And I get that but constantly talking about a problem that was already solved, defeats the purpose of talking about it and not to mention all of the AI thread have been closed this is the only one that hasn’t been closed, not to mention that the topic got so off topic so til it come back up as a relevant topic and people wanna change the rule of AI usage in Haiku then yes but I don’t see this happening since Ai assisted content is so restricted and already so hard on Ai assisted content
Philip,
It’s you who is keeping this thread going. Why don’t you just stop posting?
I should ask the same question to you and it would be nice if you could stop being off topic while you at it 
Yes of course!
Is it difficult to know if software is AI or human written isr spying on you? I certainly don’t like the idea of software trying to prompt me to buy something but is that more likely with AI written software. I suppose if it can only come via a large company, yes. Is Haiku putting itself at a disadvantage by not embracing AI in its own way? AI is getting better so as time goes on it should write better code, Human written code: is it not that just as difficult to understand by another person as AI generated code. Yes I see a problem where the developer of their own code can have more insight of what the bug is than someone else, that would likely be the case with a developer using AI. Can AI write software with less bugs or is it getting better?
It’s honestly gotten insanely good. I have Claude helping me redesign the usb_midi kernel driver to accommodate non-class-compliant MIDI devices. If you asked me 3 months ago, if I thought Claude could write kernel drivers for Haiku, and I would have asked if you were smoking rocks. Also rewriting the midi_server for additional functionality and customizability. We have been using ALSA code as a reference, for vendor, pid numbers, technical specs, and their procedures, none of which gets you into any type of license violations. Claude has strict instructions not to copy code from any project, and when referencing code from GPL licensed products, Claude also does a code audit at the end of the session. Of course, to actually be mergeable back into the OS, it would also need to go through 3rd party audits, which we can get for free.
3 Likes
I think this is the real core argument for or against AI assisted code. AI does have valuable usages. Deep bugs which are hard to trace through multiple code-bases can be located within a few minutes. Smart inputs to AI models can produce some really smart outputs.
Example:
- Probably Smart: Examine the radeon_hd driver, and find any potential null pointer de-references.
- Why it’s smart: It’s easy for an LLM to trace through code flows and find potential null pointer de-references. Zero creativity involved. It presents bugs and you can (or not) action on the findings.
- Likely Slop: Redesign the app_server to handle multiple monitors
- Why it’s slop: As soon as you move on to how to structure code, how to design API’s, how to present things to users… you’re sliding the needle towards creativity in design. You’ll get an application design matching everyone’s opinion on Stack Overflow.
I will say, if you are in this argument and haven’t used modern LLM’s, install something like codex and try them. Have an informed opinion on quality of code, and the latest AI models. Looking at the first generation of LLM’s 3-4 years ago, and forming your full permanent opinion on AI is disingenuous.
AI absolutely has the potential to mess up your code (it is definitely “confidently wrong”), but it can also surprise you with quality responses within your workflow. When AI is confidently wrong, if you push back on why it will “re-adjust” and can find better results.
7 Likes
This is only true if your objection to LLMs is literally “they cannot produce anything useful”, and pretty much no one is saying that; certainly no one here is.
From a purely utility perspective, of course they can sometimes produce useful suggestions, or code. In search results, the percentage of time that happens (i.e. results with no factual errors or misattributions) is in the 40% to 50% range, and up to around 90% with post-processing in the latest study I read. Search is easier to study than coding assistants, of course. 
The “confidently wrong” problem is the big stickler for me: even a 10% error rate is catastrophically bad for a much-used system like web search, given that the user population by definition won’t be able to tell when the answers are wrong (since if they already knew the correct answer, they wouldn’t have asked the question).
The problem is both better and worse in the context of coding assistants, since the user may be able to identify an incorrect choice from the assistant but if they don’t the results could be anything from an immediate build failure to a persistently sneaky bug that completely invalidates your security model. And the sneaky errors are far more likely to be missed and to make it into the shipped code.
To me, this renders the entire concept of LLM-based assistants worthless, since their output is inherently not trustworthy. I don’t need to see how impressive they can be when they get things right; I only need to see that they are fundamentally non-deterministic and that significant errors are well within their probability space.
And this only considers the utilitarian arguments. The moral arguments are, to many people, sufficient to dismiss LLM assistants without needing to even consider their potential utility.
I do realize that the official Haiku position thus far has been to consider only the utilitarian arguments, though.
2 Likes
Why? did anything change in the underlying techology? If not then I don’t see the need to try LSD version 4.3. “just see how good it’s gotten” completely misses the point in this entire discussion.
Audits cannot disprove you stole code and violated licenses, the fact that the model did steal code and violate the license is already a known quantitiy. So if anything this appears like a “clever hack” to try and ignore liceneses… as xkcd sais “that clever hack you just though of is called insurance fraud, and we know about it”
4 Likes
That’s why the person using the AI also has to test it thoroughly. Bugs sneak through in human code as well, and apparently go undetected for decades. The great thing about AI making mistakes, is that you can just as easily fix any errors you make. When my research and planning lead me to a dead end, instead of having wasted weeks or months of planning and coding, I could then correct course in a few hours to try out a better method.
2 Likes
The “confidently wrong” problem is the big stickler for me: even a 10% error rate is catastrophically bad for a much-used system like web search, given that the user population by definition won’t be able to tell when the answers are wrong (since if they already knew the correct answer, they wouldn’t have asked the question).
For search, the solution is to ignore what it says and follow the links it collected. This way you don’t need to trust it too much, since links and websites are real.
For coding, you consider generated code a mere example and write code yourself. Having a working example saves quite a bit of time, since you don’t need to search for things, you only need to read what things do.
And for translation tasks, they are often easy to verify but tedious to do, also LLMs tend to have a smaller error rate on those.
1 Like
That’s why the person using the AI also has to test it thoroughly.
Program testing can be a very effective way to show the presence of bugs, but it is hopelessly inadequate for showing their absence.
I wouldnt say that, given the usage frequency of the word “slop” in this forum.
3 Likes
Low quality software can be sometimes useful though.
“slop”? “low quality”? People act like Claude is just in automode all day, just hallucinating out code willy-nilly. Even the few Fable queries I was able to run before they shut it down, it didn’t have full autonomy as the media claim it can do, I interactively looked at every line of code and approved or rejected it before it was written, let alone committed. Granted my C++ skills are beginner level (I mainly use Python and web based APIs in my real job) and I haven’t touched my BeOS software in nearly 20 years. But like I said, it is very controlled. And when I find new Haiku quirks that they are getting wrong, I add them to a haiku-gotchas-md so that they don’t make the same mistakes.
2 Likes
You know what? That’s fair enough; calling it “slop” is calling it garbage, not just saying it has a significant probability of being garbage.
I have avoided using that word here, because it’s rhetorically just an insult and so I hadn’t really been reading anyone else’s use of it literally. But the literal meaning of an insult is part of the meaning of what’s said.
2 Likes
I have been trying, amidst many pressing issues (including a compulsive hospital system), to establish standards for AI behaviors.
- If a bot finds no precedent, it assumes “impossible” when merely “unptrcedented” is the case.
This is a consistent ptoblem. Can you identify others? That one is a low-hanging fruit.