Things that confuse me about the current AI market(danfrank.ca)
danfrank.ca
Things that confuse me about the current AI market
https://danfrank.ca/things-that-confuse-me-about-the-current-ai-market/
4 comments
There's really not much difference in the architecture ultimately.
The key thing is the data + compute resources. If you can afford to run a big H100 cluster, even temporarily, it's accessible.
The critical question though is: to what end is it profitable? If you're company X who's built one of these, it's obvious that another company could knock you off peak position next week. If that's the case, is there the incentive to continue ploughing money into training? For some big companies the answer is obviously yes because they want to dominate the ecosystem. But I'd question whether there's a large market for it in the medium-term.
The key thing is the data + compute resources. If you can afford to run a big H100 cluster, even temporarily, it's accessible.
The critical question though is: to what end is it profitable? If you're company X who's built one of these, it's obvious that another company could knock you off peak position next week. If that's the case, is there the incentive to continue ploughing money into training? For some big companies the answer is obviously yes because they want to dominate the ecosystem. But I'd question whether there's a large market for it in the medium-term.
ASI is the endgame where it is profitable to be in the OpenAI position, or even in the next first 20 market players capable of getting there a bit later.
But if ASI isn't achievable finally, the intelectual properties obtained in the way to it will probably be valuable, because the SOTA still works and can be re-reployed anytime in the future when the new hardware becames available and cheaper (think Cerebras stuff level). We would be in a new kind of AI winter, just waiting a couple of years till spring breaks up (cheaper, faster hardware).
Even in that winter, bigger players would still be available, think Gemini or Copilot products, getting bigger, better year after year during the winter, just as fast as the new hardware begins to be buyed/deployed. And minute by minute the market share for those bigger players will playing along with bigger reveneaus every quarter, preparing the way to full profitability in a couple of years.
Think automobile industry or oil extraction industry, going from manual work to fully machine assisted tasks, as the technology became available, from the 1900s to the 1970s. Quite a lot of years, but in the AI winter probably coming, you get even the chance to double check the countdown every now and then, just looking at what is cooking/assembling Nvidia/TSMC, Cerebras, etc.
But if ASI isn't achievable finally, the intelectual properties obtained in the way to it will probably be valuable, because the SOTA still works and can be re-reployed anytime in the future when the new hardware becames available and cheaper (think Cerebras stuff level). We would be in a new kind of AI winter, just waiting a couple of years till spring breaks up (cheaper, faster hardware).
Even in that winter, bigger players would still be available, think Gemini or Copilot products, getting bigger, better year after year during the winter, just as fast as the new hardware begins to be buyed/deployed. And minute by minute the market share for those bigger players will playing along with bigger reveneaus every quarter, preparing the way to full profitability in a couple of years.
Think automobile industry or oil extraction industry, going from manual work to fully machine assisted tasks, as the technology became available, from the 1900s to the 1970s. Quite a lot of years, but in the AI winter probably coming, you get even the chance to double check the countdown every now and then, just looking at what is cooking/assembling Nvidia/TSMC, Cerebras, etc.
Good data in good data out.
The Google dev that freaked out in that famous email that declared there is no moat was right and as your article points out, yes indeed there is no technical moat.
The moat was never technical, it was first movers.