Yes, if you're doing what everyone else is doing you can just use tensor cores and libraries which optimize for that.
Contrarily if you're doing something that doesn't map that well to tensor cores you have a problem: every generation a larger portion of the die is devoted to low/mixed precision mma operations. Maybe FGPAs can find a niche that is underserved by current GPUs, but I doubt it. Writing a cuda/hip/kokkos kernel is just soo much cheaper and accessible than vhdl it's not even funny.
AMD needs to invest in that: Let me write a small FPGA kernel in line in a python script, compile it instantly and let me pipe numpy arrays into that (similar to cupy rawkernels). If that workflow works and let's me iterate fast, I could be convinced to get deeper into it.
The primary niche of FPGAs is low latency, determinism and low power consumption. Basically what if you needed an MCU, or many MCUs but the ones in the market don't have enough processing power?
The Versal AI Edge line is very power efficient compared to trying to achieve the same number of FLOPs using a Ryzen based CPU.
> Before too long (and we already start to see this) humanoid robots will get wheels for feet, at first two, and later maybe more, with nothing that any longer really resembles human legs in gross form. But they will still be called humanoid robots.
Totally agree. Wheels are cheaper, more durable and more effective than legs.
Human would have wheels if there was an evolution pathway to wheels.
The world is full of curbs, stairs, lips, rugs, vehicles, etc. If you're a human-scale robot then your wheels need really wide base to not tip over all the time, so you are extremely awkward in any kind of moderately constrained space. I wouldn't exchange my legs for wheels. Wheelchair users have to fight all the time for oversights to be corrected. I can see maybe see a wheel-based humanoid robot, but only as a compromise.
On the other hand there is not much reason to constrain ourselves to the unstable and tricky bipedal platform or insist on having a really top-heavy human-like torso. You could have 3-4 little legs on a dog scale body with several extra long upwards reaching arms for eg.
It's hard to see one. Even a nice flat world with ample incentive and taking good "bearings" for granted, how can you evolve a wheel-organ that maintains a biological connection as well as being able to rotate an indefinite number of time?
A few difficult and grotesque endpoints:
* The wheel only rotates a fixed number of times before the creature must pivot and "unwind" in the opposite direction. This one seems most plausible, but it's not a real wheel.
* The main body builds replacement wheels internally (like tooth enamel) and periodically ejects a "dead" wheel which can be placed onto a spoke. This option would make it easier to generate very tough rim materials though.
* A biological quick-release/quick-connect system, where the wheel-organ disconnects to move, but then reconnects to flush waste and get more nutrients.
* A communal organism, where wheel-creatures are alive and semi-autonomous, with their own own way to acquire nutrients. Perhaps they would, er... suckle. Eeugh.
In one of Philip Pullmans His Dark Material novels there is a race of creatures that have a symbiosis with a tree whose huge perfectly round nut can be grasped by their fore and hind limbs and they roll around that way.
Wheels are not balls. Balls are common in nature. Wheels are not. The difference is that wheels need roads, which are not common nature and a large scale artificial objects.
Wheels are great until the robot encounters uneven surfaces, such as a stairway, or a curb. So some kind of stepping functionality would still be necessary.
I sometimes imagine wheeled creatures evolving in a location with a big flat hard surface like the Utah salt flats or the Nazca desert, but I guess there's not much reward for being able to roll around since those places are empty as well as flat. Tumbleweed found some success that way though, maybe?
The golden wheel spider lives in the sand dunes of the Namib Desert. When confronted by a spider-hunting wasp, it can perform a "cartwheeling" maneuver to escape. By tucking in its legs and turning onto its side, it can roll down a sand dune.
Is there any biological examples of freely rotating power systems? We have nice rotating joints with muscles to provide power, but I can't think of any joint that would allow the sort of free rotation while also producing torque, a wheeled animal would require.
Something internal to some shellfish, I believe, a kind of mixing rod that rotates. Hold on, I'll check if it's powered. (Also rotifers but they're tiny.)
Hmm, no, it sounds like it's externally powered:
> The style consists of a transparent glycoprotein rod which is continuously formed in a cilia-lined sac and extends into the stomach. The cilia rotate the rod, so that it becomes wrapped in strands of mucus.
Or maybe the cilia ( = wiggly hairs) could be seen as a kind of motor. Depends how you count it and exactly what the set-up is, I can't tell from this.
