Learning to Reason with LLMs

lol, my accountant is pretty darn expensive.

Here's an unpopular take on this:

"We had the chance to make AI decision-making auditable but are locking ourselves out of hundreds of critical applications by not exposing the chain of thought."

One of the key blockers in many customer discussions I have is that AI models are not really auditable and that automating complex processes with them (let alone debug things when "reasoning" goes awry) is difficult if not impossible unless you do multi-shot and keep track of all the intermediate outputs.

I really hope they expose the chain of thought as some sort of machine-parsable output, otherwise no real progress will have been made (many benchmarks are not really significant when you try to apply LLMs to real-life applications and use cases...)

I am cautiously optimistic. So much of building software is deciding what _should_ be built rather than the mechanics of writing code.

I you like coding because of the things it lets you build, then LLMs are exciting because you can build those things faster.

If on the other hand you enjoy the mental challenge but aren't interested in the outputs, then I think the future is less bright for you.

Personally I enjoy coding for both reasons, but I'm happy to sacrifice the enjoyment and sense of accomplishment of solving hard problems myself if it means I can achieve more 'real world' outcomes.

Another thing I'm excited about is that, as models improve, it's like having an expert tutor on hand at all times. I've always wanted an expert programmer on hand to help when I get stuck, and to critically evaluate my work and help me improve. Increasingly, now I have one.

"Safety" is a marketing technique that Sam Altman has chosen to use.

Journalists/media loved it when he said "GPT 2 might be too dangerous to release" - it got him a ton of free coverage, and made his company seem soooo cool. Harping on safety also constantly reinforces the idea that LLMs are fundamentally different from other text-prediction algorithms and almost-AGI - again, good for his wallet.

I was a bit confused when looking at the English example for Chain-Of-Thought. It seems that the prompt is a bit messed up because the whole statement is bolded but it seems that only "appetite regulation is a field of staggering complexity" part should be bolded. Also that's how it shows up in the o1-preview response when you open the Chain of thought section.