Great post, Keith! Barely treading water when it comes to keeping pace with the latest developments, so this was a super useful summary. Hope all is well!
Fascinating. You prompted me to get into a nuanced conversation with ChatGPT (not Pro) about the prospects for architectural convergence of the various discrete approaches towards AGI you describe. It concluded with a prediction that there are clear signs of convergence, and gave me the following formulation: AGI = LLM + Agency + Environments + Tools + Self-Improvement Loops. Is there any consensus on this amongst human commentators such as yourself?
I think there’s not consensus, too many dissenting voices, but there does seem to be convergence from the main labs on the use of LLMs as a tool in the engineering of what replaces them. A bit like the humans as the tools that built the LLMs maybe ;-)
Personally, I think prob LLM architectures get left behind by models that can interact directly with world - Deepseek’s OCR shows how vision models can learn text, Apple and Meta’s work on computer vision making rapid, largely unheralded progress. Tesla in the forefront - though I confess that as I write this I realise I haven’t been paying enough detailed attention to this side of things.
Great post, Keith! Barely treading water when it comes to keeping pace with the latest developments, so this was a super useful summary. Hope all is well!
All well John - and I think no one can keep up, me as much as anyone, so perhaps don’t feel too bad! Hope you’re doing well too.
Fascinating. You prompted me to get into a nuanced conversation with ChatGPT (not Pro) about the prospects for architectural convergence of the various discrete approaches towards AGI you describe. It concluded with a prediction that there are clear signs of convergence, and gave me the following formulation: AGI = LLM + Agency + Environments + Tools + Self-Improvement Loops. Is there any consensus on this amongst human commentators such as yourself?
I think there’s not consensus, too many dissenting voices, but there does seem to be convergence from the main labs on the use of LLMs as a tool in the engineering of what replaces them. A bit like the humans as the tools that built the LLMs maybe ;-)
Personally, I think prob LLM architectures get left behind by models that can interact directly with world - Deepseek’s OCR shows how vision models can learn text, Apple and Meta’s work on computer vision making rapid, largely unheralded progress. Tesla in the forefront - though I confess that as I write this I realise I haven’t been paying enough detailed attention to this side of things.