[implemented] Axe-Edit Fast start possible

[Chris Hurley]

I don't believe LLMs can be of any help for DSP and I don't think they will anytime soon

...

You can't get ChatGPT to come up with a new fighter jet shape that's better and that nobody found yet, you can't get chatGPT to come up with a new molecule that'll cure cancer or a way to unify physics. In the same way you can't get chatGPT to come up with novel audio algos

I saw this yesterday when I was asking questions of Grok and was reminded of this thread and the comment that "they can't do math."

While it is true that they can't generally "do math" (or make a titanium exhaust plate... or launch a missile) the whole point of "agentic AI" seems to be building tools that then can be wielded by a language model. The LLM can provide inputs to the tool which the tool can then use to produce a result. Some of the tools will be quite small but will allow the LLM to move on to the next step in the process that it predicted itself.

What gets more interesting is when the tool itself can be written using language that the LLM can handle, in which case it can build its own tools.

Here is an example screen snip just to show some of this in action. I was asking about a breakeven point on retirement calculations. I don't know whether Grok had a tool already made that could do the calculation or if it just wrote the tool itself and executed it but this gives me a glimpse and an idea of how something that is "just a predictor of words" can actual "do something". I suspect it just described the tool it needed to do the calculation in "plain english" (sorta english. It's python code). "Did the code actually work?" you might ask. I don't know because I didn't finish the exercise. I certainly have seen it make code to do calculations that matched what I did on a spreadsheet by hand.

I think that LLM will absolutely solve problems that haven't been solved yet- the question is when. The answer is probably "now" with a very large asterisk.

 

[alan.thomas]

I saw this yesterday when I was asking questions of Grok and was reminded of this thread and the comment that "they can't do math."

While it is true that they can't generally "do math" (or make a titanium exhaust plate... or launch a missile) the whole point of "agentic AI" seems to be building tools that then can be wielded by a language model. The LLM can provide inputs to the tool which the tool can then use to produce a result. Some of the tools will be quite small but will allow the LLM to move on to the next step in the process that it predicted itself.

What gets more interesting is when the tool itself can be written using language that the LLM can handle, in which case it can build its own tools.

Here is an example screen snip just to show some of this in action. I was asking about a breakeven point on retirement calculations. I don't know whether Grok had a tool already made that could do the calculation or if it just wrote the tool itself and executed it but this gives me a glimpse and an idea of how something that is "just a predictor of words" can actual "do something". I suspect it just described the tool it needed to do the calculation in "plain english" (sorta english. It's python code). "Did the code actually work?" you might ask. I don't know because I didn't finish the exercise. I certainly have seen it make code to do calculations that matched what I did on a spreadsheet by hand.

I think that LLM will absolutely solve problems that haven't been solved yet- the question is when. The answer is probably "now" with a very large asterisk.

This is a bog standard retirement calculator like there are thousands online. that could be done by students after a few months of CS101. Of course LLMs can produce code. My point is that they can't produce novel code. They can't solve unsolved math or CS problems.

I think that LLM will absolutely solve problems that haven't been solved yet- the question is when. The answer is probably "now" with a very large asterisk.

Ok, let me know what that happens.

 

[Chris Hurley]

This is a bog standard retirement calculator like there are thousands online. that could be done by students after a few months of CS101. Of course LLMs can produce code. My point is that they can't produce novel code. They can't solve unsolved math or CS problems.



Ok, let me know what that happens.

Your point that I was addressing was that they can't do math and I was showing a way that they can and do.

Your point that they can't solve unsolved math or CS problems- that may certainly be true today. We'll see where it goes. This is also an unnecessarily high bar. While they may never solve unsolvable problems. they likely will solve problems that could be solved by humans but that aren't because of capacity/resources/desire/etc..

 

[yek]

This discussion should be moved to the lounge by now.

 

[ocaptainjr2]

This discussion should be moved to the lounge by now.

Agree. The AI proponents are doing themselves no favors by derailing and hijacking thread after thread.

 

[alan.thomas]

Your point that I was addressing was that they can't do math and I was showing a way that they can and do.

Your point that they can't solve unsolved math or CS problems- that may certainly be true today. We'll see where it goes. This is also an unnecessarily high bar. While they may never solve unsolvable problems. they likely will solve problems that could be solved by humans but that aren't because of capacity/resources/desire/etc..

But it's not "doing math" it's doing the exact same word prediction it does with sentences, except using math symbols.

 

[Chris Hurley]

But it's not "doing math" it's doing the exact same word prediction it does with sentences, except using math symbols.

You're right. It isn't "doing math" but it is using word prediction and an API to allow it to feed words into a tool that can do math and then return the result back. This is a distinction but not one that matters to the consumer of the output.

The fact that it can use word prediction to write its own tools that can do math rather than just use a library of pre-made tools is also somewhat interesting. I can definitely see the value of a large library of pre-made tools that the LLM can access that are known to work correctly.

I don't think this is derailing the thread- I believe the thread began with the idea of using AI to help solve some hassly programming tasks for Axe Edit (though it turns out there is an existing feature)

I wouldn't say that I'm a proponent of AI. I think it is here whether we like it or not and it can be helpful in the right circumstances. Also an enormous buzzword. I'll do my best to let it drop from here. after Alan has his last word.