"AI" is a failed technology, in the sense that it's neither fit for purpose, nor a likely path to its stated end goal of "AGI", but a lot of people's profits currently depend on pretending it's not.
-
ok, look, you probably know who Linus Torvalds is, right? if not google it. then check it out: https://thenewstack.io/torvalds-linux-ai-stance/
Ah yes, invoking a notable person. My philosophies are all in shambles now!
Just wait until they shut off the tap on that juicy subsidized compute. When faced with the real costs you will not view this as efficient, and it is not efficient.
-
@j_s_j Mathematicians aren't freaked out because of how comprehensible and well-structured the LLM output is. It's the opposite. The models produce results (Lean code resulting in a "yes" or "no" to hard mathematical questions), accompanied by fever dreams of text that may or may not correspond to the Lean code. This is still apocalyptic for the field of mathematics because the "yes"es and "no"s themselves stand, regardless of the text quality and code quality. This approach to code isn't feasible if your software is to be used more than once, deal with the wildest inputs, and may end up running in a power plant.
@dss You’re right about *one* reason mathematicians are freaking out. Models also had trouble drawing hands a couple of years ago. There are human authored papers that can only be vetted properly by a handful of experts in related fields. When something comes along and purports to solve something humans haven’t for nearly a century, it shouldn’t be surprising the explanation needs analysis.
Mathematicians aren’t saying models aren’t making real discoveries.
-
the environmental-impact argument against AI doesn’t really hold water when you put the numbers in context. the comparison with everyday activities like eating a burger is pretty revealing.
https://welley.social/@skavish/posts/pt_01m4nt61s1ecjrvbczk5zb45z6
@skavish a few remarks:
- you choose a meat based meal, the carbon footprint of a balanced plant based meal is lower. We know eating meat every day isn't sustainable anyway, so you're not going anywhere by comparing two unsustainable things. To reduce CO2 emissions we need to reduce both of these things.
- the hardware manufacturing and the model training seem to be a big part of the footprint of AI, like the cow in your burger. We don't have reliable numbers from AI companies, but we can't consider them as negligible.
- at least the burger feeds you.. how vital is the result of those 6 hours of computation?cc @Gargron
-
open-weight models are pretty capable now
-
open-weight models are pretty capable now
@skavish Uh huh
-
@skavish a few remarks:
- you choose a meat based meal, the carbon footprint of a balanced plant based meal is lower. We know eating meat every day isn't sustainable anyway, so you're not going anywhere by comparing two unsustainable things. To reduce CO2 emissions we need to reduce both of these things.
- the hardware manufacturing and the model training seem to be a big part of the footprint of AI, like the cow in your burger. We don't have reliable numbers from AI companies, but we can't consider them as negligible.
- at least the burger feeds you.. how vital is the result of those 6 hours of computation?cc @Gargron
@eve my point is that ai's environmental impact is constantly presented as some enormous problem, while in reality, even 170 hours of agent use can have a smaller footprint than eating a burger. and that is before we even compare it with things like driving an suv, flying, air conditioning, or heating a house. somehow, those rarely get the same level of outrage. perspective matters.
-
@crazyeddie you are missing the point. his engineering expertise isn't in question. he explicitly said he's never used AI. so how can he confidently judge what AI-generated code looks like when the tools are used effectively?
@skavish Yeah, I tore that point apart. Maybe you missed it.
-
i have bad news for you: you do not know how to use it.
i do not know whose opinion you consider authoritative. how about Linus? he is famous for being strict about code quality, yet even he sees AI as a useful tool.
@skavish
I consider my experienced programming colleagues to have well grounded opinions. Those of them who have used these tools are equally skeptical of them.If the invariable response to skepticism is "You're doing it wrong," then I have bad news for YOU. You have lrft the path of wisdom.
-
@skavish Yeah, I tore that point apart. Maybe you missed it.
- of course an experienced engineer can recognize bad code. nobody disputes that.
- but seeing bad code does not tell you what a tool can produce when used well.
- to judge that, you need to understand how to use the tool, not just inspect its output.
these are three different things. you keep arguing the first while ignoring the other two.
-
- of course an experienced engineer can recognize bad code. nobody disputes that.
- but seeing bad code does not tell you what a tool can produce when used well.
- to judge that, you need to understand how to use the tool, not just inspect its output.
these are three different things. you keep arguing the first while ignoring the other two.
@skavish Yeah. #3 is your failure and sticking point. I do see that 100%. Thanks for repeating again though.
-
@skavish
I consider my experienced programming colleagues to have well grounded opinions. Those of them who have used these tools are equally skeptical of them.If the invariable response to skepticism is "You're doing it wrong," then I have bad news for YOU. You have lrft the path of wisdom.
it is genuinely mind-boggling. people who never used the tools, or know people who tried them and got shit results, confidently explain to people who use them successfully every day that they do not work. at some point, this stops being skepticism and starts looking like an ideology. the irony is fucking spectacular.
