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.
Jump in the discussion.
No email address required.
Notes -
I'll caveat that it's not clear that a sell-off solves, rather than slows, AI risk; a competitor buying all this equipment at fire sales price might be less interested in building a machine god, but they'll still be interested in building a smarter system and have a lot of spare inference or training equipment to run.
Beyond that, it depends very heavily on what the end situation you expect.
The maximally-bullish case is some form of captured recursive self-improvement producing a massive and deep moat. Claude Fable 7.2 or Kimi 5 or (more likely) some internal specialized model produces a 5x efficiency boost, which makes throwing more processing power at the question economical, which unlocks another 5x efficiency boost, so on. At the more science fiction side of things, this could be new chips that combine FPGA-like re-programmability with ASIC efficiency; at the more plausible you've got software hacks, latency reduction, and conceptual refinements.
The good news from an X-risk perspective is that almost everything going this direction so far has been slow and in hardware, which put some limits on speed, since no matter how good an AI-designed chip or network layout is, logistics takes years. The bad news is that there have been some individual human-driven efforts already in software and model design (changes to KV architecture, MTP/DFlash) already, and it's the sort of space that I'd naively expect smart-enough LLMs to 'beat' humans by brute force. If it can pop off quickly, it will do so in months rather than years, and it will be a big surprise to almost everyone else.
That's not a massive moat, since eventually information (and models themselves) leak or a competitor open-sources them. But five or ten years of selling superintelligence at a tenth the cost of what your competitors are selling 'naive intern' can cover a hell of a lot of debt, as would being able to train smarter models for a hundredth of the price of your competitors. And at the really optimistic (from a business) or pessimistic (from an x-risk) cases, you stop being in a situation where 'revenue' or even 'competitors' makes sense as a question.
The more moderate bullish case is taking existing models and refinements to regulated fields, and getting a steep enough moat that the competitors can't step in easily. Higher education's the obvious option, if not likely huge and fast enough, between the external political pressures and underlying tensions in the business models for major colleges, but there's a lot of space in medicine and compliance that are heavily licensed in ways that could make it very hard for merely-good models to be used. Even some weird cases with general-purpose robotics could end up in a state where use is generally valuable, the liability risk of using sub-cutting edge models is extreme, and thus only the nerds can get business from the big companies.
Optimistically, this could come with a massive demand-side increase -- personalized instruction making everyone able to become experts in a field they find interesting, customized entertainment, productive hobbyist work. The middle case is Nothing Ever Changes despite it all, where we end up with gambling addicts and entertainment dollars redirecting at a 1:1 ratio, or some close approximation of it. Pessimistically, it could be the NSA wanting bulk data processing capabilities, and then 'selling the business' stops being an option: when state actors are a big enough portion of your financial model, your model stops being about finances.
The weakly-bearish case is just selling inference, well. The massive expenditures for training buildouts and heavy overpowered models eventually instead allow things like global prioritization and redirection based on load and spot energy costs, at a variety of different demand levels and latency limits.
((The actually-bearish case is a financial product one, where regardless of whether the hyperscalers are making profits, the book value of their assets drops catastrophically, either because of depreciation scaling or cost of servicing debts or financing existing products makes them sell out. But that's a much more complicated case than it sounds at first.))
More options
Context Copy link