The AI Reasoning Lie

amen to “Logic is about the relationships between the ideas, not the order of presentation.

At their core LLMs compress data into a latent space where tokens are spatially related to each other. COT is an attempt to synthetically “re-tune” weights based on a specific problem (even though they are actually frozen). What we need is a model that can “think” internally and retune within the LLM’s latent space. I think this can be done and would likely qualify as reasoning

Labeling commutativity as "the crucial property" overlooks these other essential aspects. Logical reasoning is a multifaceted discipline where various properties work together to form robust systems of deduction. While commutativity is certainly useful—especially for simplifying and rearranging expressions—it is just one of many properties that contribute to effective reasoning.

This is why I am very excited by diffusion language models, those are less biased from the order of information.

17:30 I'm a bit annoyed we weren't given the real answer to the riddle. From my deductions the answer is "uncertain". The reason being none of the premises talk about how sleepy other mammals are. And none of the premises give a constraint on Jumbo being a different mammal than an elephant. So Jumbo could be both a sleepy and a non-sleepy mammal.

That kind of leakage is pernicious - we've known about those kinds of security risks for 40 years, yet modern processors continue to leak data. Now it's LLMs. Running locally is going to be the only safe choice. (Consider also the risks of cache poisoning)

What a great and indepth content. It's quite rare to see such video relating AI that's well researched. Most usually either hyping AI or dooming AI. Liked and subbed!

I don't often like but I did this time. I'm a huge fan of LLMs but this is a good perspective to have. Please make more like this!

Regarding your mention of MinionS, I went back to watch that video, and it wasn't clear that there's guaranteed obfuscation (or removal) of sensitive data before the remote LM is called. I love the concept though. The global caching issue cannot be understated: the fact that it exists at all demonstrates the privacy is not a commercial LLM priority right now. Finally, I think we're getting to a point where LMs may be considered tools for other LMs.

Good to know it's "new" with the thumbnail