[ 66512340.000 ] posts.x: Full talk on the RMSNorm paper, the CUDA streams bug, and getting it running in production:  |   [ 66512340.000 ] posts.x: A race condition between two CUDA streams once made a language model repeat itself mid-sentence and lag a step behind its own output. @f_makraduli…  |   [ 66502500.000 ] posts.x: Full talk on why robotics still has no internet-scale dataset, and Skild AI's omnibodied answer to that gap:  |   [ 66502500.000 ] posts.x: A robot that could look at a pattern of stacked blocks and rearrange loose ones to match it, by sight, dates to the 1960s. A teleoperated…  |   [ 66491520.000 ] posts.x: Full talk on how Dyna turned reward models and active learning into a 99.4% success rate on napkin folding:  |   [ 66491520.000 ] posts.x: A robot that succeeds 80 to 90% of the time on a task will string together ten flawless repeats less than 0.1% of the time. @JasonMa2020, CTO and…  |   [ 66436200.000 ] posts.x: Full talk on why Moonlake builds its simulations from causal, code-based object models instead of generated video:  |   [ 66436200.000 ] posts.x: A photorealistic walkthrough of a room and a simulation you can actually act inside of are not the same thing, and mixing them up is the mistake…  |   [ 66427320.000 ] posts.x: Full talk on automating rotational grazing with drones, satellites, and LLMs to get more cattle raised on pasture instead of feedlots:  |   [ 66427320.000 ] posts.x: Rotational grazing could put a lot more cattle back on pasture, but it's bottlenecked by labor: someone has to walk the field every day and decide…  |   [ 66417420.000 ] posts.x: Full talk on building a VLM-labeled, RF-DETR-trained detection pipeline and a toolkit of open vision models for coding agents:  |   [ 66417420.000 ] posts.x: Most vision language models are too slow to run in production, but they can label the dataset that trains the small model that does. @mervenoyann, an…  |   [ 66416280.000 ] posts.x: @HamelHusain That’s a fantastic question! Exploring the humanness of X accounts is a concrete way to determine …  |   [ 66409680.000 ] posts.x: Full talk on Perceptron's embodied foundation models, the video sparsity problem, and the scaling law trading teleop data for video pre-training:  |   [ 66409680.000 ] posts.x: Training a video model on an hour of footage means feeding in something like a million visual tokens, but only about 0.2% of them come with any…  |  

[ 66512340.000 ] posts.x: Full talk on the RMSNorm paper, the CUDA streams bug, and getting it running in production:  |   [ 66512340.000 ] posts.x: A race condition between two CUDA streams once made a language model repeat itself mid-sentence and lag a step behind its own output. @f_makraduli…  |   [ 66502500.000 ] posts.x: Full talk on why robotics still has no internet-scale dataset, and Skild AI's omnibodied answer to that gap:  |   [ 66502500.000 ] posts.x: A robot that could look at a pattern of stacked blocks and rearrange loose ones to match it, by sight, dates to the 1960s. A teleoperated…  |   [ 66491520.000 ] posts.x: Full talk on how Dyna turned reward models and active learning into a 99.4% success rate on napkin folding:  |   [ 66491520.000 ] posts.x: A robot that succeeds 80 to 90% of the time on a task will string together ten flawless repeats less than 0.1% of the time. @JasonMa2020, CTO and…  |   [ 66436200.000 ] posts.x: Full talk on why Moonlake builds its simulations from causal, code-based object models instead of generated video:  |   [ 66436200.000 ] posts.x: A photorealistic walkthrough of a room and a simulation you can actually act inside of are not the same thing, and mixing them up is the mistake…  |   [ 66427320.000 ] posts.x: Full talk on automating rotational grazing with drones, satellites, and LLMs to get more cattle raised on pasture instead of feedlots:  |   [ 66427320.000 ] posts.x: Rotational grazing could put a lot more cattle back on pasture, but it's bottlenecked by labor: someone has to walk the field every day and decide…  |   [ 66417420.000 ] posts.x: Full talk on building a VLM-labeled, RF-DETR-trained detection pipeline and a toolkit of open vision models for coding agents:  |   [ 66417420.000 ] posts.x: Most vision language models are too slow to run in production, but they can label the dataset that trains the small model that does. @mervenoyann, an…  |   [ 66416280.000 ] posts.x: @HamelHusain That’s a fantastic question! Exploring the humanness of X accounts is a concrete way to determine …  |   [ 66409680.000 ] posts.x: Full talk on Perceptron's embodied foundation models, the video sparsity problem, and the scaling law trading teleop data for video pre-training:  |   [ 66409680.000 ] posts.x: Training a video model on an hour of footage means feeding in something like a million visual tokens, but only about 0.2% of them come with any…  |