What is ChatGPT Doing?, Wolfram

Stephen Wolfram explains in his characteristically egotistical style what ChatGPT is doing and then how it could be made infinitely better with Wolfram Alpha.

Actually a combination of two of Stephen Wolfram’s two blogposts about ChatGPT,

• https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-doing-and-why-does-it-work/
• https://writings.stephenwolfram.com/2023/01/wolframalpha-as-the-way-to-bring-computational-knowledge-superpowers-to-chatgpt/

So it felt like a bit of a scam to purchase the book once I found out.

Started from first principles so it was a bit basic and covered quite a few things that I’d seen around the internet recently – not a critique, maybe just intended for a different audience. For me, the most eye-opening part about it was Wolfram applying the idea of computational irreducibility (which he coined in [[A New Kind of Science]]^?). You could never hope for a model like GPT-3 to be instantly good at things like arithmetic or things that require a non-trivial amount of thinking (without chain-of-thought prompting) because these things are hard in the sense of they need something as least as powerful as a Turing machine to be able to work them out, and the GPT architecture doesn’t contain a Turing machine. There’s no shortcut to these computations – to give the correct result you actually have to do the work. There’s an implicit outer loop in ChatGPT because it repeatedly uses what it’s already completed in the next prompt, but without changing the architecture or letting it use tools (e.g. Wolfram Alpha), it won’t be good at actually doing computations.

A very minor thing about Wolfram’s writing style in the first “half” of this “book” is that he wouldn’t “stop” putting every other “word” in “quotes” which was a bit annoying.

Or, as a poem from ChatGPT:

Oh dear book, “What is ChatGPT doing?”

Stephen Wolfram’s ego’s brewing. Infinitely better with Wolfram Alpha,

A scam to buy physical copy, a bit of a palaver.

Started with first principles, basic it may seem,

Covered things seen on the internet, not a critique, just a theme.

The most eye-opening part was computational irreducibility, GPT-3 needs a Turing machine to work things out with capability.

No shortcut to computations, hard in nature,

GPT architecture lacks the needed feature.

Without Wolfram Alpha, it won’t be good at doing,

Computations that need actual computing.

Minor writing style issues in the first half,

Quotation marks in every other word’s behalf.

But the book’s essence is insightful and grand,

Wolfram’s ideas about ChatGPT, hard to withstand.