-
@skavish Yeah. #3 is your failure and sticking point. I do see that 100%. Thanks for repeating again though.
ah yes, the classic “i have never used it, but let me explain to you how it works.” truly the pinnacle of engineering wisdom.
-
"AI" is a failed technology, in the sense that it's neither fit for purpose, nor a likely path to its stated end goal of "AGI", but a lot of people's profits currently depend on pretending it's not. I am curious to see how long the hype will last once the Nvidia ouroboros ends and money stops being pumped into marketing.
This is what happens when a #bubble like this bursts.
A thread: https://todon.eu/@dusk/116892090757525061
Tricky to concisely summarize the videos therein, but the YouTuber does an excellent job, using the UK railroad boom as an example.
-
@mellowfish @Gargron You really should clarify what “past the point of no return” means to you.
Agentic LLMs for code generation (“engineering” is far too generous a term) will continue to EXIST, yes, because we can’t un-invent something. But I think its impact on software work is immensely overblown. Just as we’ve always had shitty, unmaintainable software, LLMs will continue to ensure that we will have shitty, unmaintainable software, but at an even higher volume. A lot of code doesn’t need to be anything more than shitty, unmaintainable code to be somewhat useful (cf: *motions at shitty software that already surrounds us*), but that doesn’t make them suddenly not shitty and unmaintainable, and generally a poor business investment for people who depend on reliable, maintainable software.
I think this makes a good counter-counterpoint: https://deadsimpletech.com/blog/llms-arent-inevitable
@drahardja @Gargron as someone who cares a lot about code quality and reads and understands every line of code I ship to prod, AI coding has (still) been a fantastic tool. I ship 2-3 times as much code as I have done in recent years with no real drop in quality. No increase in number/severity of bugs (there are always some no matter who is writing/testing) and the code is still easy to change over time. I can’t imagine going back.
-
@mellowfish @Gargron “Agentic engineering” causes numerous problems, has serious quality control issues, and is extremely resource intensive. Currently, it is supported by hype, and by the widespread availability of subsidised compute. Following a market correction, those factors will evaporate; It is extremely premature and unfounded to assume the practice will continue in any form.
@janef0421 @Gargron most of those problems and quality control issues evaporate if you just *read* and review the generated code. That’s the primary difference between agentic engineering and vibe coding. The inference compute is not as subsidized as you might think, especially for properly quantized open weight models vs frontier. Will frontier training slow down post bubble and prices for compute raise noticeably? Of course. But it will almost certainly still pay for itself.
-
RE: https://beige.party/@fuzzy/117417199780071912
Hint: most of the 1,266 accounts that I block are anti-AI.
Cc @skavish
@fuzzy @mastodonmigration @skavish Interesting enough, most of the accounts are block or pro-AI slop fondlers!
-
"AI" is a failed technology, in the sense that it's neither fit for purpose, nor a likely path to its stated end goal of "AGI", but a lot of people's profits currently depend on pretending it's not. I am curious to see how long the hype will last once the Nvidia ouroboros ends and money stops being pumped into marketing.
@Gargron There's an old saying amonst traders: "The markets can remain irrational longer than you can remain solvent." When the bubble bursts, the pain will be inflicted from the bottom up.
-
@janef0421 @Gargron most of those problems and quality control issues evaporate if you just *read* and review the generated code. That’s the primary difference between agentic engineering and vibe coding. The inference compute is not as subsidized as you might think, especially for properly quantized open weight models vs frontier. Will frontier training slow down post bubble and prices for compute raise noticeably? Of course. But it will almost certainly still pay for itself.
@mellowfish @Gargron Code review hasn’t resolved issues in human generated code, so I doubt that claim immensely. Maybe you perceive that your review efforts resolve all issues, but that’s a psychological bias - You’re only aware of the issues you identify, and so it appears you’ve identified all issues.
If some series of well-designed studies come out and confirm your claim, then I’ll believe it, but for now, it’s just an assertion you made.
Also, your statements about compute subsidisation don’t contradict what I’m saying, so I see no reason to respond.
-
@drahardja @Gargron as someone who cares a lot about code quality and reads and understands every line of code I ship to prod, AI coding has (still) been a fantastic tool. I ship 2-3 times as much code as I have done in recent years with no real drop in quality. No increase in number/severity of bugs (there are always some no matter who is writing/testing) and the code is still easy to change over time. I can’t imagine going back.
@mellowfish @Gargron Your experience and mine differ.
-
Not relevant to whether or not it has failed or works poorly, which is what the original post was about
@golemwire @Nead There are plenty of times when it fails miserably too.
Maybe on this problem it works, with an expensive model, and with the help of another piece of software that does all the heavy lifting. All it must do is what computers do best, repetitive work, in this case, backtracing spaghetti code, where it can clearly validate its success.
But for many programming projects, success and failure do not have this self-proving nature handed to them on a silver platter
