Login
You're viewing the discuss.systems public feed.

Replies

  • 💬 9🔄 63⭐ 94
  • 💬 2🔄 1⭐ 15
  • 💬 0🔄 0⭐ 4
  • 💬 1🔄 1⭐ 0
  • 💬 1🔄 3⭐ 2
  • 💬 0🔄 0⭐ 2
  • 💬 1🔄 0⭐ 0
  • 💬 1🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • Reiner Jung 🇬🇱 🇺🇦 🇪🇺prefec2@norden.social
    Oct 16, 2026, 6:05 PM

    @nixCraft AI code assistance is best for mondaine tasks, like boiler plate code, solutions used by many programmers, like the stuff you can find on stack exchange and friends. It is less useful for very mathematical issues. At least this is my impression. Most likely, Claude could create my Python code faster than I do, but I usually don't know what I want to build.

    I would love to see studies on how well these tools perform with code close to machine hardware.

    💬 3🔄 1⭐ 5
  • 💬 1🔄 6⭐ 20
  • ShiroBlooShiroBloo@mastodon.social
    Oct 16, 2026, 10:22 PM

    @typhon @prefec2 @nixCraft You're being a bit unfair, I had it make some small Blender addons that just automate some feature, reducing 20 clicks into one. Choice quality didn't have to be perfect here.

    💬 0🔄 0⭐ 0
  • Leon the Lonerleontheloner@defcon.social
    Oct 16, 2026, 10:04 PM

    @prefec2 @nixCraft Your impression is correct. I could be wrong about this, maybe there's a switch in there somewhere that says "Do deterministic things now, become calculator" but to my understanding the output will always have some level of randomness involved. Really not a great fit for deterministic needs. Like math.

    💬 2🔄 0⭐ 1
  • Andreasandreasio@mastodon.social
    Oct 16, 2026, 10:19 PM

    @leontheloner @prefec2 @nixCraft there is a bunch of switches ("tools"). Basically, if the language model, reasons that the user is asking a math question, then the calculation is sent to a calculator app.

    💬 1🔄 0⭐ 1
  • 💬 0🔄 0⭐ 2
  • Heckin' ChonkerHeckinChonker@mastodon.social
    Oct 16, 2026, 10:26 PM

    @leontheloner @prefec2 @nixCraft There is a 'temperature' setting that controls how random the output of AI models is, and its set very low for coding tasks. When choosing the next token, models ran with a higher temperature will pick from the top choices randomly, where low temp will typically go with the top choice.

    I don't think I would count low temperature as being deterministic, but maybe somewhat less random.

    💬 2🔄 0⭐ 1
  • Heckin' ChonkerHeckinChonker@mastodon.social
    Oct 16, 2026, 10:32 PM

    @leontheloner @prefec2 @nixCraft We use claude for most coding tasks at my work. Engineers are still responsible for code they merge. Overall its maybe a 2x speedup. Reviewing PRs generated by Claude isn't really any different than reviewing a PR generated by my senior devs. Its correct a lot of the time, and the quality is generally good. Its not always correct and its still critical to have domain experts involved providing guard rails.

    All of the ethical concerns are still absolutely valid.

    💬 1🔄 1⭐ 0
  • argv[-1]argv_minus_one@mastodon.sdf.org
    Oct 17, 2026, 4:43 PM

    @HeckinChonker

    Either AI is as incompetent as I have personally found it to be, in which case damn near everyone else is suffering from mass psychosis, and every software I use is going to be ruined.

    Or AI is somehow actually good, in which case my days of solving problems with creativity and logic are numbered, I have no future, and my entire existence will soon be unnecessary.

    Either way makes me want to jump off a f***ing building. Jesus F***ing Christ.

    @leontheloner @prefec2 @nixCraft

    💬 2🔄 0⭐ 0
  • Reiner Jung 🇬🇱 🇺🇦 🇪🇺prefec2@norden.social
    Oct 17, 2026, 6:08 PM

    @argv_minus_one @HeckinChonker @leontheloner @nixCraft I had similar thoughts, but maybe it's different oder something in the middle. And you still have to understand the problem so you can explain it in a prompt. My normal way of solving issues is to express my thoughts in code, fail, and by doing this better understand what the problem is. Had I typed in my problem directly in an AI, it would have provided a solution for my limited understanding of the issue.

