Everything's a robot | Cofounder/CEO @PrismaXai
San Francisco
Joined August 2024
- Tweets381
- Following72
- Followers3.6K
- Likes190
weekend (actually, more like 3 weekends) project: we used Claude and a mass spec to reverse engineer a proprietary cryogenic refrigerant, successfully reconstructing it out of commodity gasses
I was excited to shit on the launch-day RTX Spark reviews (just to be clear, the N1X is good, the reviews are not) and was not disappointed:
- These guys windowscentral.com/hardware/… opened with "well, the CPU is disappointing because it is 10% slower than a Macbook Pro", benchmark Llama 3.1 8B and Stable Diffusion 1.5 (hello? is it suddenly 2023 again), run Hermes Agent with local Qwen-27B (but never mention performance numbers?), and, after failing to install OpenClaw, conclude that local AI is too convoluted
- Tom's Hardware runs this suspicious benchmark from AA (github.com/ArtificialAnalysi…) which replays agent traces to a depth of 64K tokens, carefully hiding prefill/decode times. Further compounding the issue is they run Qwen-27B in q4 which is not a valid setting, the model is too small and accuracy degrades measurably at 4bpw. After failing to run a AAA game at 1440p max settings on the poor 5060-class GPU in the N1X, they conclude with "it is too expensive" (the Tom's review is really bad and is an example of why we can't have nice consumer products)
- These guys hothardware.com/reviews/nvid… run the standard obsolete Llama + SD1.5 suite (seriously, why does this even exist?) and then run MLPerf Flux-2, where the N1X really shines...but perhaps a bit too much. I am not convinced that the Strix Halo results are anywhere close to optimal, especially given how much slower it is than the B390 integrated GPU on the LLM benchmark
Meanwhile, modern workloads such as H3 video generation, Flux/Qwen image editing, and 120B-class text generation, where competing hardware scores poorly or zero, are completely omitted.
good time to tweet about the @PrismaXai IROS demo, this was an pi-0.5 fine tuned (expert + r=128 LoRa) on about 1K episodes of data collected on-embodiment but off-rig (collected at one of our production data facilities), trained on about 60 A100-hours (actual full experiment was about 1K GPU-hours, there are a lot of checkpoints that we discarded...), ablated in sim + our real lab environment, then deployed sight-unseen into the venue against the very busy background
deployment was very hassle-free, we kind of just eyeballed the camera and arm spacing and it worked fine. stressing about calibrated rigs is fake news
the smooth performance comes from (1) very good data, 1x speed, no pauses, no mistakes and (2) a fast TensorRT inference stack we wrote which cut end-to-end step times to about 87 ms in fp16 -- apparently this is somewhat interesting so we will open source it after the conference
I think the robustness also comes from the good data, if you deploy into a new environment performance will be degraded by some amount and having strong baseline performance definitely helps
This is one of the rare instances where I see a vla model like pi 0.5 actually deployed to run at speeds I expect. Much better performance than most rollouts out there. Impressive work from @PrismaXai #IROS2026
did you know the insurmountable fortress that is CUDA very nearly fell? forums.developer.nvidia.com/…
I was there. this was right around 2012-2013. back then, NVIDIA's enterprise revenue was hundreds of times less than it is now and almost all of their sales came from gamers. unfortunately, the features that make a GPU good at CUDA (large shared memories, hardware for thread sync and communication, flexible dispatchers) are all dead weight for gaming.
the previous-gen GTX 580 was built for compute. this made it rather uncompetitive for graphics, and NVIDIA was bleeding market share and reputation to AMD. on the 600 series, they walked back, reorganizing the SM's and caches to build a part that was incredibly fast in games but kind of a joke for compute.
general purpose compute was saved by the Titan supercomputer, a US DOE project which provided the resources to design and tape out the GK110 GPU. GK110 later became the Tesla K40, and its legacy lives on in the -100 series of AI chips we know and love today.
lesson learned? consumer products suck. CUDA very nearly became unviable because of some line charts drawn by 20-something-year-old kids playing video games which showed that AMD was cheaper per frame. without CUDA, GPUs with fast interconnects would have shown up in 2021, not 2016. five years is a lot, especially right now, and we might not have gotten AI at all.
the Nvidia-Huggingface acquisition has people all riled up because it means Huggingface will stop supporting non-Nvidia hardware (I cannot imagine an Nvidia-branded Huggingface supporting AMD), but the real question people should be asking is why AMD doesn't support itself. They are worth almost a trillion dollars...
what in tarnation are these conclusions...
PI never seriously considered selling to anyone and they are deploying, they just don't talk about it because well, they have several billion dollars and don't need to spend time talking about things.
