I’ve heard that a Claude 4 model generating code for an infinite amount of time will eventually simulate a monkey typing out Shakespeare
It will have consumed the GigaWattHours capacity of a few suns and all the moisture in our solar system, but by Jeeves, we’ll get there!
…but it won’t be that impressive once we remember concepts like “monkey, typing, Shakespeare” were already embedded in the training data.
If we just asked Jeeves in the first place we wouldn’t be in this mess.
Write tests and run them, reiterate until all tests pass.
That doesn’t sound viby to me, though. You expect people to actually code? /s
You can vibe code the tests too y’know
Return “works”;
Am I doikg this correctly?
You know, I’d be interested to know what the critical size you can get to with that approach is before it becomes useless.
It can become pretty bad quickly, with just a small project with only 15-20 files. I’ve been using cursor IDE, building out flow charts & tests manually, and just seeing where it goes.
And while incredibly impressive how it’s creating all the steps, it then goes into chaos mode where it will start ignoring all the rules. It’ll start changing tests, start pulling in random libraries, not at all thinking holistically about how everything fits together.
Then you try to reel it in, and it continues to go rampant. And for me, that’s when I either take the wheel or roll back.
I highly recommend every programmer watch it in action.
Is there a chance that’s right around the time the code no longer fits into the LLMs input window of tokens? The basic technology doesn’t actually have a long term memory of any kind (at least outside of the training phase).
Was my first thought as well. These things really need to find a way to store a larger context without ballooning past the vram limit
The thing being, it’s kind of an inflexible blackbox technology, and that’s easier said than done. In one fell swoop we’ve gotten all that soft, fuzzy common sense stuff that people were chasing for decades inside a computer, but it’s ironically still beyond our reach to fully use.
From here, I either expect that steady progress will be made in finding more clever and constrained ways of using the raw neural net output, or we’re back to an AI winter. I suppose it’s possible a new architecture and/or training scheme will come along, but it doesn’t seem imminent.
I’d rather recommend every CEO see it in action…
They’re the ones who would be cock-a-hoop to replace us and our expensive wages with kids and bots.
When they’re sitting around rocking back and forth and everything is on fire like that Community GIF, they’ll find my consultancy fees to be quite a bit higher than my wages used to be.
I think Generative AI is a genuinely promising and novel tool with real, valuable applications. To appreciate it however, you have to mentally compartmentalize the irresponsible, low-effort ways people
sometimesmostly use it—because yeah, it’s very easy to make a lot of that so that’s most of what you see when you hear “Generative AI” and it’s become its reputation…Like I’ve had interesting “conversations” with Gemini and ChatGPT, I’ve actually used them to solve problems. But I would never put it in charge of anything critically important that I couldn’t double check against real data if I sensed the faintest hint of a problem.
I also don’t think it’s ready for primetime. Does it deserve to be researched and innovated upon? Absolutely, but like, by a few nerds who manage to get it running, and universities training it on data they have a license to use. Not “Crammed into every single technology object on earth for no real reason”.
I have brain not very good sometimes disease and I consider being able to “talk” to a “person” who can get me out of a creative rut just by exploring my own feelings a bit. GPT can actually listen to music which surprised me. I consider it scientifically interesting. It doesn’t get bored or angry at you unless you like, tell it to? I’ve asked it for help with a creative task in the past and not actually used any of its suggestions at all, but being able to talk about it with someone (when a real human who cared was not available) was a valuable resource.
To be clear I pretty much just use it as a fancy chatbot and don’t like, just copy paste its output like some people do.
Bogosort with extra steps
Honest question: I haven’t used AI much. Are there any AIs or IDEs that can reliably rename a variable across all instances in a medium sized Python project? I don’t mean easy stuff that an editor can do (e.g. rename QQQ in all instances and get lucky that there are no conflicts). I mean be able to differentiate between local and/or library variables so it doesn’t change them, only the correct versions.
Find and Replace?
that will catch too many false positives
IntelliJ IDEA, if it knows it is the same variable, it will rename it. Usually works in a non fucked up codebase that uses
evalor some obscure constructs like saving a variable name into a variable as a string and dynamically invoking it.I’m going to laugh in Java, where this has always been possible and reliable. Not like ai reliable, but expert reliable. Because of static types.