    💬 1🔄 0⭐ 0
  • argv[-1]argv_minus_one@mastodon.sdf.org
    Oct 17, 2026, 6:45 PM

    @prefec2

    On the rare occasions when I had a problem and could not find any solution, bots have proven profoundly unhelpful.

    I posted about one such incident last month. Here's the thread: mastodon.sdf.org/@argv_minus_o

    I'm sure the AI bros would tell me I was just prompting it wrong, but…well, see for yourself. I described the problem with exemplary precision and thoroughness, and in return I got unmitigated garbage.

    @HeckinChonker @leontheloner @nixCraft

    💬 1🔄 0⭐ 0
  • argv[-1]argv_minus_one@mastodon.sdf.org
    Oct 17, 2026, 6:47 PM

    @prefec2

    My email inbox is also daily bombarded with confidently-incorrect slop from my meat-proxy coworker, which has further soured my opinion of AI.

    If it's this disastrously wrong this often about things I know, what *else* is it frequently disastrously wrong about?

    So, seeing the likes of Linus Torvalds sing AI's praises has made me feel like I'm the only remaining sane man in a world gone completely mad.

    @HeckinChonker @leontheloner @nixCraft

    💬 2🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 1🔄 0⭐ 0
  • 💬 0🔄 0⭐ 2
  • 💬 0🔄 0⭐ 0
  • 💬 1🔄 0⭐ 0
  • argv[-1]argv_minus_one@mastodon.sdf.org
    Oct 18, 2026, 7:12 PM

    @stuart

    However, while anyone can plainly see that AI is incompetent within their own field, most people still perceive AI as shockingly competent in all other fields.

    Which is fallacious. Its “confidently incorrect” behavior in your field should have proven to you that it is not competent in *any* field. It's a con man, not an expert.

    @chemaunch @prefec2 @HeckinChonker @leontheloner @nixCraft

    💬 1🔄 0⭐ 0
  • Leon the Lonerleontheloner@defcon.social
    Oct 18, 2026, 7:49 PM

    @argv_minus_one @stuart @chemaunch @prefec2 @HeckinChonker @nixCraft

    And this is only so effective because confidently incorrect is also rewarded so regularly.

    It's also strange how pervasive it is. Cause I would go so far to say I, and many many of my colleagues over the years, have effectively been trained to be confidently incorrect (or at least seem knowledgable of something we are maybe vaguely aware of) even though we (well, some of us) are already competent at what we do, we're just demanded to be even more than what's reasonable.

    It's not just IT either.

    💬 0🔄 0⭐ 1
  • argv[-1]argv_minus_one@mastodon.sdf.org
    Oct 18, 2026, 8:01 PM

    @leontheloner

    Yeah, that's what makes me different from most people. I *hate* “confidently incorrect” behavior. As soon as I see someone behave that way, I write them off as untrustworthy in all matters.

    And yes, that would make the vast majority of people untrustworthy. Which…well…look around at what's happening in the world. Most people are indeed untrustworthy, and we've got the global socio-techno-economic collapse to prove it.

    @stuart @chemaunch @prefec2 @HeckinChonker @nixCraft

    💬 0🔄 0⭐ 0
  • Martin Taibrmartin_t@mastodon.social
    Oct 17, 2026, 8:48 PM

    @argv_minus_one @HeckinChonker @leontheloner @prefec2 @nixCraft Wrong way to look at it, IMO.

    There's no AI. There are models of text trained on a lot of code but also on what parts of the thought process which went into the code that they could get their hands on.

    They keep getting more data. Github opted people into training on their interactions. (Wonder whether those cunts never understood consent or don't think they're gonna get punished.) Even worse, people keep using the models 1/n

    💬 1🔄 0⭐ 0
  • Martin Taibrmartin_t@mastodon.social
    Oct 17, 2026, 8:51 PM

    @argv_minus_one @HeckinChonker @leontheloner @prefec2 @nixCraft
    and thus feeding the companies extremely fine grained data.

    (Some of them opt in to improving the model, some of the interaction data is used without consent or eventually will be when the mean expected loss from lawsuits is less than mean expected gain from better models.)

    I see very competent people saying LLMs don't help them at all in the long term and I see very competent people saying they're a massive productivity gain.2/n

    💬 0🔄 0⭐ 0
  • Martin Taibrmartin_t@mastodon.social
    Oct 17, 2026, 8:54 PM

    @argv_minus_one @HeckinChonker @leontheloner @prefec2 @nixCraft
    These data points are meaningless without knowing the type of project and what type of changes they made to gauge how represented the tasks are in the training data.

