krasin
This is almost twice better than RTX 5090. This GPU will not be prefill constrained, I feel.
But I am surprised that they were able to pack so much flops into 300W.
To a degree, this is what copyright was supposed to do.
I really fail to see a problem with these tiny amounts of non-brittle material embedded into a solid case. It's still very dangerous, but it's locally dangerous (meters away), not at the scale of whole countries.
The high-power unit had 300 grams of Pu-238 in 1965. Given its 87.7 years half-life, only 187g of Pu-238 remaining. It's very hard to do much damage with this amount of radioactive material.
How large is the amount of plutonium in there? I highly doubt that it has the claimed potential.
Obviously, it's still not done, and yet to prove to be profitable, but their reactor design does suggest that they have a chance to make it work and replace a lot of CO2 emissions.
1. Original article, but paywall: https://www.scmp.com/news/china/science/article/3331031/chin...
2. Derivative (+based commentary), but no paywall: https://www.chosun.com/english/world-en/2025/11/07/MND3QUGUT...
They support FP8/BF8 with F32 accumulate and also IU4 with I32 accumulate. The max matrix size is 16x16. For comparison, NVIDIA Blackwell GB200 supports matrices up to 256x32 for FP8 and 256x96 for NVFP4.
This matters for overall throughput, as feeding a bigger matrix unit is actually cheaper in terms of memory bandwidth, as the number of FLOPs grows O(n^2) when increasing the size of a systolic array, while the number of inputs/outputs as O(n).
1. https://www.amd.com/content/dam/amd/en/documents/radeon-tech...
2. https://semianalysis.com/2025/06/23/nvidia-tensor-core-evolu...
The employees were robbed from having a big cash exit. Illiquid stock options from Windsurf were converted to illiquid stock options of Devin.
What's worse is that the well is now poisoned. I would advise against joining startups from now on, because I think that there's no upside for employees anymore.
The linked whitepaper is pretty useless, and I am saying as a big fan of diffusion-transformers-for-not-just-images-or-videos approach.
Also, Gemini Diffusion ([1]) is way better at coding than Mercury offering.
From their "order" page:
> UPERFECT Omega is currently in the concept stage and will soon be put into research and development.
>With these innovations, the team fabricated a 1.09 cm2 flexible monolithic perovskite/CIGS tandem solar cell. Competing with top rigid counterparts, the device achieved an impressive stabilized efficiency of 24.6% (certified at 23.8%), one of the highest reported values for flexible thin-film solar cells to date.
>After 320 hours of operation and 3,000 bending cycles at a radius of 1 cm, the device retained over 90% of its initial efficiency, demonstrating exceptional mechanical durability and long-term stability.
I always recommend quilter.ai: https://www.quilter.ai/ - they "eat" Kicad or Altium files, and produce autorouting and/or autoplacement. I tried them about 9 months ago and made a couple of boards that worked well.
So, yes, you can run DeepSeek R1 on it, but there are cheaper options (if we only talk about inference).
They just need a few more iterations.
TL,DR: Unitree is making a real attempt to make their robots to be at least somewhat repairable, but these robots are still very difficult to work with due to glued connectors, tight cable managements, hundreds of screws, lack of availability of spare parts, lack of documentation, etc.
They use llama2 to demonstrate that their compression method works. There are potential cases:
1. The method works on all / most LLMs. In this case, it does not matter on which model they demonstrated the effect.
2. The method only works on llama2, but not on other models. Given that they published the code, I expect that people will quickly test the method on many other models, so we will know that soon. And yet - there would be a scientific significance even if it works only on llama2, as it would mean that there's some special and good in that architecture.
But I would bet it's #1 - the method works on most of the models and they just picked whatever they had already had code bindings to, to save the effort.
I have not read it yet (it's being released in May), so can't give a true recommendation, but I did preorder it. The state of art is changing fast, and the book is expected to capture at least some snapshot of it.
Source: I am a father of two daughters with less than 2 years difference; and my brother is also less than 2 years apart from me. So, I lived my advice in both aspects: as a parent and as a child.
I hope to see new Jetsons based on Blackwell sometime in 2026 (they tend to be slow to release those).
I believe it will take on the order of 100M hours of training data of doing tasks in real world (so, not just Youtube videos), and much larger models than we have now to make general-purpose robotics working, but I also believe that this will happen.
I've saved your comment to my favorites and hope to revisit it in 10 years.
In the past two years two very important developments appeared around imitation learning: robotics transformer, that "eats" text + video and spits robot joints commands; and diffusion policies for robot control. Some starting points for this rabbit hole:
1. HuggingFace LeRobot: https://github.com/huggingface/lerobot
2. ALOHA: https://aloha-2.github.io/