MidniteMikeWrites(1/10) I think by now everyone knows that AI (LLMs) is a bubble. Just a day after I published this post #Sora was shut down. The real question that needs to be answered is what type of bubble this is. A thread 🧵:
(2/10) The "real" bubble is in borrowing to build data centers (and buy GPUs). In my blog post I screenshot an industry marketing post eagerly writing about the premise of borrowing money to build tenantless data centers.
https://misaligned.markets/what-to-expect-ai-bubble/#2-abandoned-ai-data-center-projects
(3/10) Somewhat separately, the current crop of GenAI companies are unprofitable, but if that were it I don't think the bubble would be as big a problem. A lot of discussion I've seen about the bubble does fixate on this, though. That's where the notion of a "good" bubble comes from.
(4/10) Signs that GenAI is a "good" bubble are at best mixed. Scaling infrastructure (annotators, orchestration, more compute, tooling, workflows) is what is driving value. Costs are falling if you ignore this https://misaligned.markets/what-to-expect-ai-bubble/#what-i%E2%80%99d-expect-if-generative-ai-weren%E2%80%99t-a-bubble
(5/10) GenAI doesn't go away after the data center and AI startup bubbles, but the future looks closer to driverless cars. Basically infrastructure intensive tech that looks impressive because it hides its dependencies from view. https://misaligned.markets/ai-hype-market-fundamentalism/#the-%E2%80%9Cleft-brained%E2%80%9D-nature-of-ai-boosterism-and-consumer-capitalism
(6/10) Discussions about AI automation hinge on the unnamed premise of capital-as-labor: a perfect substitution of human for machine. But if models alone aren't the locus of LLMs' efficacy, then this idea is just wrong. It'll take labor to manage AI.
https://misaligned.markets/llm-intelligence-is-a-dark-pattern/
(7/10) Something else that the "good" bubble framing misses is that bubbles aren't neutral. There are other things society could have done with the resources that went into a bubble. Our current information landscape is incapable of considering those counterfactuals. This is what motivated my post. https://misaligned.markets/what-to-expect-ai-bubble/
(8/10) In my post I argue that governance + material constraints shape a technology. There would have been internet without the dotcom dark fiber bubble. There would have been transformers without GPTs and LLMs. https://misaligned.markets/what-to-expect-ai-bubble/#internet-counterfactuals-as-parallel-worlds
(9/10) LLMs are a non-neutral form of AI built around the AGI delusion and desire to enclose transformers which were public knowledge. User psychosis and model hallucination aren't incidental. This "good" bubble cost us a better scoped technology. https://misaligned.markets/what-to-expect-ai-bubble/#which-path-is-most-likely-and-what-have-we-lost
(10/10) While I worry about LLMs as a poorly scoped technology, I don't think they are going away. As naive as this sounds, I want responsible, thoughtful people in spaces where they can push back or emphasize LLM failure modes. It's an open question as to what that looks like, but I think it does include discouraging much of today's use. Or at least making people more critical users. https://misaligned.markets/what-to-expect-ai-bubble/#how-should-you-navigate-this-future