Comment by chasd00
14 days ago
Don't forget, the bubble talk is more around "AGI is right around the corner, don't get left out!" and not so much the application of an LLM to business process flows. There's definitely value in some use cases to using an LLM but I don't see the massive bets on AGI paying off in timelines Wall St. is comfortable with. Further, my feeling is the AGI promise and investment is what is keeping some of the model providers in business. From what I understand, inference cost per user doesn't scale the way a webserver does. With an LLM, more users means significantly more infrastructure cost and Google, Meta, MSoft can afford to run/train models at a loss because they get revenue elsewhere but not everyone is in that boat.
Yes the problem is not whether LLM have value, but whether they will add enough value in the next two three years to pay for the hundred of billions of trillions needed to match the expectations. OpenAI has to find more than a trillion by the end of the decade, project Stargate still has 500 billions to find, etc. Also add fast depreciation on everything invested. The ZIRP playbook of building a monopoly for years before turning a profit cannot work here, models are commoditizing fast so no moat there either, and capabilities as a function of compute have plateaued.