Bayley

Everything's a robot | Cofounder/CEO @PrismaXai

[Bayley @castorhat]

this, my friends, is why there is a sim2real gap

[Max Sumrall @MaxSumrall]

I've gotten Microduck @pollenrobotics up to 1.8 / 1.9m/s, if the sim is to be believed.

[Bayley @castorhat]

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...

[Bayley @castorhat]

@NVIDIARobotics pushing for running the model on the robot is great, but AGX Thor costs more than the BoM of most robots now :(

[Bayley @castorhat]

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?)

[Zhen @zhen_do_ob]

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.

[Bayley @castorhat]

continuing my crusade against AI generated reports: https://t.co/b5yUNE4FKN

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.

[Bayley @castorhat]

business has been great but more time with customers means less time to tweet (plus IROS is coming, catch us there!).

in the meantime here is a teaser of my next piece: frontier AI (deploy big LLMs, train robotics models) on a used Ethereum miner

[Bayley @castorhat]

this is just embarrassing https://t.co/1xoN3xHPXI

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...

[Bayley @castorhat]

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

[Bayley @castorhat]

fun fact: if you switch from Opus 5 to Fable 5 mid session, everything is bad

[Bayley @castorhat]

the fascinating conclusion of the "ban open source" saga: https://t.co/QkZNbKoLKN

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

[Bayley @castorhat]

bootleg silver A100's - full parameter finetunes on openpi for under $1000!

[Bayley @castorhat]

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?

[Bayley @castorhat]

The weights for GLM-5.2 have been downloaded over 3 million times in the past month. Given that there are fewer than 3 million installations that can run GLM-5.2, I guess most of the downloads are weight hoarders?

Feels like an untapped market somehow...

[Bayley @castorhat]

continuing the story: some of you guys may know that my weekend project for the past few weekends has been getting a frontier model running on a friend's pool heater. the pool heater is two nodes - a head node which also has the dev harness deployed, and an aux node which is just compute.

the harness accesses the aux node via passwordless ssh from the head node. this works great - Claude's ability to chain commands over ssh is uncanny.

during an...incident today, the aux node ran out of memory which caused it to start swapping. unfortunately, the aux node is not a real computer - it is a virtual machine with the root volume mounted over 10G NFS. the resultant thrashing destroyed one of the SSDs in the NFS volume which, due to some poor life choices we were well aware of, killed the array.

this happened at an unfortunate time because I had to step out for a couple ours, but before I did, I instructed Claude to "wait until the aux node came back up and then continue" (it knows how to do that, by polling SSH) and a friend that "the aux node seems to have OOM'ed, can you reboot it?". of course, in the Uber I find out that the aux node wasn't just OOM, it was outright gone.

the friend kindly rebuilt the array and the VM. Claude (which had been waiting patiently in the 90 minutes of downtime) logged into the rebuilt VM, discovered there was nothing left of the old instance, read its own notes, rsync'ed the files it needed over the internode 100G Infiniband, and proceed happily as if nothing happened. by the time I got home, it had figured out how to run GLM-5.2 with cudagraphs.

moral of the story: don't fuck up, but if you do, the AI's got your back

[Bayley @castorhat]

I gave Claude Fable access to one of my HPC nodes and it managed to swap so hard it killed (?) one of the disks. Not sure if impressed or annoyed...

[Bayley @castorhat]

I gave Claude Fable access to one of my HPC nodes and it managed to swap so hard it killed (?) one of the disks. Not sure if impressed or annoyed...

[Bayley @castorhat]

So the big news today is how the guys in China finally did another Top500 submission, at #1. To be clear - China has had a ~Top10 HPC system for the past few years - the upgraded TaihuLight - but has refrained from making a submission for fear of reprisals in the form of sanctions. This is probably a good idea - nation-scale systems are unashamedly used for weapons design and flaunting this capability in front of the world when you are not using fully domestic technology will generally lead to a bad ending.

The new submission - LineShine - achieves 2.2EFLOPS on HPL using 45360 Huawei LX2 CPUs connected using a proprietary fat-tree network. The interesting thing here is the LX2 CPU, which is the latest in a series of Chinese HPC processors built with a explicitly-memory-managed, DSP-like architecture. We briefly dive into the architecture below, as well as some historical comparisons.

