https://github.com/libsdl-org/SDL/issues/15350#issuecomment-4255050646
-
@pojntfx it doesnt help that it’s overhyped. opus 4.6 is genuinely useful in a way these tools weren’t a year ago.
but listening to the hype you’d think 4.6 came out a year ago.
it also doesn’t help that most people are not being exposed to frontier models
@phillmv That totally makes sense to me. Some people also just have a hard ideological aversion to LLMs that I imagine developed gradually over the last three years given that the people behind those companies are some of the most annoying people on the planet hyping up each other's startups
-
@phillmv That totally makes sense to me. Some people also just have a hard ideological aversion to LLMs that I imagine developed gradually over the last three years given that the people behind those companies are some of the most annoying people on the planet hyping up each other's startups
@phillmv To me it just looks like the same schisms that happened when systemd, Wayland, Btrfs, Kubernetes, Electron, VSCode or any other tech like that was introduced
Yeah the first iterations were not working and did not live up to the hype but people are clearly finding a lot of use from them, and if your brain is still under the assumption that things are the way they were years ago then today you're just _wrong_
-
LLM Policy? · Issue #15350 · libsdl-org/SDL
I've noticed the use of Copilot within a few reviews (13277 and 12730) which concerns me given the vast amount of issues associated with this technology (ethical, environmental, copyright, health, etc) so I was hoping a policy could be p...
GitHub (github.com)
I stg the entire anti-LLM crowd is just using tooling from two years ago and living in a parallel world as a result
Like wdym "I asked ChatGPT to make a change for me" that's not how it's worked for like more than a year, have you not engaged with the topic for like 5 minutes
I pay for, and regularly test, the most recent Claude code. It can be coaxed to produce working code, but the process remains about as fun as sifting through turds for nuggets of corn.
And that's the goal; the entire point of Claude is to make the process of writing code feel like micromanaging an idiot savant, but cheaper. As Anthropic releases new versions, the further you get from puzzle solving and the closer you get to management.
The reason I write code is to solve puzzles, and the details are important to coming up with a good solution.
-
I pay for, and regularly test, the most recent Claude code. It can be coaxed to produce working code, but the process remains about as fun as sifting through turds for nuggets of corn.
And that's the goal; the entire point of Claude is to make the process of writing code feel like micromanaging an idiot savant, but cheaper. As Anthropic releases new versions, the further you get from puzzle solving and the closer you get to management.
The reason I write code is to solve puzzles, and the details are important to coming up with a good solution.@ori I mean I totally agree re:details in implementations, I rarely if ever actually _commit_ anything LLM-generated, but have you not found it useful from an analysis perspective? Some things like that incredibly annoying trial-and-error loop while figuring out why the bindings generator isn't working properly for that one specific GObject class I can really short-circuit down from hours to a minute and then spend the remaining time getting the details right
-
@ori I mean I totally agree re:details in implementations, I rarely if ever actually _commit_ anything LLM-generated, but have you not found it useful from an analysis perspective? Some things like that incredibly annoying trial-and-error loop while figuring out why the bindings generator isn't working properly for that one specific GObject class I can really short-circuit down from hours to a minute and then spend the remaining time getting the details right
@ori > the further you get from puzzle solving and the closer you get to management
Idiot savant is a good way to put it, but idk, I also feel like I'm thinking quite a bit more now given that I can hack together shitty experiments for different ways of solving something much more easily. For example, if I'm trying to find out which CRDT makes the most sense to use with a P2P messaging framework I can just offload the process of trying the three different implementations in parallel
-
@ori > the further you get from puzzle solving and the closer you get to management
Idiot savant is a good way to put it, but idk, I also feel like I'm thinking quite a bit more now given that I can hack together shitty experiments for different ways of solving something much more easily. For example, if I'm trying to find out which CRDT makes the most sense to use with a P2P messaging framework I can just offload the process of trying the three different implementations in parallel
@ori In the past I'd have been far too lazy to actually try out all the options before making a decision, but maybe that's just me or specific to my problem space
-
Like, I can understand if you think these things are utterly useless if you use the completely wrong tool for things! This is like someone trying to use a dishwasher to run a Wayland compositor!
I think a lot of the pushback too comes from folx who are using modern tools etc, but are literally just generating all the code and shipping it, no reviews etc. When I see this happen so often, I can see how it would slowly push people towards a "no LLMs" rule
-
@ori > the further you get from puzzle solving and the closer you get to management
Idiot savant is a good way to put it, but idk, I also feel like I'm thinking quite a bit more now given that I can hack together shitty experiments for different ways of solving something much more easily. For example, if I'm trying to find out which CRDT makes the most sense to use with a P2P messaging framework I can just offload the process of trying the three different implementations in parallel
That's three times the management and almost no reasoning about the details. I wouldn't trust the assessments that come out of the process, and I won't use software that was written that way unless someone pays me to.
I think I'm probably going to get pushed out of this industry soon.
-
That's three times the management and almost no reasoning about the details. I wouldn't trust the assessments that come out of the process, and I won't use software that was written that way unless someone pays me to.
I think I'm probably going to get pushed out of this industry soon.@ori In the example here none of the code would actually be written by anyone other than a human tbc. I'm not sure about the "no reasoning" part honestly ... evaluating different implementations of things and comparing them against each other, finding out how and if the bindings would work ... that's something that at least in the contexts I'm aware of management would already have offloaded to two or three teams and pit them against each other. In small teams, that was not possible but now is
-
@ori In the example here none of the code would actually be written by anyone other than a human tbc. I'm not sure about the "no reasoning" part honestly ... evaluating different implementations of things and comparing them against each other, finding out how and if the bindings would work ... that's something that at least in the contexts I'm aware of management would already have offloaded to two or three teams and pit them against each other. In small teams, that was not possible but now is
@ori But yeah I agree re:management if someone's not comfortable with that this must absolutely suck. I didn't think of it that way before.
Sometimes I actually use them the other way, where I just let it prompt me to implement things instead without any changes. That way you still get a mental map of what changes are actually being made. Letting myself be guided by an autocomplete model also really does sound like how ever sci-fi book I've read describes the start of human enslavement by AI
-
@ori But yeah I agree re:management if someone's not comfortable with that this must absolutely suck. I didn't think of it that way before.
Sometimes I actually use them the other way, where I just let it prompt me to implement things instead without any changes. That way you still get a mental map of what changes are actually being made. Letting myself be guided by an autocomplete model also really does sound like how ever sci-fi book I've read describes the start of human enslavement by AI
I'm comfortable with management, but telling people what to do is the part of it that sucks. Being in a position to nurture the growth of their abilities is what's fun.
Claude doesn't have that. It's a seeded deterministic function. How do you feel about trying to nurture the personal growth of an overly talkative calculator?
-
R relay@relay.infosec.exchange shared this topic