    LLMs are also a moving target. People who were saying LLMs completely changed the way they develop (they always use the verb "code" fsr) a year ago were either full of BS, so bad at dev they couldn't program almost at all or were doing boilerplate-heavy stuff. 3/n

    💬 0🔄 0⭐ 0
  • Martin Taibrmartin_t@mastodon.social
    Oct 17, 2026, 8:57 PM

    @argv_minus_one @HeckinChonker @leontheloner @prefec2 @nixCraft
    The people saying they see gains now look credible. I've played with free LLMs (I cannot bring myself to pay to those who organize large scale #LaborTheft ) and I see the improvements. It's not hard to extrapolate those improvements into seeing how the paid SOTA models today could genuinely be useful.

    I also review code generated by LLMs, sadly, and it's clear devs who were shit back them use them to be shit at scale. 4/n

    💬 0🔄 0⭐ 0
  • Martin Taibrmartin_t@mastodon.social
    Oct 17, 2026, 8:59 PM

    @argv_minus_one @HeckinChonker @leontheloner @prefec2 @nixCraft
    Many people who oppose LLMs focus on the output of these shit devs and their echo chamber doesn't show them people who were previously competent and see modest gains.

    Here's the worst part:

    A process will likely continue for about as long as it has been running.

    As much as I hate them, LLMs will likely keep getting better for at least a few years.

    At the start of 2026 I figured based on METR graphs that 2026 might be the last5/n

    💬 0🔄 0⭐ 0
  • Martin Taibrmartin_t@mastodon.social
    Oct 17, 2026, 9:03 PM

    @argv_minus_one @HeckinChonker @leontheloner @prefec2 @nixCraft
    yeah when I can stay competitive without using them.

    That prediction has not changed.

    The exponential curve is still exponential. Supposedly, one prompt can do 16 hours of human expert work with SOTA models today. Having a rough feeling for how free models keep evolving and seeing output from SOTA models used by others, I can believe it.

    I'll probably be forced to start paying for them or find another career within a year. 6/n

    💬 1🔄 0⭐ 0
  • Martin Taibrmartin_t@mastodon.social
    Oct 17, 2026, 9:06 PM

    @argv_minus_one @HeckinChonker @leontheloner @prefec2 @nixCraft
    I'll probably pick the first option, keeping in mind that the output is the work of people whose data was used for training and they deserve compensation (both money and ownership).

    I honestly hope this tech will send a shock through society big enough to trigger another revolution and we hopefully settle into a more fair and just system. Then the people whose work was stolen can get what they deserve, retroactively. 7/n

    💬 2🔄 0⭐ 0
  • Martin Taibrmartin_t@mastodon.social
    Oct 17, 2026, 9:09 PM

    @argv_minus_one @HeckinChonker @leontheloner @prefec2 @nixCraft
    Slavery used to be legal and normal for most of human history, until it wasn't. (Well, except some places mostly outside the developed world).

    Then came serfdom and indentured servitude.

    Now we have employment.

    I hope in the future, we'll have cooperatives as the only legal way to govern companies and dividing profits according to hours worked and skill level the only legal way to distribute profit.

    8/n

    💬 1🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • argv[-1]argv_minus_one@mastodon.sdf.org
    Oct 17, 2026, 9:31 PM

    @martin_t

    People being suddenly forced to reveal their faces and IDs just to have a chat online should've sent a shock through them too, and I don't see any riots. Do you?

    Hell, I don't even see any riots over the global financial crisis that's impoverishing people left and right. The old adage that “no country is ever more than three missed meals away from a revolution” was pure fiction, apparently.

    @HeckinChonker @leontheloner @prefec2 @nixCraft

    💬 1🔄 0⭐ 0
  • Martin Taibrmartin_t@mastodon.social
    Oct 17, 2026, 9:47 PM

    @argv_minus_one @HeckinChonker @leontheloner @prefec2 @nixCraft
    Have they actually missed 3 meals though? How many people? And how many of them have rifles?

    It's not a binary thing but a multi dimensional optimization problem.

    I hate the conclusion I've come to but there's a lot of injustice people will tolerate. Rioting or, better, purposefully overthrowing the gov in an organized manner is risky and the cost benefit analysis of most people isn't there as long as they literally can eat.

