It won't be about tokens in the long run, and I think the labs know that. It will be about long-term memory and user relationships, and who gets to own that data. The disruption and upheaval has barely even started, IMHO.
So it still gets back to the question of who needs or wants the truly frontier models, because the memory and user relationships will only be about that small category?
Maybe they win market share for the large category in some other way?
> The government pays for surveillance? Companies pay for customer service replacements and target advertisements?
Yes definitely. All together that's probably within striking distance of a trillion dollars in annual revenue. Just TV subscriptions + cell phone plans in the U.S. alone is about 400 billion (leaving aside any other country).
Most large, mature tech companies are at 5-10x revenue to valuation ratios (and 10-20x EBIDTA to valuation ratios).
Of course whether you get there is a separate question, but any company that fulfilled what you're asking would pretty easily justify multi-trillion dollar valuations.
That $400B for video and phones though. That's like a $1k average for every adult and child in the country. It's twice what I spend, and I'm not on a family plan.
Video is a bit trickier to nail down since there isn't a single authoritative group putting out numbers, but most whitepapers converge around $150-200 billion (see e.g. https://www.prnewswire.com/news-releases/total-subscription-...).
Thank you for the detailed links though!
That's significantly underestimating the potential. Companies typically spend $200k/year for developers (salary, benefits, taxes, equipment, office space...). If AI can make them even 25% more productive, that would justify $50k/year in tokens.
Regulation wouldn't solve the problem, but many frightened people would be in favor of it anyway.
Similarly, I think it's a good bet that in the relatively near future, pretty much all consumer software, down to piddling little phone games, will be built with the presumption of some LLM/inference layer.
One very simple reason for this is that AI could dramatically change people's expectations for UI/UX across the board. I can see a future where the expectation is to have some natural language interface at minimum, or have UI driven a great deal by inference under the hood.
In which case, applications without these features become about as appealing as command line tools to the general public.
Personal assistants dealing with your email, filling out paperwork, and booking reservations don't use a lot of tokens.
If every software developer paid $200/month, every academic paid what they pay for Matlab or Mathematica licenses, every YouTuber paid for whatever video processing, and every normal person paid a subscription comparable to their TV subscriptions and cell phone plans - does that justify the valuations?
The government pays for surveillance? Companies pay for customer service replacements and target advertisements?
It seems like there are a few small categories that use a lot of tokens, and a large category that doesn't need many at all.