Totally independently afaict, the Royal Institute just had thier Fellow of Mathematics give a talk which made basically the same points

Modern mathematics is too large; AIs plus Aitomated Theorem Solvers are the path forward

https://youtube.com/watch?v=oOYcPkBaotg

Indeed. LRMs are now the norm

I think this is a good perspective, it's important to recognize that things are always changing. Newer LLMs when provided enough context are much more capable than something like GPT-4 which is what I think most people were exposed to. I think a lot of people tried using ChatGPT, GPT-4 first released in March 2023 was the huge initial hype. I have no doubt that LLMs are able to assist with math successfully.

Using hyperbole like “destroy society” is unhelpful.

Well, you watch. You ain't seen nothing yet. The hyperbole will soon prove to be totally understated. Not that this is helpful though, you're right about that.

PDP-11 here 🙂

Using hyperbole like "destroy society" is unhelpful. Equally low-effort, low-precision to say "social media has destroyed society"

Arguably so, but does nothing to advance conversation nor understanding

For the record, I learned on system/360 So yeah... the Internet is new

Remember when the Internet was new and the Windows TCPIP stack was fragile AF?

Actually I'm old enough to remember the internet very much more reliable than that at that time, and before. The Windows TCP/IP stack being fragile and buggy was just a Microsoft problem. Those of us under Unix had no issues - at least nowhere near as many.

But I digress...

There's a fundamental difference between the birth of the internet and the birth of AI: the former didn't destroy society to become a mature technology.

I have no doubt AI will one day deliver on its promises. But I question who will be left in any capacity to enjoy it when it does.

LLM for code generation depends a LOT on which model you use, what you put in its context window, and what framework you are working with. We are absolutely still figuring out how to use the tools for serious software engineering and there is absolutely overhype.

Remember when the Internet was new and the Windows TCPIP stack was fragile AF?

It's like that

Incidentally, I didn't downmod you. I don't think that's fair. Those who did are probably just AI haters having a gut reaction.

Here, I'll upmod you in fact 🙂 Just because I disagree with the whole concept doesn't mean you're responsible for it 🙂

Honest question: have you published in math journals?

No, but I know enough about AI - for having to correct the seemingly good but actually deeply flawed code it generates at work - to know that if you don't triple-check what it does, it'll create unreliable shit in any field of exact science.

If a human can triple-check a math proof done by AI, chances are the human could have come up with the proof themselves. Conversely, a proof so complex that only an AI can come up with it, that's beyond a human mind's grasp - something you would very much like help from a reliable AI with - will be untrustworthy almost by definition.

That's the long version of my snarky remark.

Important point: most of your visual field is being hallucinated via interpolation between points recently sampled by your foveae. Hallucination is as fundamental to animal cognition as it is to machine cognition.

I don't do math with my fovea. The whole point of math is to escape our animal senses and manipulate universal truths.

Honest question: have you published in math journals? That is where this particular lecturer is coming from.

Important point: most of your visual field is being hallucinated via interpolation between points recently sampled by your foveae. Hallucination is as fundamental to animal cognition as it is to machine cognition.

Yeees... Let's do math with hallucinating bullshit machines.

Thanks, I hate it.

How is it a grift if it actually helps?

AI 's new grift