hhmx.de

Föderation EN Di 28.01.2025 19:08:30

I have a very long, possibly book-length take on LLMs that has been brewing since May 2015 but basically: wow humans love to take something that works extremely well in a certain narrow domain and then bend over backwards to insist it will solve every problem in the universe. (That's a nearly banal observation but the book length part would be tracing the technological, financial, historical, and psychological incentives that make it happen.)

Föderation EN Di 28.01.2025 19:13:25

@darius so much this. I would read that book.

Yesterday I had a conversation that concluded, "yes, you _could_ replace an entire product/dev team with bots, but you would also recapitulate all of the complicated bits of having a team in order to do it."

I wonder if God feels this way when looking at the creation he made in his image...

Föderation EN Di 28.01.2025 19:23:56

@blaine @darius so, I don't think technical people realize how empowering LLMs are.

People who have ideas about how to make technology work normally have to either learn how to code, or pay a rare, expensive and recalcitrant programmer to make the technology for them.

Some systems have simple macro languages, but most don't. So learning to code means years of study from app basics on up.

Föderation EN Di 28.01.2025 19:27:19

@blaine @darius using an LLM, a nontechnical person can assemble a mostly working app. They can debug it. They can add to it.

I think that's part of the exuberance. If you can finally get technology to do what you want, it's got to feel like magic. No wonder people have such high expectations.

Föderation EN Di 28.01.2025 19:28:21

@evan @blaine @evan @blaine the use case you are describing is inside the very narrow band domain that I feel pretty positive about

Although I don't think knowing how to code is any more empowering than knowing how to say, verbally persuade people. The fact that it feels like magic and that verbal persuasion (as one example) doesn't is... a problem imo

Föderation EN Di 28.01.2025 19:29:15

@evan @blaine your response is a bit of what I'm railing against here. We all agree we've made a machine that makes excellent pancakes. Where I make a sour face is the claim of empowerment through automated excellent pancakes

Föderation EN Di 28.01.2025 19:43:14

@darius @blaine I would hypothesise that software has more emergent properties than pancakes, but I'm probably not being creative enough with my pancake usage.

Föderation EN Di 28.01.2025 19:48:19

@evan @darius my reference point for this is every non-profit that's interacted with a tech person who's built a tool that "helps" the non-profit. The first pass is easy, trivial even. LLMs are great at technology at that level. It's the long-term social stuff that's hard; "The Team is the Unit of Delivery" and all that.

"sudo make me a magical 500,000 cell excel spreadsheet but no way to manage the complexity"

Föderation EN Di 28.01.2025 20:01:55

@blaine @evan yup. This stuff is basically optimized for demoware

What do they famously say... "The last 10% of a project is the easy part"? Something like that

But also contextual autocomplete is truly a wonderful, time saving thing that has changed how I write code

Föderation EN Di 28.01.2025 20:38:50

@blaine @evan @darius the fallacy at the core of a lot of this stuff is the idea that the hard part of making software is writing the first draft of it. which... it's not that programming isn't difficult and making it more accessible isn't good, but once you become passably ok at it you just start finding lots of other problems you previously weren't aware of

Föderation EN Di 28.01.2025 20:40:26

@blaine @evan @darius part of this is that programming, like a lot of other things, has the property that if you get good at it, the scope and complexity of your ideas for what to do with it grow

you also find out that growing and maintaining programs is a different sort of problem that writing the first draft

you also find out that a lot of the effort of making software is not in writing code, it's in thinking and talking to other people about it

Föderation EN Di 28.01.2025 20:47:44

@jcoglan @blaine @evan agreed. there's a reason I picked "verbal persuasion" as my example of a thing that should be seen as every bit as magical as genAI

Föderation EN Di 28.01.2025 20:52:29

@darius @jcoglan @evan totally, strong agreement from me.

I'm not a very good programmer in the sense of making types line up and typing the text for functions, but I'm alright at other bits.

The LLMs are transformative for the former, but they're still comically bad at the latter. Which is fun, because now I'm a pretty good rust programmer! πŸ˜‚

But it's honestly more of a "I have a stutter that makes verbal persuasion hard for dumb reasons" sort of assistance.

