Important questions on AI addressed by this series.
My comments shared with the -somewhat more quiet- bluesky audience:

- Should we teach AI like children? Learning like children needs the proper cognitive architecture, which AIs lack (similar to raising a chimp as a child)
1/

AGI understood as “universal human replacement technology” is technically, socially and epistemologically rubbish

Huge “foundation” models are the antithesis of a general problem solving intelligence: in its solipsistic thinking only one perspective is pushed, while new discoveries are based on novel approaches to data

Any computational system is limited by its resources and architecture. Some (eg LeCun) say this means there is no such thing as general intelligence (GI), including in humans. In fact, GI is one capable of creating tools to overcome its own architectural limitations… like humans

So glad that here there is much less of the hyped AGI/ASI positive and negative forecasting compared to the ex-Twitter. People seem more focused on research. 👍

Almost AGI?

- Took a simple planning task in a chessboard (steps to reach a given square per piece)

- Ran it 3 times: original, mirrored positions (same result), again original

- None of:
GPT-4o
o1-mini/preview
Sonnet-3.5-new
Gemini-1.5-Pro-002

got it right in any of the trials and...

AI systems that learn from first principles are good, systems that only learn from first principles are bad.

.@swarat.bsky.social‬ do you have a different perspective on AI reasoning?

🤔We are still taking shortcuts on AI reasoning with the mindset of stepwise progress. It will be useful for many tasks but will not solve the reasoning challenge.
Ultimately, we'll need to face it based on a conceptual understanding. It isn't as difficult as a road not taken

Current AIs don't even have a criteria for deciding if it achieved its target! That's what you get with pure empirical approaches. Blind and futile reliance on right architecture being emergent

Imagine having a definition of understanding and intelligence under which we could assess LLMs. Excluding or attributing those properties is relevant only as a way for estimating current and future capabilities of these systems under the applied definition and its underlying mechanisms.

People say AI is about to reach and surpass human intelligence, measured by the way AI answer questions.

Wonder if intelligence isn’t better measured by the way you ask questions facing new problems.

How close are we there?