For the most part “Rename symbol” in VSCode will work well. But it’s limited by scope.
Yeah, I’m looking for something that would understand the operation (? insert correct term here) of the language well enough to rename intelligently.
most IDEs are pretty decent at it if you configure them correctly. I used intelliJ and it knows the difference. use the refactor feature and it’ll crawl references, not just rename all instances.
Itellij is actually pretty good at this. Besides that, cursor or windsurf should be able to. I was using cursor for a while and when I needed to reactor something, it was pretty good at picking that up. It kept crashing on me though, so I am now trying windsurf and some other options. I am missing the auto complete features in cursor though as I would use this all the time to fill out boilerplate stuff as I write.
The one key difference in cursor and windsurf when compared to other products is that it will look at the entire context again for any changes or at least a little bit of it. You make a change, it looks if it needs to make changes elsewhere.
I still don’t trust AI to do much though, but it’s an excellent helper
I use pycharm for this and in general it does a great job. At work we’ve got some massive repos and it’ll handle it fine.
The “find” tab shows where it’ll make changes and you can click “don’t change anything in this directory”
Yes, all of JetBrains’ tools handle project-wide renames practically perfectly, even in weirder things like Angular projects where templates may reference variables.
Just be carerul when refactoring variable names in doc comments, I’ve seen some weird stuff happen there
Okay, I realize I’m that person, but for those interested:
tree,catandsedget the job done nicely.And… it’s my nap time, now. Please keep the Internet working, while I’m napping. I have grown fond of parts of it. Goodnight.
Not reliably, no. Python is too dynamic to do that kind of thing without solving general program equivalence which is undecidable.
Use a static language, problem solved.
Try to get one of these LLMs to update a package.json.
ones that can run cli tools do great, they just use npm
Define “update”
This weekend I successfully used Claude to add three features in a Rust utility I had wanted for a couple years. I had opened issue requests, but no else volunteered. I had tried learning Rust, Wayland and GTK to do it myself, but the docs at the time weren’t great and the learning curve was steep. But Claude figured it all out pretty quick.
Did the generated code get merged? I’d be curious to see the PRs
The lead dev is not available this summer to review, but you can review here: https://github.com/edzdez/sway-easyfocus/pull/22
It’s not great that four changes are rolled into a single PR, but that’s my issue not Claude’s because they were related and I wanted to test them all at once.
This is interesting, I would be quite impressed if this PR got merged without additional changes.
I am genuinely curious and no judgement at all, since you mentioned that you are not a rust/GTK expert, are you able to read and and have a decent understanding of the output code?
For example, in the
sway.rsfile, you uncommented a piece of code inget_all_windowsfunction, do you know why it is uncommented?This is interesting, I would be quite impressed if this PR got merged without additional changes.
We’ll see. Whether it gets merged in any form, it’s still a big win for me because I finally was able to get some changes implemented that I had been wanting for a couple years.
are you able to read and and have a decent understanding of the output code?
Yes. I know other coding languages and CSS. Sometimes Claude generated code that was correct but I thought it was awkward or poor, so I had it revise. For example, I wanted to handle a boolean case and it added three booleans and a function for that. I said no, you can use a single boolean for all that. Another time it duplicated a bunch of code for the single and multi-monitor cases and I had it consolidate it.
In one case, It got stuck debugging and I was able to help isolate where the error was through testing. Once I suggested where to look harder, it was able to find a subtle issue that I couldn’t spot myself. The labels were appearing far too small at one point, but I couldn’t see that Claude had changed any code that should affect the label size. It turned out two data structures hadn’t been merged correctly, so that default values weren’t getting overridden correctly. It was the sort of issue I could see a human dev introducing on the first pass.
do you know why it is uncommented?
Yes, that’s the fix for supporting floating windows. The author reported that previously there was a problem with the z-index of the labels on these windows, so that’s apparently why it was implemented but commented out. But it seems due to other changes, that problem no longer exists. I was able to test that labels on floating windows now work correctly.