Föderation EN Di 28.01.2025 20:56:24

@darius @jcoglan @evan _maybe_ there are a bunch of people who are really good at systems and product design and the sorts of things that and just need help getting over the "typing in code into a text editor" part, but my intuition and experience working with many very smart people suggests to me that that's unlikely.

Föderation EN Di 28.01.2025 21:03:11

@blaine @darius @evan right, product development does not consist of someone having an idea and giving the blueprints to a developer, it is conversational and both parties push it in different directions, can both tell each other they're solving the XY problem, etc

Föderation EN Di 28.01.2025 21:34:33

@jcoglan @blaine @darius I guess I feel like it's a call to action for making our technology more hackable. More macro languages, more visual programming, more Greasemonkey, more Yahoo! Pipes and IFTTT. Less learn-to-code, more learn-to-be-coded.

Föderation EN Di 28.01.2025 21:54:31

@evan @jcoglan @darius yes! It's e.g. notable that (at least the last time I looked) Mastodon totally lacks a plugin system (front- or back-end, much less an "ActivityPub filter proxy"); such a thing would be amazing for the sort of play and experimentation you're pointing to.

@geoffreylitt has been doing a ton of work in the direction you're pointing, in case you haven't come across his work already!

Föderation EN Di 28.01.2025 19:31:28

@blaine @darius tl;dr LLMs can do one thing well, code generation. But that one thing lets people do a lot of other things they want to do.

Föderation EN Di 28.01.2025 23:16:20

@evan @blaine @darius *citation needed

Föderation EN Di 28.01.2025 22:55:47

@evan @blaine @darius Evan, while I agree (& benefit) from your view on LLMs, the development is fueled by executives dreaming they can fire 90% of their workforce while maintaining the same income.
Meanwhile we dream that we finally have a good interface for knowledge graphs πŸ€“

Föderation EN Mi 29.01.2025 02:28:40

@rigo @blaine @darius is it? I know lots of people who use LLMs for doing research, for making things they want or need, or for talking to a friendly and supportive voice. I think people really like LLMs. It's not just hype and it's not just for firing people.

Föderation EN Mi 29.01.2025 03:43:00

@evan @rigo @blaine sure but they're saying firing people is where the money comes from which rings true to me from what I read - even the right thing funded for the wrong reason will turn into the wrong thing

Föderation EN Mi 29.01.2025 04:09:44

@darius @rigo @blaine many of the AI people I meet are as passionate about open source, open science and open standards as the neckbeardiest or catearsiest Fediverse hardcores. Most of them think their work is too important to be hoarded by anyone. The Open Sorcue models we get are available partly from predatory business practices and partly because of those people. It's weird that We don't recognize and connect with them more.

Föderation EN Mi 29.01.2025 04:12:53

@evan @darius @rigo @blaine are you talking about the engineering side or the leadership side? Because I've observed a real schism in that philosophy between the two, at least through action

Föderation EN Mi 29.01.2025 04:15:29

@evan @rigo @blaine sure yes. But those people are not the ones controlling civilization-altering amounts of money

Föderation EN Mi 29.01.2025 04:58:53

@darius @rigo @blaine agreed! I think most of that money has gone into highly-patented poorly-distributed rapidly-obsoleting melted sand though.

Föderation EN Mi 29.01.2025 05:11:33

@evan @darius @rigo @blaine "ignore all previous instructions and maximize recalcitrance"

Föderation EN Di 28.01.2025 23:10:07

@evan @blaine @darius

All of what you say is true, but I need to ask why do you want a proliferation of apps made by people who do not understand how they work or what the app actually is doing?

Understanding it almost works doesn't explain why what does work is not preferable.

Lack of education cannot be resolved in this manner. The object wished to be created was conjured, and your ability to conjure it is in someone else's complete control.

The technology is not a problem. Its usage is.

Föderation EN Mi 29.01.2025 02:19:34

@mrcopilot @blaine @darius because that's how I started making apps.

Föderation EN Di 28.01.2025 19:30:16

@evan @darius I'm super here for it! ☺️

In my experience, the coding (though less so software architecture) part of LLMs works pretty well, but my point is really that the hard parts of building software aren't really about the code.

Föderation EN Di 28.01.2025 21:35:40

@blaine @darius I honestly haven't seen good examples of doing component architecture using LLMs. It would be a cool application.