Sunday doesn't compete with any of these guys, they are a vertically integrated product company. why is Sunday on the list but not 1X?
Skild has a smart team and is kind of insane, but they are at best on par with Generalist/PI in the pure model, low hardware space.
What about DeepMind? OpenAI? Meta (Sangbae Kim works there and Sangbae is very good, his lab did the original low-cost actuator stuff at MIT that all of the modern robots are based out of?)
So uh, the winners that will likely emerge will likely be Dyna and Skild going head to head:
1) PI is cooked because they don’t deploy well, don’t have good research techniques and the talent flow there is not S tier - no wonder they were evaluating selling to Ant
2) Generalist is gonna be slower because they need to partner with startups to deploy and do evals which means their speed and knowledge won’t compound as quickly since the deployment partners have competing priorities
3) Genesis is just getting to lab demos and just starting on hardware so they’re behind
4) Sunday is full stack but they also haven’t truly deployed into real consumer environments and is competing vs Figure - the 99.1% solve of folding was ok but the demos clearly show they can’t get tput up.
So yeah, it’s pretty clear to me this will be a Dyna vs Skild two horse race since both have deployed, both have real revenue, but afaik only Dyna has successfully grinded down the 9s on quality, throughput, and cost for multiple real world customers while having all of the speed, knowhow, etc to compound both the data flywheel + b2b network effects.
QED I rest my case.
continuing my crusade against AI generated reports: pugetsystems.com/labs/articl…
I don't even know where to begin on this one. AI was clearly involved ("the working multi-GPU path" is a classic Claude-ism for "it didn't work, so I gave up" during infrastructure work).
After a short opener in which $3000 of GPUs are paired with $10000 of CPU, they proceed to benchmark Qwen3 8B - a bad model from late 2024 - in 16bpw on a 32GB card, when any recent small MoE, even quantized to 4bpw, beats it in speed, KV usage, and capability. Furthermore, the benchmarks were run with an input/output ratio of 500/500 tokens, a configuration not representative of anything.
Concluding that old models do in fact run, they proceed to benchmarking Qwen-27B, upon which vLLM immediately died due to a RCCL error (classic AMD nonsense). Normally, "the software failed to run" would be a good reason to leave a scathing review, but apparently determined to shill their R9700, they switch to llama.cpp...with a 4bpw GGUF that only uses a quarter of their very expensive GPU's capacity.
After mentioning that SGLang works (yet not using it?) and an incorrect infographic starring llama.cpp's --split-mode=row (which no one ever used), we get very un-stellar results of 23 tok/sec at conc=1 and 56 tok/sec at conc=8 with no prefill numbers whatsoever. They conclude with a nice pat on the head claiming Q4 is what gives them usable speed, when in fact Q4 is unusable for anything on this model.
They then dive into the most horrific cost analysis I've ever seen, comparing the cost of Claude Opus in the cloud to the cost of running Qwen-27B at home (why don't they compare it to the cost of Qwen-27B in the cloud?). Concluding nothing, they tell you that buying some R9700 is a good deal, just trust them, and deftly segway into the classic Z-Image-Turbo, vaguely achieving the performance of a midrange gaming laprop.
Finally (there's actually more but this article is so full of errors it is exhausting to list them all), they provide an oranges-to-watermelons comparison against a quad B70 (which is apparently "enough for 35B MoE models" because of course that's what people run, Qwen-35B in bf16 on their 128GB systems) followed by some rather poor buying advice.
The sad thing is, this site was once a go-to site for enterprise workstation benchmarks. A lot of small companies and IT professionals are going to use this post as buying advice, making some terrible mistakes along the way.
this is just embarrassing together.ai/blog/glm-5-3-vs-…
clearly written by AI, and the entire premise is wrong - 5.3 Flash isn't a 5.3 distillation, it is a preview of the GLM-next architecture
come on guys. I expect this slop from random YC startups, but Together AI is one of the most respected teams in the industry...
Apple announced their realigned Mac pricing and now everyone is complaining the top Studio costs more than a car, which brings up the question, why is a Honda Civic the pinnacle of engineering in consumer eyes?? no one says "wow this phone costs more than a refrigerator" despite the fact that refrigerators are damn useful objects
the fascinating conclusion of the "ban open source" saga: archive.ph/jEVK6
tl;dr - Dario has managed to sign himself and no one else up for government review. Overall, I'd call this a win for open source
while I concur that the cybersecurity implications of frontier LLMs are worth considering (if only to not invoke the wrath of the all-knowing overlords at Visa and MasterCard), in what way does banning Americans from using Chinese models help protect us from Russian hackers?