10 Comments
Mar 13Liked by Rohit Krishnan

Utterly sensible and absolutely the stance more of us should be taking.

And yet, Heidegger also said that "language is the house of being". If this is true, then entire patterns of being are latent in language. Why couldn't these patterns be training data for the same statistical techniques to extract and reify as some ground of LLM "being"?

This prevents me from slamming the door entirely on generality.

Expand full comment
author

It might be, and indeed is somewhat, the case in that LLMs clearly have learnt statistical regularities on all sorts of things while learning to speak. Question's more about the fact that if it's always one giant forward pass over the entire network for every token that has limitations.

Expand full comment
Apr 2Liked by Rohit Krishnan

Love the way you put it "When Descartes said “I think, therefore I am”, it didn't automatically mean the reverse too, where we could point to another thing and say “you thought, therefore you are”.

I would like to push back on your conclusion 'LLMs have given us the world's knowledge already, it's unfair to expect it to also develop a soul.' LLM's have given us a way to tap into the knowledge that EXISTS ALREADY; I'm not sure they are inspired or have intuition to create New Knowledge. Like Einstein (or any other discoverers) did when he proposed the TOR.

Expand full comment

Love the way you put it "When Descartes said “I think, therefore I am”, it didn't automatically mean the reverse too, where we could point to another thing and say “you thought, therefore you are”.

I would like to push back on your conclusion 'LLMs have given us the world's knowledge already, it's unfair to expect it to also develop a soul.' LLM's have given us a way to tap into the knowledge that EXISTS ALREADY; I'm not sure they are inspired or have intuition to create New Knowledge. Like Einstein (or any other discoverers) did when he proposed the TOR.

Expand full comment
author

I think that's a fair point, though to what extent interpolation between existing knowledge creates some new knowledge isn't a trivial question to answer.

Expand full comment
Apr 2Liked by Rohit Krishnan

Love the conclusion when Descartes wrote you thought therefore you are

Expand full comment
Apr 2Liked by Rohit Krishnan

It might be, and indeed is somewhat, the case in that LLMs clearly have learnt statistical regularities on all sorts of things while learning to speak. Question's more about the fact that if it's always one giant forward pass over the entire network for every token that has limitations.

Expand full comment
Apr 2Liked by Rohit Krishnan

This is incredible

Expand full comment
author

Thank you!

Expand full comment

The emergence of Claude-3, a Large Language Model (LLM) by Anthropic, marks a significant advancement in artificial intelligence. Claude-3 surpasses its predecessors, including GPT-4, and boasts capabilities that exceed the average human IQ. It can perform tasks ranging from economic modeling to scientific understanding and code writing, demonstrating its remarkable ability to process complex information.

However, Claude-3's performance also reveals its limitations. Despite its impressive feats, it struggles with mundane tasks like solving crosswords or mazes, highlighting a disparity between its abilities and human cognition.

The discussion delves into the nature of intelligence, questioning whether Claude-3's capabilities align with human understanding. It argues that LLMs possess a specialized form of intelligence derived from vast data ingestion, which differs from human cognition. This "Theory of Special Intelligence" suggests that LLMs excel at connecting insights across domains but lack the generalization abilities inherent in human intelligence.

Furthermore, the discourse explores the multifaceted nature of intelligence, drawing parallels to diverse forms of animal intelligence and Howard Gardner's theory of multiple intelligences. It critiques traditional intelligence tests like IQ tests, which may not adequately capture the complexity of intelligence, especially at higher levels.

The narrative also contemplates the possibility of unconscious allusions and insights generated by LLMs, reflecting humanity's collective knowledge. Despite lacking consciousness or a soul, LLMs offer valuable insights derived from the vast information they process.

In summary, Claude-3 represents a significant milestone in AI development, challenging conventional notions of intelligence and prompting reflection on the nature of cognition. While LLMs possess remarkable abilities, they operate differently from human cognition, emphasizing the need for nuanced understanding and continued exploration in AI research.

Expand full comment