    💬 1🔄 0⭐ 0
  • argv[-1]argv_minus_one@mastodon.sdf.org
    Oct 17, 2026, 10:05 PM

    @martin_t

    My partner in Paarthurnax, have you seen food and housing prices lately? We're all gonna starve to death in a matter of months at this rate!

    Unfortunately, most people seem completely incapable of extrapolating from current trends. They refuse to foresee what's about to happen until it actually happens, even when you outright tell them.

    @HeckinChonker @leontheloner @prefec2 @nixCraft

    💬 1🔄 0⭐ 0
  • Martin Taibrmartin_t@mastodon.social
    Oct 17, 2026, 10:15 PM

    @argv_minus_one @HeckinChonker @leontheloner @prefec2 @nixCraft
    Yes, I get paid less than most devs (tradeof for working on interesting shit) and I can afford rent,food,hobbies easily. I have enough left over to invest. Do I deserve more? Absolutely. But I am nowhere near starving and I have enough buffer to find another job.

    Now, *buying* an apartment, that's another matter. But that's exactly my point, people will let the rich extract wealth from them as long as they don't *literally* starve

    💬 0🔄 0⭐ 0
  • argv[-1]argv_minus_one@mastodon.sdf.org
    Oct 17, 2026, 9:16 PM

    @martin_t

    If LLMs are that good, then why did they fail this badly at answering my socket programming question? For an expert, it should've been trivial. mastodon.sdf.org/@argv_minus_o

    And why do they keep vomiting up meaningless slop at my coworker? An expert would've asked him a pile of follow-up questions and then called him on his bulls***.

    So no, I can't believe it.

    And I certainly won't give them access to my computer. I'm not completely nuts.

    @HeckinChonker @leontheloner @prefec2 @nixCraft

    💬 2🔄 0⭐ 0
  • Martin Taibrmartin_t@mastodon.social
    Oct 17, 2026, 9:36 PM

    @argv_minus_one @HeckinChonker @leontheloner @prefec2 @nixCraft
    I hate to sound like the hypers but 1) you used shit models - any model with small or mini in name is gonna be shit. 2) you used them as chat, not in a loop with deterministic tools (as "agents")

    Truth be told, I don't use them in loops, I also only use chat but I can already see how sometimes it takes 5 or more prompts to get the right output. It's frustrating for a human but in a loop where it can run tests, it'll only 1/n

    💬 1🔄 0⭐ 0
  • Martin Taibrmartin_t@mastodon.social
    Oct 17, 2026, 9:39 PM

    @argv_minus_one @HeckinChonker @leontheloner @prefec2 @nixCraft
    the final solution.

    The trick is they're monkeys on typewriters with a somewhat good heuristic. If they only produce correct results 1 out of 10 tried but they can try things 100x faster than a human, they're a net economic win. It also explains why they're really good at stuff where they can mechanically verify the output but then completely go off the rails in ways no human would.

    Out of curiosity, can you try better models?2/n

    💬 1🔄 0⭐ 0
  • Martin Taibrmartin_t@mastodon.social
    Oct 17, 2026, 9:42 PM

    @argv_minus_one @HeckinChonker @leontheloner @prefec2 @nixCraft
    Anthropic lets me use Sonnet 5 with a free account. Try setting the effort level to medium or high, see what happens.

    "Open""AI" doesn't let me select a model in their interface but when prompted, it generated that it is GPT-6 which sounds like BS to me but worth trying your prompt there as well.

    (I'd do it myself but I can't verify whether the answer is correct or not and I find it rude to submit other people's work to LLMs)

    💬 0🔄 0⭐ 0
  • argv[-1]argv_minus_one@mastodon.sdf.org
    Oct 17, 2026, 9:51 PM

    @martin_t

    You definitely sound like the hypers.

    Obviously I'm not going to fund Anthropic, so no.

    Nor am I insane enough to give a bot permission to execute arbitrary code on my computer. They can verify their work on their own damn computers.

    I don't need a bot to throw s**t at the wall to see what sticks. If I needed that, I'd hire a teenager. I need code that's carefully designed, with all the consequences and edge cases carefully handled.

    @HeckinChonker @leontheloner @prefec2 @nixCraft

    💬 2🔄 0⭐ 0
  • Martin Taibrmartin_t@mastodon.social
    Oct 17, 2026, 9:57 PM

    @argv_minus_one @HeckinChonker @leontheloner @prefec2 @nixCraft
    That's why I said to try free models. Even the free ones I proposed will have more stolen work in them than the small/mini you used. You're likely not even giving them any data they don't already have since you posted the prompts publicly.

