Really engaging overview of the potential AI trajectory! I largely agree with the emphasis on RL's resurgence and the rapid advancements in areas like coding, the timeline you've sketched out feels plausible in many respects.
My main thought is that Google DeepMind might be a bit underestimated in the 2025 timeframe. While you rightly highlight their later monumental achievements in maths and physics (the Millennium Prize, unifying physics frameworks), the initial portrayal of Gemini-3 as primarily a 'culmination of pretraining scaling' perhaps downplays their concurrent deep work on unique reasoning architectures, especially those 'not reliant on external CoT' as you mention. Their historic strengths and current focus on foundational reasoning, even if it doesn't immediately manifest as the same type of agentic CoT RL product as others, could yield significant, perhaps less immediately flashy but ultimately profound, advantages. Although you do acknowledge GDM is making progress on these distinct paths, which I believe could be even more pivotal than depicted in the earlier stages of this projected timeline.
That said, your points about different labs developing distinct specializations (Anthropic's focus on quality/safety, OpenAI's product drive, the niches of xAI/DeepSeek) and the overall challenges and nuances of scaling RL are very well articulated and insightful.
With respect to roll-out and consumer v business. Stateside Anthropic may be doing well with business, but in the EU the landscape is much different. Google have had some major business wins, eg CMA CGM as have OpenAI. So I do not see OpenAI as consumer only over here... even in the US OpenAI has stellar wins with business, Morgan Stanley, John Deere, etc.
Nice
Really engaging overview of the potential AI trajectory! I largely agree with the emphasis on RL's resurgence and the rapid advancements in areas like coding, the timeline you've sketched out feels plausible in many respects.
My main thought is that Google DeepMind might be a bit underestimated in the 2025 timeframe. While you rightly highlight their later monumental achievements in maths and physics (the Millennium Prize, unifying physics frameworks), the initial portrayal of Gemini-3 as primarily a 'culmination of pretraining scaling' perhaps downplays their concurrent deep work on unique reasoning architectures, especially those 'not reliant on external CoT' as you mention. Their historic strengths and current focus on foundational reasoning, even if it doesn't immediately manifest as the same type of agentic CoT RL product as others, could yield significant, perhaps less immediately flashy but ultimately profound, advantages. Although you do acknowledge GDM is making progress on these distinct paths, which I believe could be even more pivotal than depicted in the earlier stages of this projected timeline.
That said, your points about different labs developing distinct specializations (Anthropic's focus on quality/safety, OpenAI's product drive, the niches of xAI/DeepSeek) and the overall challenges and nuances of scaling RL are very well articulated and insightful.
With respect to roll-out and consumer v business. Stateside Anthropic may be doing well with business, but in the EU the landscape is much different. Google have had some major business wins, eg CMA CGM as have OpenAI. So I do not see OpenAI as consumer only over here... even in the US OpenAI has stellar wins with business, Morgan Stanley, John Deere, etc.
A thought-provoking piece!
" except for China gradually experimenting with increasingly more AI in their censorship bureaucracy”.
Unlike us, China has no covert censorship. Beijing publishes the rules and 80% of folks say they're fair and reasonable.
Over 80% say their government and media tell the truth, and they do. So there's not much to censor, really.
Still can't believe the times we're living in.
Excellent article:)