What was i supposed to take away from that demo?

Except sometimes you can't seem to stop the prose.

I've come close to blocking the site on my network many times but it's an absolute goldmine of interesting info too... I'm not really sure if there's a solution, other than to practice emotionally disengaging from internet discussions.

I suck at spreadsheets. I know they can do both useful and amazing things, but my daily life does not revolve around spreadsheets and I simply do not understand most of the syntax and operations required to make even fairly basic things work. It requires a lot of time and effort for me to get simple things done with a spreadsheet on the rare occasion that I need to manipulate one.

There are things in life that I am very good at; spreadsheets are simply not amongst them.

But do I know what I want, and I generally even have a ballpark idea of what the results should look like, and how to calculate it by hand [horror]. I just don't always know how to articulate it in a way that LibreOffice or Google Sheets or whatever can understand.

LLMs have helped to bridge that gap for me, but it's a pain in the ass: I have to be very careful with the context that I give the LLM (because garbage in is garbage out).

But in the demo, the LLM has the context already. This skips a ton of preamble setup steps to get the LLM ready to provide potentially-useful work, and moves closer to just making a request and getting the desired output.

Having one unified interface saves even more steps.

(And no, this isn't for everyone.)

Yes, "satisfaction"

Tried it once. Didn't get "satisfaction"; instead felt deeply irritated by the "backseat driver". Maybe it works better if you're just churning out mediocre boilerplate.

That "satisfaction" vanished pretty damn quickly, once I realised that I have often more work correcting the stuff so generated than I would have had writing it myself in the first place.

LLMs in programming absolutely have their uses, Lots of them actually, and I don't wanna miss them. But they are not "thinking ahead" of the code I write, not by a long shot.

A "wow, that's a great start" to one could be a "damn there's an issue I need to fix with this" to another. To some, that great start really makes them more productive. To others that 80% solution slows them down.

For some reason, programmers just love to be zealots and run flamewars to promote their tool of choice. Probably because they genuinely experience it's fantastic for them, and the other guy's tool wasn't, and they want them to see the light, too.

I prefer to judge people on the quality of their output, not the tools they use to produce it. There's evidently great code being written with uEmacs (Linux, Git), and I assume that, all the way on the other end of the spectrum, there's probably great code being written with VSCode and Copilot.

Web server work in Go, Python, and front end work in JavaScript - it's pretty good. Only when I try to do something truly application specific that it starts to get tripped up.

Multi threading python work - not bad, but occasionally makes some mistakes in understanding scope or appropriate safe memory access, which can be show stopping.

Deep learning, computer vision work - it gets some common pytorch patterns down pat, and basic computer vision work that you'd typically find tutorials for but struggles on any unique task.

Reinforcement learning for simulated robotics environments - it really struggles to keep up.

ROS2 - fantastic for most of the framework code for robotics projects, really great and recommended for someone getting used to ROS.

C++ work - REALLY struggles with anything beyond basic stuff. I was working with threading the other day and turned it off as all of its suggestions would never compile let alone do anything sensible.

There are a few things that really help the AI to understand what you want to do, otherwise it might struggle and come up with not so good code.

Not to say it gets it right everytime, but definitely often enough for me not to even consider turning it off. The time save has been tremendous.

“Oh you couldn’t take a train to work? Must have been something you did, the Palantir is usually great and helps our society. It always works great for me and my friends.”

Absolutely. LLMs let you make more programming mistakes faster than any other invention with the possible exceptions of handguns and Tequila.

To be fair, it is also really good at spewing out industrial levels of boilerplate. As we all know, 99% of the effort in coding is the writing of code and the more boilerplate in your code base the better. /s

https://i.redd.it/xr8uxqayv68d1.jpeg

For demos sure, but I'm not super hopeful on this frankly. LLMs are inherently about generating next char in sequential order. Nothing about real world spreadsheets is linear like that - they're all interlinked chaos.

I’m not sure it’s so much ‘satisfaction’; it felt more like I was having a stroke until I turned it off. Its suggestions were, like, plausibly code, but completely contextually nonsensical in general; frankly IntelliJ’s old autocomplete-with-guessing functionality was better, as it at least _knows_ a certain amount about the codebase. Now, this was on a very large old codebase; no doubt it’s better if writing trivial new things.

I did not get that feeling from CoPilot. I usually got the feeling that it was interrupting me to complete my thought but getting it wrong. It was incredibly annoying and distracting. Instead of helping me to think it was making it harder to think. Pair programming with an LLM has been great. Better than with most humans. But autocomplete sucks for me.