I think I would count internal power created by the rotating component itself. I hadn't though of that possibility, since human made machinery usually has the power producing component located in the main body and transferring that power to a freely rotating component is quite hard. Biological systems wouldn't necessarily look like that, and could feasibly be powered by the wheels themselves deforming as if the wheels were a separate, but connected, biological system.
Its evaluation now purely relies on NNUE neural network.
So it's an good exmaple of the better lesson. More compute evently won against handwritten evaluation.
Stockfish developers thought old evaluation would help neural network so they kept the code for a few years, then it turned out that NNUE neural network didn't need any input of human chess knowledge.
AFAIK this is only a part of it, it still has its opening library, as well as the end-game one (IIRC chess is a solved game for up to seven remaining pieces on the board).
Also, the PR you link says the removed code in fact had a performance impact, just too low to justify its code size 25% of all Stockfish).
WARP is a full VPN tunnel. The above poster was suggesting if someone wanted to avoid a VPN tunnel, DNS64+NAT64 is a nice "hack" that just uses DNS as the tool to reestablish IPv4 connectivity. (It's also how most mobile/cell traffic today reaches IPv4, via DNS64+NAT64 gateways.)
If you have worked or lived in China, you will know that Chinese open-source software industry is a totally shitshow.
The law in China offers little protection for open-source software. Lots of companies use open-source code in production without proper license, and there is no consequence.
Western internet influencers hype up Chinese open-source software industry for clicks while Chinese open-source developers are struggling.
These open-weight model series are planed as free-trial from the start, there is no commitment to open-source.
> Western internet influencers hype up Chinese open-source software industry for clicks while Chinese open-source developers are struggling.
That kind of downplays that Chinese open weights are basically the only option for high quality weights you can run yourself, together with Mistral. It's not just influencers who are "hyping up Chinese open-source" but people go where the options are.
> there is no commitment to open-source
Welcome to open source all around the world! Plenty of non-Chinese projects start as FOSS and then slowly move into either fully proprietary or some hybrid-model, that isn't exactly new nor not expected, Western software industry even pioneered a new license (BSL - https://en.wikipedia.org/wiki/Business_Source_License) that tries to look as open source as possible while not actually being open source.
They took a simple technique (normalizing flows), instantiated its basic building blocks with the most general neural network architecture known to work well (transformer blocks), and trained models of different sizes on various datasets to see whether it scales. Looks very bitter-lesson-pilled to me.
That they didn't scale beyond AFHQ (high-quality animal faces: cats, dogs and big cats) at 256×256 is probably not due to an explicit preference for small models at the expense of output resolution, but because this is basic research to test the viability of the approach. If this ever makes it into a product, it'll be a much bigger model trained on more data.
EDIT: I missed the second paper https://arxiv.org/abs/2506.06276 where they scale up to 1024×1024 with a 3.8-billion-parameter model. It seems to do about as well as diffusion models of similar size.
"""The bitter lesson""" is how you get the current swath of massively unprofitable AI companies that are competing with each other over who can lose money faster.
I can't tell if you're perpetuating the myth that these companies are losing money on their paid offerings, or just overestimating how much money they lose on their free offerings.
If it costs you a billion dollars to train a GPT5 and I can distill your model for a million dollars and get 90% of the performance, that’s a terrible deal for you. Or more realistically, whoever you borrowed from.
The bitter-er lesson is that distillation from bigger models works pretty damn well. It’s great news for the GPU poor, not great for the guys training the models we distill from.
Maybe for a big llm, but if they add some gpu cores and added a magnitude or 2 more unified memory to their i devices, or shoehorned m socs into high tier iDevices (especially as their lithography process advances), image generation becomes more viable, no? Also, I thought I read somewhere that apple wanted to infer simpler queries locally and switch to datacenter inference when the request was more complicated.
If they approach things this way, and transistor progress continues linearly (relative to the last few years) maybe they can make their first devices that can meet these goals in… 2-3 years?
somewhat hard to say how the cards fall when the cost of 'intelligence' is coming down 1000x year over year while at the same time compute continues to scale. the bet should be made on both sides probably
I believe the 1000x number I pulled is from SemiAnalysis or similar, using MMLU as the baseline benchmark and the cost per token from a year ago to today at the same score. Model improvements, hardware improvements and software improvements all make a massive difference when combined to make much greater than 10x gains in terms to intelligence/$
After Transformer took over AI, FPGA for AI is totally dead now. Because Transformer is all about math matrix calculation, ASIC is the solution.
Modern Datacenter GPU is nearly AISC now.
reply