I don't think this is where you were going with your comment, but I'll mention this just because you're somewhat adjacent to a routine mistake in business: Uber is a people delivery company, but they've had a lot of bright engineers working for them on their infrastructure and software over the years, and that work has rippled out through the industry. Amazon (in VMWare's words) is "a company that sells books", and t…
And Google is the ad factory.
Doubleclick buying google's brand was the smartest business move ever.
> Both the full training run and the large-scale deployment are built entirely on AI ASIC superpods. Pretraining spans millions of accelerator-days across more than 35 trillion tokens, To think that Nvidia would not have any competition is quite laughable and Jensen knew that China would catch up. This is the reason why restricting GPUs as a temporary blockade does not work and they would just make all the Chinese AI…
US restriction on China is not just GPU, but total blockade of anything semi related all the way from fab equipment to final chips. Same restriction will work on any other country. But it does not work on China. Not only there are astonishing number of crazy smart AI researchers in China, but also the entire semi supply chain, from fab machines to GPU/XPU chips and software ecosystems, is advancing extremely rapidly. China will be the only country where every step of the AI supply chain from materials, fab equipment, lithography, 7nm to 3nm logic fabs, HBM, packaging, photonics, GPU/CPU/XPU, software ecosystem, frontier AI labs, power production and power generation equipment are all within a single country.
For comparison allow me to add chatGPT 5.5: "Choose U-235 if the goal is safe, boring, practical electricity generation. Choose Pu-241 only if the goal is specifically to consume/recycle plutonium in a reactor designed and licensed for that fuel. In brutal shorthand: Pu-241 is a better “fissile isotope” in some nuclear-physics ways, but U-235 is a much better reactor fuel in the real world." If only I knew anything a…
Am i the only one feeling my soul being emptied every time i read another "brutal shorthand" or "honest take"?
Dwarkesh Patel has AI/ML guests on his podcast. BoorishBears may have been referring to the Jensen Huang episode where they discuss TPUs: https://youtu.be/Hrbq66XqtCo?t=982
Specifically Dwarkesh couldn't understand that GPUs are not enough: it's GPUs plus multiple ecosystems to leverage them at massive scale during training vs inference. Instead of giving China open access to US controlled chips and creating a misalignment between labs that want to train a model on whatever is best, and hardware manufacturers that need labs to suffer the growing pains for their new ecosystems built from…
I don't listen to Dwarkesh but I'm aware of who is and his influence. I was baffled that he could not understand it...Don't know if he had his own agenda or just not intelligent (which is scarey for someone with influencec), but I sensed the frustration in Jensen Huang for something that is fairly obviously.
The same scenario happens all the time when the US takes away something from China and China doubles down, gets into survival mode and then beats the US.
Yeah, for me it seems like a if you have to ask you can't run it" type question. In general the TL;DR is that anything above 35B needs hardware you buy basically only to run large LLMs, and if you have that hardware you don't need to ask the question.
That's simply not true. ~70B models can run fine (albeit somewhat slow) on consumer hardware with 64GB RAM. There are heavily quantized (Q1.x) models that are still usable on similar hardware. Granted recently there haven't been a lot of models of this size, but still, 35B isn't really the practical limit. 35B is mostly the limit if you're using consumer grade GPUs with limited RAM and need the model to run fast. Peo…
Yes, this is true, but that's not what I'm saying.
I'm saying that 64GB+ personal computers are vanishingly rare outside builds that were specifically done with AI in mind.
Gamers never saw the need for them, and even in software development 32GB was the standard until AI came along.
Yes, there were specialized use cases where they did exist, and yes, some people just wanted to max out the Macbooks but.. it was rare.
1024 Huawei Ascend superpods = 50K 910C chips. That is a tiny tiny system. OpenAI uses _milions_ of GPUs for training On the other hand, this probably reuses the existing deepseek v4 architecture and weights. Maybe didn't need that much compute.
I'm sure it also takes more compute effort to be at the frontier, rather than being able to distill and poach ideas from the frontier. No mistake that it's the same handful of labs taking turns at or near the frontier.
I don't understand why people keep repeating this nonsense.
Anthropic claims deepseek has made 150K requests to their servers. Even if this number is correct, it takes far more requests to distill from a 3.2T model into a 1.6T model. 150K is closer to running a few benchmarks.
If anything, deepseek together with googles deepmind are the ones innovating while Anthropic and openAI are spending money and time on politics to try to hinder or ban competition.
Am i the only one feeling my soul being emptied every time i read another "brutal shorthand" or "honest take"?
It's genuinely depressing.
Yeah, but the default is extreme yap mode. I prefer "in a sentence" or "rows, not paragraphs." "Brutal" sounds like bait from an influencer trying to sell me a $400 course about how 4AM workouts will make me a millionaire.
There was some earlier speculation this is the model behind the stealth-released openrouter/owl-alpha model, that's been free for the last month.
Not speculation - they said it was.
Oh, usually when that is announced the stealth model updates to the unmasked one. They did that with "Ling 2.6 flash" which was elephant-alpha. I guess we may see an update today then
I asked about tiananmen square and it said "Too many requests, try again later" - this was my first question. I understand this is one data point but still ;/
I asked a question with "Search" enabled, with the app set to English, and got results back in Chinese. Interesting view into how the LLM responds to its context.
I can’t get any tool calls working. Seems to use a `` wrapper which the current harnesses I’m using don’t support