Really interesting to visit this post in 2025. Many of your early points about multimodal models, hype cycles, and chip competition turned out to be stunningly accurate. I think that the trajectory around diplomacy and bureaucracy architectures feels a little less central compared to how things actually unfolded, with agent frameworks, RAG pipelines, etc, taking the spotlight. I still feel that the propaganda concerns resonate; the information environment continually feels warped by AI-assisted persuasion.
If we were to extend this trajectory today, do you envision future AI progress having to do more with scaling architecture (agents, tool use) or more as a story of governance and incentives (who is deploying + regulating these systems)? Or perhaps a combination of both?
Thanks! I basically agree with your assessment, though I think “agent frameworks, RAG pipelines, etc.” is an example of what I was talking about with bureaucracy architectures. I just chose the unfortunate clumsy name “bureaucracy” which never took off.
AI-2027.com is reasonably close to what extending this trajectory would look like. These days I expect things to take a little longer, like 2029 instead of 2027.
Really interesting to visit this post in 2025. Many of your early points about multimodal models, hype cycles, and chip competition turned out to be stunningly accurate. I think that the trajectory around diplomacy and bureaucracy architectures feels a little less central compared to how things actually unfolded, with agent frameworks, RAG pipelines, etc, taking the spotlight. I still feel that the propaganda concerns resonate; the information environment continually feels warped by AI-assisted persuasion.
If we were to extend this trajectory today, do you envision future AI progress having to do more with scaling architecture (agents, tool use) or more as a story of governance and incentives (who is deploying + regulating these systems)? Or perhaps a combination of both?
Thanks! I basically agree with your assessment, though I think “agent frameworks, RAG pipelines, etc.” is an example of what I was talking about with bureaucracy architectures. I just chose the unfortunate clumsy name “bureaucracy” which never took off.
AI-2027.com is reasonably close to what extending this trajectory would look like. These days I expect things to take a little longer, like 2029 instead of 2027.