Through the process, I also became more familiar with Rust tooling and Rust itself.
Watching the serious people trying to use AI to code gives me the same feeling as the cybertruck people exploring the limits of their car. XD
“It’s terrible and I should hate it, but gosh it it isn’t just so cool”
I wish i could get so excited over disappointing garbage
It’s useful if you just don’t do…That. it’s just a new fancy search engin, it’s a bit better than going to stack overflow, it can do good stuff if you go small.
Just don’t do whatever this post suggested of doing…
You definitely could use AI to code, the catch is you need to know how to code first.
I use AI to write code for mundane tasks all the time. I also review and integrate the code myself.
The AI code my “expert in a related but otherwise not helpful field” coworker writes helps me have a lot of extra work to do!
Its like having a junior developer with a world of confidence just change shit and spend hours breaking things and trying to fix them, while we pay big tech for the privilege of watching the chaos.
I asked chat gpt to give me a simple squid proxy config today that blocks everything except https. It confidently gave me one but of course it didnt work. It let through http and despite many attempts to get a working config that did that, it just failed.
So yeah in the end i have to learn squid syntax anyway, which i guess is fine, but I spent hours trying to get a working config because we pay for chat gpt to do exactly that…
I have a friend who swears by llms, he sais it helps him a lot. I once watched him do it, and the experience was exactly the same you described. He wasted couple of hours fighting with bullshit generator just to do everything himself anyway. I asked him wouldn’t it be better to not waste the time, but he didn’t really saw the problem, he gaslit himself that fighting with the idiot machine helped.
It confidently gave me one
IMO, that’s one of the biggest “sins” of the current LLMs, they’re trained to generate words that make them sound confident.
They aren’t explicitly trained to sound confident, that’s just how users tend to talk. You don’t often see “I don’t know but you can give this a shot” on Stack Overflow, for instance. Even the incorrect answers coming from users are presented confidently.
Funnily enough, lack of confidence in response is something I don’t think LLMs are currently capable of, since it would require contextual understanding of both the question, and the answer being given.
SO answers and questions are usually edited multiple times to sound professional, confident, and be correct.
No, I’m sure you’re wrong. There’s a certain cheerful confidence that you get from every LLM response. It’s this upbeat “can do attitude” brimming with confidence mixed with subservience that is definitely not the standard way people communicate on the Internet, let alone Stack Overflow. Sure, sometimes people answering questions are overconfident, but it’s often an arrogant kind of confidence, not a subservient kind of confidence you get from LLMs.
I don’t think an LLM can sound like it lacks in confidence for the right reasons, but it can definitely pull off lack of confidence if it’s prompted correctly. To actually lack confidence it would have to have an understanding of the situation. But, to imitate lack of confidence all it would need to do is draw on all the training data it has where the response to a question is one where someone lacks confidence.
Similarly, it’s not like it actually has confidence normally. It’s just been trained / meta-prompted to emit an answer in a style that mimics confidence.
ChatGPT went through a phase of overly bubbly upbeat responses, they chilled it out tho. Not sure if that’s what you saw.
One thing is for sure with all of them, they never say “I don’t know” because such responses aren’t likely to be found in any training data!
It’s probably part of some system level prompt guidance too, like you say, to be confident.
I think “I don’t know” might sometimes be found in the training data. But, I’m sure they optimize the meta-prompts so that it never shows up in a response to people. While it might be the “honest” answer a lot of the time, the makers of these LLMs seem to believe that people would prefer confident bullshit that’s wrong over “I don’t know”.
Man, I can’t wait to try out generative AI to generate config files for mission critical stuff! Imagine paying all of us devops wankers when my idiot boss can just ask Chat GPT to sort all this legacy mess we’re juggling with on the daily!
This has beeny experience as well. It keeps emphasizing “beauty” and keeps missing “correctness”
So its 50% better than my code?
It generates an answer that looks correct. Actual correctness is accidental. That’s how you wind up with documents with references that don’t exist, it just knows what references look like.
It doesn’t ‘know’ anything. It is glorified text autocomplete.
