Chat-GPT 5 has just been released. Following it there should be new versions for other AI models soon as well - e.g., Grok 5 will be released later this year.
What’s new:
Lower hallucination rate.
Safe completions.
More advanced reasoning.
PhD-level expert - where GPT-3 was like talking to a high school student, and 4 was like a college student.
Vibe coding - can create software from a short prompt. e.g., a demo created two different web apps in seconds.
This looks good for using AI to help with coding applications - that can then be adjusted by a developer. For Haiku, perhaps it could improve on a debug app, or porting software, or creating some code needed for drivers.
As an assistant or helper to aid rookie developers to code good drivers to catch any potential memory leak…, yes. But to dump an entire blob of code without any human insight to be submitted, no.
I could find it useful, for example, to prompt to it “This is my repo: {repo url}. Please find any memory leaks, buffer overflows, security issues, race conditions… Do not just paste whole chunks of code here.”. Sometimes a bad code could get missed and for that, the AI could be useful.
Funnily,people already complain that it’s far worse than the previous models,which are now no longer available: ChatGPT-Nutzer sind mit der Einführung von GPT-5 richtig unzufrieden
Not that I’d personally care,never used that overhyped nonsense and never will,but I had to immediately think of that article I read yesterday when I found this thread.
Well, let’s see how “smart” AI is…
Absolutely dumb answers when asked anything concerning Numerical Analysis, completely erroneous code that uses library functions which don’t exist, or have the wrong kind/number of arguments. And even if the functions are correct, their use isn’t, and the code either doesn’t work at all, or compiles but gives wrong results. All answered in a “reasonable” way that could convince the naive user.
The motif is always the same: wrong code, even if it’s just a short subroutine. You tell AI this is wrong, it gives another answer, which is also wrong. You help AI telling where the errors are, it politely answers “you are absolutely right!”, then it “corrects” them with yet another wrong code. This goes on, and after several iterations, if you are lucky you may get an acceptable answer (which is not optimized, but at least it works). Usually, however, it falls into an endless loop, and after some point you get the same (wrong) code as before. You tell it “this is the same code you gave previously”, it says you re right and gives you… the exact same code again.
Whoever trusts AI for anything else than a glorified search engine is a fool shooting their own foot.
@syd : I’ve moved the topic to the “Off-Topic” category because there is no direct relevance to Haiku.
If you want to spark an interesting discussion about AI and Haiku you could wait until the new models go online, try them extensively and show us some code with relevance to Haiku instead of essentially just copying OpenAI’s press release. This way we would have a base of facts and data to talk about instead of arguing whether OpenAI’s claims are actually true.
I was tinkering with Grok concerning Haiku related application programming here recently. It claimed to have a good amount of knowledge of the Haiku APIs. So I asked for some example code for an application idea. It failed to achieve it first go when attempting to compile what it produced, but debugging and the dialogue was promising.
Those may be current problems, but @orignal said it couldn’t, implying there will never once be a possibility that they could ever be solved in the future, so that just begs the question, how do you know that to be the case?
I don’t disagree with that, but you’re treating AI as though the way it is currently is the way it will always be in the future, but in the last few years, we went from AI models being able to recreate only a vaguely bear-shaped image to being able to create relatively realistic video, and that trend is only going to continue for the forseeable future. Now you said it couldn’t serve those functions listed in the OP, implying it will never have that capability, so whence do you get this objectively factual information that AI will never be able to complete those functions or that it won’t be able to produce maintainable code?
Yes, I’m pretty confident that current so called “AI” couldn’t be used for any real development. Because in my personal opinion, it’s nothing, but scam.
However I don’t recall that a real AI might be implemented in future and not LLM-based.
Sure, that’s your claim, but I’m asking for you to substantiate it – as in explain why it is objectively never going to happen. If it’s just a matter of opinion, then you’re defeating your entire argument, as there’s no non-circular, non-arbitrary reason why anyone ought believe it, and thus arguing said position is entirely pointless.
Because it’s not based on new inventions or technologies. It’s only base on declaration “let’s consider this as an AI” and plenty of money to convince people, that it’s true. