3 ms·

I gotta be honest, I looked at chat gpt, thought just doing LLM inference is not that hard, someone will produce an ASIC that destroys nvidia's margins here. Y

I gotta be honest, I looked at chat gpt, thought just doing LLM inference is not that hard, someone will produce an ASIC that destroys nvidia's margins here. You could honestly make a hedge fund by listening to me pontificate things and then betting opposite of me. I have been bearish on MS, twitter, dropbox, bitcoin, the list goes on. It's admirable how I am so consistently wrong.

Workaccount2 3y ago
I guess to be fair, I was also bearish on Nvidia when it hit $250. So probably would have been dealing with pain since I likely would have sold after those 50% gains. Same thing with BTC. I tried to buy 20,000 in 2010. Was too much hassle back then so I gave up. But even if I went through with it. I would have sold the moment I turned $20 in $200. Much less hold past $500 or $1000.
- c22 3y ago
  I feel this. I sold the bitcoin I had mined in 2010 for around $800 in 2012 and thought I was an absolute genius.
tanseydavid 3y ago
>> someone will produce an ASIC that destroys nvidia's margins here I am no expert but it sure seems like Groq might be on to something with their LPU (language processing unit).
Sohcahtoa82 3y ago
> someone will produce an ASIC that destroys nvidia's margins here. From what I understand, if someone made an ASIC for LLM inference, such a device would effectively have the specific model locked in. You wouldn't be able to update the ASIC for newer versions of the model that alter the architecture. At best, you might be able to make one that allows you to update the weights. I'm also not sure how cheaply you could produce such an ASIC. One that runs a 7B parameter model probably wouldn't be too expensive, but something that could rival GPT-4 would be.
- chii 3y ago
  i imagine that ASIC of such types would be analogue, rather than digital. Or use a different substrate than silicon - may be light/laser based? It'd have to be quite revolutionary, rather than evolutionary.

Workaccount2 3y ago

I guess to be fair, I was also bearish on Nvidia when it hit $250. So probably would have been dealing with pain since I likely would have sold after those 50% gains. Same thing with BTC. I tried to buy 20,000 in 2010. Was too much hassle back then so I gave up. But even if I went through with it. I would have sold the moment I turned $20 in $200. Much less hold past $500 or $1000.

c22 3y ago
I feel this. I sold the bitcoin I had mined in 2010 for around $800 in 2012 and thought I was an absolute genius.

tanseydavid 3y ago

>> someone will produce an ASIC that destroys nvidia's margins here I am no expert but it sure seems like Groq might be on to something with their LPU (language processing unit).

Sohcahtoa82 3y ago

> someone will produce an ASIC that destroys nvidia's margins here. From what I understand, if someone made an ASIC for LLM inference, such a device would effectively have the specific model locked in. You wouldn't be able to update the ASIC for newer versions of the model that alter the architecture. At best, you might be able to make one that allows you to update the weights. I'm also not sure how cheaply you could produce such an ASIC. One that runs a 7B parameter model probably wouldn't be too expensive, but something that could rival GPT-4 would be.

chii 3y ago
i imagine that ASIC of such types would be analogue, rather than digital. Or use a different substrate than silicon - may be light/laser based? It'd have to be quite revolutionary, rather than evolutionary.