Just wait till LLMs come prebuilt with world models. That will be the underpinnings, it’s innate cognitive structure, that will help it understand language - this time.

I hope so too. But, sanity seems to be scarce these days.

I dislike how we still can’t edit posts here - read as “Could be reading …”

Could you be reading your tweets instead, that would keep him busy full time.

For digital only products - yes. Unfortunately, it’s going to manifest itself in physical products as well. These models are still under human supervision when used in safety critical products. Most physical products are self use - that’s where any breakdown in reliability will be widely felt.

This shift will run downstream too. Once “works some of the time” is acceptable for models, it will seep into product pipelines. From marketing gloss to development then eventually in the supply chain. Reliability becomes optional not just for the models but in the end product itself. It’s coming.

LLM’s internal monologue:

Real world mode: “Bro … this decision tree looks like a whole forest.”

Eval mode: “Ahh, a single pruned leaf … perfect boys - I got this.”

Some prefer spoken words over reading either by choice or sometimes handicap - so, I prefer both exist.

Just read the transcript. Multiple ways to get it, here’s a couple:

www.wikihow.com/Get-Transcri...

While most of it is based on solving for his vivid imagination regarding AGI, there are some interesting tidbits sprinkled around. Like - while RL optimizes for the reward function, it also distorts the latent space such that the global performance actually degrades - was probably the best one.

Because whenever she’s reporting “news” - it’s debatable.

The darn thing is a tool that’s easily lost in the matrix woods without handholding from an expert in the domain at hand. The only reason for these comparisons, is so these companies can claim - look our “AI” can replace those 5 PhDs you were going to hire, for only $20k a month.