site banner

Small-Scale Question Sunday for January 26, 2025

Do you have a dumb question that you're kind of embarrassed to ask in the main thread? Is there something you're just not sure about?

This is your opportunity to ask questions. No question too simple or too silly.

Culture war topics are accepted, and proposals for a better intro post are appreciated.

2
Jump in the discussion.

No email address required.

Thoughts about what's going on in markets right now? Nvidia and AVGO (the big AI compute production companies) just took a 17% dump.

People are saying that this is due to R1 proving that less compute is needed for good results, so there is less need for more capex (but OpenAI announced 500 billion of capex over the next four years, just a few days ago). I don't understand this though, if you can do good work with little compute, you should be able to do great work with more compute. If anything this should mean an acceleration towards singularity territory, timelines shortening. The hyperscalers will be forced to up their game and fight back, accelerate their releases. Deepseek is just at this moment down due to too high demand or what they call 'malicious activity'. As I write this I opened up a music video on youtube and found it was all AI for the visuals. Everyone needs more compute.

Also why is Facebook up when their open-source models have been so thoroughly trounced? Google and OpenAI/Microsoft have stuff comparable to R1.

I heard an alternate thesis that this was due to japan raising interest rates and upsetting the carry trades that support US tech valuations.

I hold both of these hard-hit companies so perhaps I'm coping a bit. I guess I've long believed that markets are irrational in the short term, so it doesn't matter if stocks go back to where they were on October 1st. It's all karmic consequences for buying Marvell (which also makes hardware for AI) just a few days ago. I heard that little voice inside telling me to wait for a red day and ignored it. It's the same old story: https://youtube.com/watch?v=61Q6wWu5ziY

Up until Deepseek the conventional wisdom was that the bottleneck of the AI revolution will be Nvidia. So they would command prices and margins. After deepseek it was less certain. Due to the way the human mind works movements arounds probabilities around 0 and 1 - even small movements in probabilities away from 0 and 1 could trigger disproportional reactions. Nvidia will produce a shitload of silicone and sell it. But the reduced computational demands means that there is price points and applications at which ati/intel or some asics could become viable.

I read the other day that China has been obtaining 4090 chips through clandestine means, somehow stealing them from graphics card boards and then returning them so they end up in the supply chain GPU-less. Is it evident exactly what silicon Deepseek is using, or is it a possibility that the underlying silicon is mostly nvidia chips that have been obtained despite export restrictions?

I think it is too complicated. And if done on large enough scale people will notice. And if done that way it will be for Chinese military/Intelligence use.

Also why is Facebook up when their open-source models have been so thoroughly trounced? Google and OpenAI/Microsoft have stuff comparable to R1.

Because Meta's AI strategy was trying to prevent the other big players from building a moat, and Deepseek seems to be doing their work for them.