What is the cost of the most expensive individual foundation models in this world? I think going in the $50-$500 billion territory requires notable improvement at $3-$20 billion scale, possibly agentic behavior for longer term tasks with novel subproblems, otherwise scaling investment stops and a prediction along the lines in this post makes sense (assuming RL/search breakthroughs for reasoning are somewhat unlikely).
Meaning 30% of >$5bn, that is 70% of <$5bn? What needs to happen for investment to stay this low through 2029, given that I’m guessing there are plans at present for $1bn-$3bn runs, possibly this year with GPT-5 and then Claude and Gemini? When say OpenAI has a valuation of $80bn, it makes sense to at some point invest some percentage of that in improving the backbone of their business and not losing the market niche to competitors. (ITER and ISS are in $20bn-$150bn range, so a $50bn model is feasible even without a demonstrable commercial motivation, but possibly not by 2029 yet.)
What is the cost of the most expensive individual foundation models in this world? I think going in the $50-$500 billion territory requires notable improvement at $3-$20 billion scale, possibly agentic behavior for longer term tasks with novel subproblems, otherwise scaling investment stops and a prediction along the lines in this post makes sense (assuming RL/search breakthroughs for reasoning are somewhat unlikely).
edit for clarity
I am not very good at sizes. But I guess that it’s gonna keep increasing in price, so yeah, maybe >$5bn (30%).
Meaning 30% of >$5bn, that is 70% of <$5bn? What needs to happen for investment to stay this low through 2029, given that I’m guessing there are plans at present for $1bn-$3bn runs, possibly this year with GPT-5 and then Claude and Gemini? When say OpenAI has a valuation of $80bn, it makes sense to at some point invest some percentage of that in improving the backbone of their business and not losing the market niche to competitors. (ITER and ISS are in $20bn-$150bn range, so a $50bn model is feasible even without a demonstrable commercial motivation, but possibly not by 2029 yet.)
To be clear, your post is assuming TAI is still far away? AI is just what it is now but better?