tight latency targets, and let it run. Sol did some amazing things. First, it scanned benchmarks and shortlisted base models: Qwen 3.5, Gemma 4, Liquid LFM 2.5. It found a dataset on HuggingFace for typed text. Then it built a simulator for fingers striking a Mac keyboard, modeling the physical layout with a Gaussian distribution around each key. It simulated striking the wrong key, wrong order, fat-fingering, etc. With the models + data + simulator, it fine-tuned using MLX right on my MacBook. It had a working prototype within an hour! But accuracy was pretty poor. — Problem 1: Tokenization Sol read papers, ran tests, and identified that the tokenizer was the bottleneck. Tokenization makes typos hard for the model to see, so it memorizes mappings instead of using its language priors. Sol tried ByT5, Google’s tokenizer-free byte-level LLM. This made a big improvement, but the model is old and lacked the knowledge needed to reach Sol performance. Sol dug deeper and realized a tokenizer-free model isn’t needed; instead, it used T5Gemma, an encoder-decoder model. This can understand the input deeply before producing output, and furthermore, Sol could post-train the encoder to improve performance. This gave a much higher ceiling. — Problem 2: Loss function Now the model was correcting some typos perfectly, but ignoring most. Sol realized that standard cross-entropy loss was teaching the model to avoid edits, because the vast majority of characters in the training data were left unmodified. The fix was wild: Sol wrote a custom loss function that byte-aligns the source and target strings, uses a dynamic programming algorithm to compute the minimum edits between the two, then weights correct edits much higher than copies. After a lot of tuning, this dramatically improved accuracy. — Problem 3: Autoregression One failure mode remained: if the model made a mistake, it couldn’t backtrack. It could only predict the next token. Teaching it to “think” like a reasoning model would solve this, but would be far too slow. Sol found a beautiful solution: instead of greedily predicting the next token, beam search over all possibilities. This parallelizes the exploration instead of one linear chain-of-thought. At the end, choose the path with highest cumulative log probability. This worked great, but made the experience worse, since the user wouldn’t see progress until the whole search was done. To fix this, Sol made a clever observation: after each search step, the longest common prefix among surviving branches is guaranteed to appear in the final result, so it can be displayed immediately. As the search progresses, weaker paths are dropped and the prefix grows, so the user sees continuous progress. Sol built all this as a custom MLX pipeline that does the parallel decoding on the MacBook GPU, with just ~40ms TTFT. It’s crazy fast and entirely local. — Final eval (error reduction rate, higher is better): - Apple autocorrect: 49.66% - GPT-5.6 Luna: 82.47% - GPT-5.6 Terra: 87.64% - GPT-5.6 Sol: 90.56% - Our model (1.7B): 91.02% Final cost: - 1 quota reset (thanks Tibo) - $0 (And yes, I verified there's no cheating. In fact, we test words scrubbed from the training data to prove the model isn’t memorizing) There were a ton more details and tangents I could write about: contrastive learning, GRPO, DPO, dynamic masking, and more. Sol is a fascinating and creative model. It blew my mind so many times. Don’t let a lack of experience stop you: Sol makes AI experiments accessible to anyone!show more

Gemini Omni + GPT Images 2 + Claude Code... is f*cking cracked i just built an AI animation ad generator that turns any product into a fully scripted AI 3D explainer video drop in your product photos and a one-line pitch. the system analyzes your brand, pitches you 4 proven ad concepts with hooks and shot lists, then renders the whole thing end-to-end. scripted, voiced, stitched, ready to post. if you're still paying editors $200+ per video or waiting days for revisions this replaces that entire workflow. here's how it works: > set up your brand kit with product photos and a one-line pitch > hit analyze and the AI pitches you 4 different proven ad concepts > pick one and hit generate the system then scripts every shot, renders each clip in parallel, adds voiceover, QC checks and stitches it into a finished vertical ad 3 minute setup and 4-12 min per ad. each 30 second ad costs under $3 in API credits which is cheaper than any editor and 10x faster. RT + reply "ANIMATION" and i'll send you the full app + setup guide (must follow so i can dm)show more

A Letter to Our Community: The Road Ahead for... Robotics To our Community and Partners, As we step into 2026, our mission at Axis is clearer than ever: Constructing the definitive End-to-End Scaling Layer for Robotics. Our goal is to accelerate the transfer of diverse human intelligence into Robotics General Intelligence (RGI). By owning the critical path of intelligence creation, we are turning the physical limitations of robotics into a scalable, software-driven future. Here is our strategic outlook and roadmap for the year ahead. The Core Thesis: Simulation is the Only Way Out The path to RGI is currently blocked by Data Scarcity, Generalization Fragility, and Hardware Fragmentation. At Axis, we believe Simulation is the only way out. Our Simulation Data Platform and Data Augmentation Engine transform raw data into "Synthetic Gold". Backed by academic milestones like Roboverse, Skill Blending, and GraspVLA, we have proven that pure simulation can achieve the generalization required for the real world. We don’t just collect data; we architect it. The Engine: Why Crypto? We believe RGI should come from all, not a few. Crypto is not just a feature; it is the primitive that powers our entire ecosystem flywheel: - Incentive Mechanism: Democratizing contribution and rewarding the trainers and developers. - Assetization: Turning proprietary data and refined models into liquid, ownable assets. - Verifiable Workflow: We are opening the "Black Box" of AI. By bringing total transparency to the Task Generation → Data Collection → Model Training pipeline, we ensure every byte of intelligence is verifiable, traceable, and secure. 2026 Strategic Deliverables This year, we are committed to delivering three foundational pillars: - The World's Largest Training Dataset for Robots: A robot training set—diverse, high-quality interaction data at an unprecedented scale. - A Robotics Foundation Model: A universal robotic brain trained on our pure simulation and synthetic data, capable of robust cross-embodiment transfer and open-world adaptability. - Evolvable Robot Hardware: Robots deployed with Axis models that autonomously evolve through continuous interaction, turning every deployment into a self-improving node within our RGI network. The Ultimate Vision We are building more than models; we are architecting the Distributed Machine Economy. A future where every dataset, model, and robotic embodiment is a verifiable asset in a global, autonomous network. Thank you for building the future of intelligence with us✌️📷show more

So I’ve lived in Hillsboro, Oregon for 10 years.... Drove around today and tonight and shot this myself. This is what the “Data Center Plains” looks like.👇 My town sits at the end of 6 transpacific sea cables connecting the US to Asia. That’s why 30+ data centers landed here. They’re everywhere. Spread across the entire north and west end of the city. Road after road. Building after building. Miles of it. And they keep building. Pushing further west every year into farmland that’s been here for generations. Buying up land, Giving mass amounts of money to home owners to move, Tearing down homes. Tearing down historic sites. $7.2 billion in exempted property taxes. Some of these finished buildings are literally sitting completely dark… PGE told them no power for 3-5 years. They still built them anyway. A power plant is now going up right next to the data centers because they maxed the local grid. There are families still living next to construction zones. Old farmhouses directly across the street from data center walls. Nobody asked the people who already lived here. These are the families who refused to leave, so they said we are just gonna put them up next to your houses anyway. A pioneer homestead from 1865, 190 years of continuous farming is about to be gone, NTT Global Data Centers got that land tax-free until 2051. Signed in a single day at City Hall. Intel, the employer that actually brought thousands of real jobs here is laying off locals at the same time. This sound runs 24 hours a day. 7 days a week. My electricity bill went from $80 to $150. Water rates are set to increase 105% over 5 years, critics say to fund data center infrastructure, not residents. Data centers aren’t a joke. Just wanted to share my first hand experience with them. #datacentersshow more