    I am on your side, I literally have a half-written blog post titled Artificial Intelligence and my Penis. I feel exactly like you. But sticking your head in the sand doesn't solve anything.

    💬 1🔄 0⭐ 0
  • argv[-1]argv_minus_one@mastodon.sdf.org
    Oct 17, 2026, 10:12 PM

    @martin_t

    I mean, you yourself described it as a monkey with a typewriter that only produces correct results once out of every ten attempts, and attempts to solve its failures by brute force instead of logic and learning. I have no need for that.

    Anyway, feel free to feed it my prompts yourself, if you're curious as to what it will do. They are, as you say, public knowledge now. But I am not curious.

    @HeckinChonker @leontheloner @prefec2 @nixCraft

    💬 0🔄 0⭐ 0
  • Martin Taibrmartin_t@mastodon.social
    Oct 17, 2026, 10:10 PM

    @argv_minus_one @HeckinChonker @leontheloner @prefec2 @nixCraft
    BTW, I absolutely hate when people who've been hyping LLMs for 3 years keep saying "try the latest model". They're like a broken clock. Except a broken clock is right twice a day, then goes back to being wrong. LLMs still produce a lot of shit but keep getting less bad. I bet latest free models can gen a sudoku solver 99 times out of 100. Bigger tasks soon. And once those assholes stop being wrong, they'll stop being wrong forever.

    💬 1🔄 0⭐ 0
  • argv[-1]argv_minus_one@mastodon.sdf.org
    Oct 17, 2026, 10:27 PM

    @martin_t

    Mm. In that case, maybe my family can use this future LLM to replace me, and I can finally check out of this godforsaken hopeless hellscape.

    But so far, I'm not seeing any evidence of what you're saying. My coworker continues to email me a firehose of asinine slop from full-size Claude that no intelligent human expert would ever write.

    @HeckinChonker @leontheloner @prefec2 @nixCraft

    💬 1🔄 0⭐ 0
  • 💬 1🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • Heckin' ChonkerHeckinChonker@mastodon.social
    Oct 17, 2026, 11:49 PM

    @argv_minus_one @martin_t @leontheloner @prefec2 @nixCraft I've only used copilot and claude, as part of my companies subscription. Copilot was garbage. Clade has been much better. Its typically only wrong when its missing some domain knowledge that was not in its context, and its trivial to resolve those cases.

    I understand not wanting to use them for ethical and moral reasons. I'm only using them at work myself because it very much feels like the alternative is losing my source of income.

    💬 0🔄 0⭐ 0
  • 💬 2🔄 0⭐ 0
  • Martin Taibrmartin_t@mastodon.social
    Oct 17, 2026, 2:19 PM

    @f4grx @HeckinChonker @leontheloner @prefec2 @nixCraft
    Yes, but it will produce useful output often enough to be a net productivity gain *when used for small enough tasks*. The labor theft is a separate topic from whether it works.

    Mastodon is an echo chamber which claims it's unethical and (implicitly) therefore it doesn't work. Which is BS.

    Same with the bug in Linus' code. It probably saved time on average. We should focus on the stolen patterns of work, not whether it works.

    💬 2🔄 0⭐ 0
  • 💬 0🔄 0⭐ 1
  • Sunday Midnightsundaymidnight@mastodon.social
    Oct 17, 2026, 2:53 PM

    @martin_t @f4grx @HeckinChonker @leontheloner @prefec2 @nixCraft
    It's you who keeps it all running, y'know? You and what you actually understand. Thing is, the more you lean on the machine, the less you really get the code you're supposed to be looking after. That's exactly what I'm saying. And hey, maybe you're super disciplined but most people aren't. And that's playing out everywhere right now.

    💬 0🔄 0⭐ 1
  • Heckin' ChonkerHeckinChonker@mastodon.social
    Oct 17, 2026, 3:10 PM

    @sundaymidnight @martin_t @f4grx @leontheloner @prefec2 @nixCraft If I could find a remote job with comparable pay that did not involve LLMs I would. I've looked, and I don't think they exist. The change from slop machine to producing quality code happened much faster than I expected.

    Genuinely, what would you do here? The options seem to be to use it begrudgingly, or to try and change careers which would put my family at risk.

    💬 0🔄 1⭐ 2