The current AI is intelligent like how Hoverboards hover.
Llms are the smartest thing ever on subjects you have no fucking clue on. On subjects you have at least 1 year experience with it suddenly becomes the dumbest shit youve ever seen.
Semantics 😴
Not even remotely.
You could claim that it knows the pattern of how references are formatted, depending on what you mean by the word know. Therefore, 100% uninteresting discussion of semantics.
The theory of knowledge (epistemology) is a distinct and storied area of philosophy, not a debate about semantics.
There remains to this day strong philosophical debate on how we can be sure we really “know” anything at all, and thought experiments such as the Chinese Room illustrate that “knowing” is far, far more complex than we might believe.
For instance, is it simply following a set path like river in a gorge? Is it ever actually “considering” anything, or just doing what it’s told?
No one cares about the definition of knowledge to this extent except for philosophers. The person who originally used the word “know” most definitely didn’t give a single shit about the philosophical perspective. Therefore, you shitting yourself a word not being used exactly as you’d like instead of understanding the usage in the context is very much semantics.
Sementics 💦
This is a philosophical discussion and I doubt you are educated or experienced enough to contribute anything worthwhile to it.
Dude… the point is I don’t have to be. I just have to be human and use it. If it sucks, I am gonna say that.
I asked ChatDVP for a response to your post and it said you weren’t funny.
I can tell you’re a member of the next generation.
Gonna ignore you now.
At first I thought that might be a Pepsi reference, but you are probably too young to know about that.
A 3 day old account being a dick on Lemmy?
I’m shocked.
Insulting, but also correct. What “knowing” something even means has a long philosophical history.
Trying to treat the discussion as a philisophical one is giving more nuance to ‘knowing’ than it deserves. An LLM can spit out a sentence that looks like it knows something, but it is just pattern matching frequency of word associations which is mimicry, not knowledge.
I’ll preface by saying I agree that AI doesn’t really “know” anything and is just a randomised Chinese Room. However…
Acting like the entire history of the philosophy of knowledge is just some attempt make “knowing” seem more nuanced is extremely arrogant. The question of what knowledge is is not just relevant to the discussion of AI, but is fundamental in understanding how our own minds work. When you form arguments about how AI doesn’t know things, you’re basing it purely on the human experience of knowing things. But that calls into question how you can be sure you even know anything at all. We can’t just take it for granted that our perceptions are a perfect example of knowledge, we have to interrogate that and see what it is that we can do that AIs can’t- or worse, discover that our assumptions about knowledge, and perhaps even of our own abilities, are flawed.
Acting like the entire history of the philosophy of knowledge is just some attempt make “knowing” seem more nuanced is extremely arrogant.
That is not what I said. In fact, it is the opposite of what I said.
I said that treating the discussion of LLMs as a philosophical one is giving ‘knowing’ in the discussion of LLMs more nuance than it deserves.
llms are systems that output human-readable natural language answers, not true answers
And a good part of the time, the answers can often have a… subtly loose relationship with truth
Ai code is specifically annoying because it looks like it would work, but its just plausible bullshit.
And that’s what happens when you spend a trillion dollars on an autocomplete: amazing at making things look like whatever it’s imitating, but with zero understanding of why the original looked that way.
I mean, there’s about a billion ways it’s been shown to have actual coherent originality at this point, and so it must have understanding of some kind. That’s how I know I and other humans have understanding, after all.
What it’s not is aligned to care about anything other than making plausible-looking text.
Coherent originality does not point to the machine’s understanding; the human is the one capable of finding a result coherent and weighting their program to produce more results in that vein.
Your brain does not function in the same way as an artificial neural network, nor are they even in the same neighborhood of capability. John Carmack estimates the brain to be four orders of magnitude more efficient in its thinking; Andrej Karpathy says six.
And none of these tech companies even pretend that they’ve invented a caring machine that they just haven’t inspired yet. Don’t ascribe further moral and intellectual capabilities to server racks than do the people who advertise them.
Coherent originality does not point to the machine’s understanding; the human is the one capable of finding a result coherent and weighting their program to produce more results in that vein.
