Comment by bilbo0s | Hacker News Reader

Comment by bilbo0s | Hacker News Reader

Parent Full thread

I had analyzed the paper as well, and gave some quick thoughts on a few of the low level things that stood out and would affect NVDA here:

https://news.ycombinator.com/item?id=42856652

I have to agree. I'm not sure the architectural and engineering changes here don't put NVDA further ahead. Some stuff all gpu makers can do. Like quantize on transfer. But NVDA is so far ahead in other areas and DeepSeek leans heavily on that industry leading tech. At least at the low level.

But I'm betting the average investor doesn't have the level of facility with ML to properly analyze the information that's come out. Heck, I'm betting most of the people working in AI are really just TF or PyTorch monkeys. The real AI experts have probably been hovered up by MS or FB or AMZN or OpenAI or what have you. So it feels like we could get lucky here and the AI related market could drop like a rock taking the good with the bad. Which I think would be awesome for most people in a position to wait for it to happen.

I can't wait for the 2000 dot com crash for AI.

What gets me about the deepseek work is that it's neither novel or theoretically difficult.

I've been using a similar synthetic data pipeline to generate examples for sota bert embedding models.

That low-hanging fruits like this are left unexplored by meta or alphabet - and if we're to be believed lead to crisis meetings for a week straight - makes one wonder wtf the engineers there are doing?

This is just a baseless conspiracy theory that I have, but I do wonder if they intentionally avoided certain avenues of research because it could erode the “moat” major tech firms have by dramatically reducing the capital costs required for training and inference. If you focus your research around work that mandates high-end hardware at scale you can lock out tons of potential incumbents

There is nothing in the deepseek paper that suggests you can't use the order of magnitude in hardware costs you saved to just train models that are ten times as large.

But there are thresholds of commercial viability in all of this. DeepSeek's technology gets you over that line with less sophisticated hardware.

There's already some pretty impressive work being done with folks using just a pair of M2 Ultras with r1 in a "home lab" context that goes way beyond what you could previously do with llama.

Yes, but the r1 model is shiny and new. I've been using it locally for a day and I'm already finding odd spots. One thing that I've seen it do a number of times is get the correct answer in the thinking tokens, then ignore it, generate thousands more tokens until it reaches it context length then give a wrong answer.

There are things, like doc strings for functions, that llama models give better results for.

…and even if you can’t, you could train models that are ten times more specific, or update them ten times more frequently.

Reply on news.ycombinator.com