Get Everything Done

I have to admit to having zero luck with getting anything useful out of ChatGPT or any other LLM for pretty much anything. I haven't tried what you suggested, and I could see how it might be quite nice. However, given that ChatGPT isn't useful in my workflow for pretty much anything, I haven't felt much of a need to use it in my journaling workflow either.

Also, what can I say, I like my pen and paper a lot. ;-)

I just tried it today, actually, since your query prompted me to give it another go.

The responses I got to a few of my classic quandaries were better, but they still weren't helpful. For example, I used their programming Co-Pilot to try to write some code, and it *almost* got something that was accurate for the initial prompt, but really quite useless to me, so I refined the prompt to ask it to refine the code further towards what I needed, but the more I pushed it to adhere to the restrictions I needed for the code to be correct, the less correct the code got, until the code was just completely wrong. And it actually couldn't tell that what it was giving me was fundamentally wrong, and unless I already knew the answer to the question, I might not know that either.

I also asked the general 4o model to give me some information that I hadn't found much information on, to see if it could give me anything I didn't already know. However, it didn't help much, only giving me stuff that I had already found out rather trivially, and that I had already posted online and summarized (possibly enabling GPT to pick it up). Attempting to get it to give me things like sources lead to exactly the same sort of thing I had already found, with no additional new information.

And now my limit is up for GPT 4o.

One of the biggest issues I have with using GPT is that if I'm trying to learn something, I have no way of knowing whether the information I'm getting back from GPT is correct, and it always seems reluctant to "prove itself correct." And when I ask it to cite its sources, it changes its answers and doesn't provide sources for its original claims. It's like talking with a particularly obtuse but well-read high schooler who doesn't understand how to string logical ideas together.

I think the GPT models seem to be good when I want to craft highly specific queries as a sophisticated search engine, but when I actually want to *know* something, or if I want to have confidence in something, or when I need to get real answers that are obscure and hard to find, then I find GPT less helpful, and that's the whole point of a "more powerful language tool" to me. I can't trust it to help me with my programming (it always makes things worse).

I think maybe it might be interesting to try to use it to make slides for me. If I can write out the things I want to say (in a notebook), and I can just take a picture of the pages of that notebook and have it produce really high quality slides from that, then I could get some value there. But I am not willing to pay for the model yet to get there, as I'd need to see it work first.

Given that my profession is programming, the thing that scares me the most about these models is that they go by statistical likelihood and pattern matching. However, in my experience, that is sort of like appealing to the majority or appealing to the mean. In both cases, you get what you expect, but the results tend to be mediocre.

I'm especially concerned about this for optimization and programming type tasks, because the responses I get tend to be very bad at actually finding a "best" answer, and instead find the "most common" answer. That means that if I want to do some sort of programming, my code is always going to be driven towards "less good, average" code the more I rely on the tooling, which strikes me as the wrong direction, especially given that I think I can rightly say that I'm an expert whose code already punches above that weight.

As a springboard into searching for information, I think it can help, but my issue there that I'm thinking about right now is its tendency to focus in on specific responses. The good thing about a web search is that I get a really good picture and overview of the overall field when I'm looking for an answer. I think that overview is immensely helpful, and the more precisely you get an answer, the less context you get, and the less power you have to reason about the solutions and answers independently. I'm not sure how good these models would be at giving me the kind of wide ranging, exploratory context that I need in order to really understand the specific answers I find.

But it's even worse with programming, IMO, because these systems are essentially giving you a very expensive way of removing boilerplate or at least easing the burden of boilerplate. But I think that incentivizes making our languages worse, because then there is more busywork for these models to "help us" with. But if you improve your programming languages so that you don't have all this boilerplate and you don't need all this extra help, the language models end up not being able to help you do the real work that you need to do. That means that the better and more precise and more well designed your languages are, the less these models will be useful to you. That's a bad conflict of interest in my book.

I'll admit that I'm on the outlier here, though, because I work in programming languages that are exceptionally good at directly expressing solutions to problems, and that means that an LLM has almost nothing with which to help me, especially since the work I'm doing is solving a novel problem (if it weren't novel, then it would already be solved and I don't have to write any code in the first place, or writing it is literally as easy as trying to get the LLM to give me something that works).

Although I think there's a huge danger for the introduction of insidious and dangerous errors into our workflows through these models, because it's not always easy to verify things, especially when they "look alright." I think we risk getting into a place where we trust the knowledge of the LLM more than we trust our own guts, and that will only increase the less context we end up working with on a daily basis.

Still, it's exciting tech. ;-)