Nice read, Dean. As an Indian I am proud of the success of local model builders like Sarvam and at the same time worried that we are too busy patting ourselves on the back. Yes, we have a demographic dividend and unless we are going to plug ourselves Matrix-like and become meat computers, we will need the same elements (compute, capital, and data) to build our sovereign Frontier models.
Small models may be enough to meet the bulk of task towards the head of distribution but all our aspirations to solve healthcare etc are long tail problems and require AGI-ish frontier models. Even if we concede that the models are hitting a wall right now, and assume real money in AI is in applying small models to solve the bulk head problems, how hard is it for OpenAI or Anthropic to quantize their SOTA model or leverage their compute to drown the market and blow these local AI companies out of the market.
I loved the optimism coming out of the summit but am scared of the underlying delusional assumptions.
Well put. The nature of India is such that every individual bit of complexity that is in front, the "never-ending need to readjust and remeasure one's surroundings", this is what I wrote of as a series of bilateral negotiations, this also acts as an almost coasean brake in some instance, and while I'm quite hopeful too that we will figure this out, it is a hard problem.
My impression is that the frustration and pessimism regarding Africa's lackluster progress along the formerly common developmental pattern will hit many formerly success stories of the "global south" and eventually creep up to the fully developed laggards too, leaving a few superpower competitors, until the "all bets are off " stage reaches them too.
For all their colocated compute, Grok still seems like a default outcome for Middle Powers that do chase the frontier: a good model on benchmarks but outside the conversation for the leading use cases of AI. I don't believe the Bitter Lesson is done, but their experience in our present place on the curve does suggest that it's a necessary but not sufficient condition for competing at the frontier.
Nice read, Dean. As an Indian I am proud of the success of local model builders like Sarvam and at the same time worried that we are too busy patting ourselves on the back. Yes, we have a demographic dividend and unless we are going to plug ourselves Matrix-like and become meat computers, we will need the same elements (compute, capital, and data) to build our sovereign Frontier models.
Small models may be enough to meet the bulk of task towards the head of distribution but all our aspirations to solve healthcare etc are long tail problems and require AGI-ish frontier models. Even if we concede that the models are hitting a wall right now, and assume real money in AI is in applying small models to solve the bulk head problems, how hard is it for OpenAI or Anthropic to quantize their SOTA model or leverage their compute to drown the market and blow these local AI companies out of the market.
I loved the optimism coming out of the summit but am scared of the underlying delusional assumptions.
Well put. The nature of India is such that every individual bit of complexity that is in front, the "never-ending need to readjust and remeasure one's surroundings", this is what I wrote of as a series of bilateral negotiations, this also acts as an almost coasean brake in some instance, and while I'm quite hopeful too that we will figure this out, it is a hard problem.
https://www.strangeloopcanon.com/p/life-in-india-is-a-series-of-bilateral
My impression is that the frustration and pessimism regarding Africa's lackluster progress along the formerly common developmental pattern will hit many formerly success stories of the "global south" and eventually creep up to the fully developed laggards too, leaving a few superpower competitors, until the "all bets are off " stage reaches them too.
Dean, I just love your writing. It reminds me of Mircea Cărtărescu 😝
For all their colocated compute, Grok still seems like a default outcome for Middle Powers that do chase the frontier: a good model on benchmarks but outside the conversation for the leading use cases of AI. I don't believe the Bitter Lesson is done, but their experience in our present place on the curve does suggest that it's a necessary but not sufficient condition for competing at the frontier.