Latest / Black Box Theory Podcast / BBT #13 Snippet | The AI Arms Race Is Costing Billions to Train
Transcript
- Thomas: I hope it's the beginning of the end for the sake of everyone who is in the workforce or trying to get in the workforce or actually wants to do something they enjoy with their life. I hope this is the beginning of a great reset. ⁓ I hope it all just comes crashing down. like I will say, I keep going back to the dot-com bubble because I've seen comparisons to that. Obviously I wasn't there. ⁓ I know some of you might think that I was there. People tell me I look old. ⁓ Thank you. Question mark. But. Malcolm: I do too. It is crazy. Thomas: Yeah, no.com bubble. Basically, that's what companies were doing. They would slap.com to their business name or say, ⁓ we have a website. We're dot com, blah, Stock will shoot up. Now we're seeing that with AI. ⁓ And like Malcolm said, this is a completely competitive space. Very competitive, it's constantly changing. You can talk a little bit more about the Claude models and things, but it's like right now, getting into that as a, you're trying to build. I hope they're really trying to figure out a way to support or do kind of like what companies like Microsoft does where you. Malcolm: Mm-hmm. Thomas: try to work with these AI companies to create offerings running on their models because if you're trying to create a model, you're... Malcolm: 39 million ain't going to cut it, Thomas: Yeah, it's over with. mean, we saw what was the thing, ⁓ not lobster, what was the thing that, open claw, yeah, stuff gets pot and popping for a week or two max, and then somebody come out with something better. Malcolm: Open, open quote. Yeah. ⁓ yeah. Yeah. I don't think people realize brother, like it's like training LLMs, like frontier LLM models, bro. These can range from 60 to a hundred million dollars, bro. Like GPT-4 is like in that range, but it's a massive power. It's a lot like across like for training costs. Like, so you have your compute infrastructure, the GPUs themselves call something crazy. The large expense with... like rental rates, they can charge anything on an hourly basis. You have your model size and data, the more parameters and with large models with more parameters and large data sets require more training time and higher cost. You have your energy consumption, which everyone knows about the four, five, six, seven. gigawatts of power that are needed, data creation, like all this stuff, bro, kind of goes into these high training calls. But I would, like I said, I'd just be curious to see how well, one, how well positioned this company is to deal with this crazy pivot, how they will rebrand themselves, how they will change, like adjust the hiring of their current kind of like talent. I'm like, bro, I just don't understand how you, like, what do you tell the guys that are in there? making the shoes now. If I'm a guy that designs our next shoes and you're me, yeah, we doing all GPU stuff now, brother. You can go ahead, pack your stuff up, bro. All of that. Thomas: Hmm. Malcolm: I don't know man. Thomas: show designers in 2026. It's not looking good. Not looking good. Malcolm: Woo! Yeah man, but yeah, it's crazy times we living in bro. SMP on a run, all time high, shoe companies trying to make LLMs bro. It's a cut.