Greg Brockman interview [video]

https://medium.com/@vishalmisra/shannon-got-ai-this-far-kolm...

This article explains what's missing in terms of two kinds of complexity that oppose: Shannon complexity vs. Kolmogorov complexity.

It introduces the opposition by an example of driving the value of pi as decimal number, which has no pattern and high complexity, and a formula for deriving pi that does have a pattern with low complexity, then observing that mind can work from the patternless high-complexity back to the patterned low-complexity without prior examples, while AI can't.

LLMs encode and retrieve patterns in the training data, and doing so can connect data to the terminology of known principle, but mind can observe inconsistencies in data and to reason from first principles to resolve the inconsistency.

The distinction between these two modes can seem blurry as AI can traverse the patterns of the known in ways that are extraordinarily revealing, but it's not structured to reason about the unknown.

Inference is not sufficient for reason.

For example, a conventional algorithm can search for patterns in text at a scale many orders of magnitude beyond a mind's capacity, and this can be very revealing, but to do so this algorithm need not read the text with comprehension.

Regarding the question: can genAI be enhanced to reason? The answer is assumed to be "no", due to the categorical opposition of the two kinds of complexity and the lack of understanding of structures within genAI to handle the reasoning.

Read the article, which includes other examples including a jump from Newtonian to Einstein physics in the history of astronomy, and a noodling on how to talk about the edge of the unknowable in AI.

I think the Chinese labs have a fundamentally different viewpoint: they’re building infrastructure, and looking at it more like how a US corporation might like having some of its employees making core contributions to compiler ecosystems like LLVM/clang and so on. The payoff is down the road, partially reputational, but also having a great compiler is good for everyone in the computational business world. The rentier-finance capitalist instead wants to privatize the compiler and extract rents for access.

The thing about infrastructure is this: you don’t get a direct financial return on investment in infrastructure (think roads, which make other economic activity possible) unless you have some ridiculously corrupt system controlled by rent-collectors (which is how the US electricity grid and fiber optic backbone works). That’s all the major US LLM providers are doing: trying to collect rents on systems that were built using the global human knowledge base as inputs.

At the very least OpenAI should be releasing their older models on a steady timetable. Sure it might reduce some revenue streams but it would be good for their reputation.

The latest efforts like agents are clearly showing the limitations and are nowhere near AGI.

We’ve now reached the buzzwords and bullshit stage of the bubble where they cast around for problems shaped like the solution.