All HPC devices have a shared memory-global memory architecture, where shared memory is a fast local SRAM accessible by a subset of the processing units and global memory is DRAM which is accessible by all processing units. Where they differ is how this hierarchy is managed:

- CPUs use the shared memory as cache, hiding it entirely from the programmer
- GPUs use the shared memory as cache, but additionally allows explicit shared memory loads
- DSPs only allow computation on shared memory, relying on explicit loads to transfer data from global memory

Additionally, CPU/DSPs and GPUs differ in how they expose parallelism to programmers:

- CPU/DSPs explicitly reveal cores and SIMD lanes: there are N cores, each of which an execute m operations per cycle
- GPUs abstract cores x SIMD lanes as threads: to the programmer, there are N threads, each of which work on one element at a time. Internally, GPUs are a complex share-resource SIMD multicore processor - GPU "cores" are not fully independent of each other, but are more so than the SIMD lanes in a CPU

The LineShine/Huawei LX2 is a 304-core ARM DSP which is physically two 152-core processors and electrically 8 38-core core groups. Each core group has access to 4 GB of HBM and each 152-core processor has access to 128 GB of DDR5. While it appears that cores do have access to memory directly via load/store instructions (as an ARMv9-compliant device, it would be impossible not to), the primary way of memory access is through the DMA engines, a block which lives in the uncore and asynchronously moves data between memory levels.

The most famous DMA-driven processor was the Cell Broadband Engine, a 1+8 core DSP where the 8 cores could only access data via DMA transfers. Its spiritual successor is the Sunway SW26010P, a 384-core processor built as 6 64-core complexes, each in the same style as Cell. LX2 is a divergence from the pure DMA architecture - there is no management core - but the idea remains the same: explicit programmer management of data transfer leads to high performance in the hands of a skilled programmer.

[PrismaX @PrismaXai]

The First 100 is live.

We partnered with 11 of the best teams in crypto and AI to get here. They tested the platform, scored real robot data, and helped shape the standard before launch.

Meet our Day 1 Launch Partners.

[PrismaX @PrismaXai]

The First 100 begins now.

Verify Quality is live on PrismaX. For the first time, anyone can score the robot training data that models learn from, earn points, and compete to become one of The First 100.

Better data. Better models. The standard starts with you.

[Bayley @castorhat]

huge congrats to the awesome team at @ambient_xyz ! even if you don't care about the philosophy of open access, here's why what they're doing is important:

post-2026 LLM workflows rely not just on LLM access, but consistent, repeatable behavior from the same model, call after call, month after month. we're not talking about some customer service chatbot that sits on your webpage doing nothing, but serious, internal tooling designed to act like a junior developer or junior marketing associate, taking over repetitive tasks so that your humans can focus on being productive human beings.

these tools need to be well-characterized - it's honestly OK if they make mistakes once in a while, but their users need to know what mistakes can happen and how often in order to make the correct decisions. random infrastructure changes that may feel benign or even beneficial for run-of-the-mill chat users can wreak havoc on carefully-tuned agentic pipelines.

this brings me to my point: in tech, nothing lasts forever. it may feel unthinkable that Anthropic, or Fireworks, or Google, will be gone someday, but if you look at this industry's history the only proprietary closed-source standard that has survived until 2026 is the Microsoft Windows API. if you built your company on HP-UX, or SGI, or Netscape, or DEC Alpha, you'd be out of luck today - either the standards or the companies themselves have ridden off into the sunset.

Ambient is a huge step in the right direction. even if all the organizations named in this post vanished today, Ambient could still live on. being a proper, permissionless system means that as long as there is a single person interested in the project, the project can be revived. projects like Ambient are a vital part of solving a hard problem: how do we bring the same kind of decentralized resilience that the human-generated Internet enjoys to knowledge stored in language models, while managing the fact that unlike the Internet, creating and using these models consumes tremendous physical resources?

[ambient.xyz @ambient_xyz]

In just two weeks: some of the largest supply-chain attacks in history, a major commercial LLM sabotaging legitimate ML research requests, and a US export ban kneecapping the world's access to a major model release. More than ever, we need Open, Verifiable Inference.

Today we start shipping direct responses. A🧵:

[Bayley @castorhat]

It's amazing how quickly technology progresses when there's money to be made. About a decade ago the state of the art in rack-scale interconnects was NUMALINK 7, which dated back to the SGI of the mid-90s and was used by weird scientific computing people for their aero/weather/nuclear simulations. The total number of customers was about 10 (SGI would put out a press release every time they sold a system). NUMALINK 7 was good for about 100 gbits/second per cable.

The spiritual successor to NUMALINK is definitely rack-scale NVLINK, which is similarly a copper-based, performance-oriented, short-range interconnect. NVLINK 5 is good for an incredible 14400 gbits/sec - over two orders of magnitude faster than NUMALINK in 10 years. And yes, people complain about the cost of NVL72 systems, but a rack of Blackwell in 2026 costs less than a rack of SGI in 2016...