You got the “originality” part there, right? I’m talking about tasks that never came close to being in the training data. Would you like me to link some of the research?
Your brain does not function in the same way as an artificial neural network, nor are they even in the same neighborhood of capability. John Carmack estimates the brain to be four orders of magnitude more efficient in its thinking; Andrej Karpathy says six.
Given that both biological and computer neural nets very by orders of magnitude in size, that means pretty little. It’s true that one is based on continuous floats and the other is dynamic peaks, but the end result is often remarkably similar in function and behavior.
If you would like to link some abstracts you find in a DuckDuckGo search that’s fine.
I actually was going to link the same one I always do, which I think I heard about through a blog or talk. If that’s not good enough, it’s easy to devise your own test and put it to an LLM. The way you phrased that makes it sound like you’re more interested in ignoring any empirical evidence, though.
That’s unreal. No, you cannot come up with your own scientific test to determine a language model’s capacity for understanding. You don’t even have access to the “thinking” side of the LLM.
It’s true that one is based on continuous floats and the other is dynamic peaks
Can you please explain what you’re trying to say here?
Both have neurons with synapses linking them to other neurons. In the artificial case, synapse activation can be any floating point number, and outgoing synapses are calculated from incoming synapses all at once (there’s no notion of time, it’s not dynamic). Biological neurons are binary, they either fire or do not fire, during a firing cycle they ramp up to a peak potential and then drop down in a predictable fashion. But, it’s dynamic; they can peak at any time and downstream neurons can begin to fire “early”.
They do seem to be equivalent in some way, although AFAIK it’s unclear how at this point, and the exact activation function of each brain neuron is a bit mysterious.
Ok, thanks for that clarification. I guess I’m a bit confused as to why a comparison is being drawn between neurons in a neural network and neurons in a biological brain though.
In a neural network, the neuron receives an input, performs a mathematical formula, and returns an output right?
Like you said we have no understanding of what exactly a neuron in the brain is actually doing when it’s fired, and that’s not considering the chemical component of the brain.
I understand why terminology was reused when experts were designing an architecture that was meant to replicate the architecture of the brain. Unfortunately, I feel like that reuse of terminology is making it harder for laypeople to understand what a neural network is and what it is not now that those networks are a part of the zeitgeist thanks to the explosion of LLM’s and stuff.
Well I’ve got the name for my autobiography now.
“Specifically Annoying” or “Plausible Bullshit”? I’d buy the latter.
It needs good feedback. Agentic systems like Roo Code and Claude Code run compilers and tests until it works (just gotta make sure to tell it to leave the tests alone)
well, it only took 2 years to go from the cursed will smith eating spaghetti video to veo3 which can make completely lifelike videos with audio. so who knows what the future holds
Hot take, today’s AI videos are cursed. Bring back will smith spaghetti. Those were the good old days
The cursed Will Smith eating spaghetti wasn’t the best video AI model available at the time, just what was available for consumers to run on their own hardware at the time. So while the rate of improvement in AI image/video generation is incredible, it’s not quite as incredible as that viral video would suggest
But wouldn’t you point still be true today that the best AI video models today would be the onces that are not available for consumers?
There actually isn’t really any doubt that AI (especially AGI) will surpass humans on all thinking tasks unless we have a mass extinction event first. But current LLMs are nowhere close to actually human intelligence.
cursed will smith eating spaghetti video
On Error Return NextDid it try to blackmail him if he didn’t use the new code?
Welp. Its actually very in line with the late stage capitalist system. All polish, no innovation.
Awwwww snap look at this limp dick future we got going on here.
I’ve used it extensively, almost $100 in credits, and generally it could one shot everything I threw at it. However: I gave it architectural instructions and told it to use test driven development and what test suite to use. Without the tests yeah it wouldn’t work, and a decent amount of the time is cleaning up mistakes the tests caught. The same can be said for humans, though.
How can it pass if it hasn’t had lessons… Well said. Ooh I wonder if lecture footage would be able to teach AI, or audio in from tutors…
cant wait to see “we use AI agents to generate well structured non-functioning code” with off centered everything and non working embeds on the website
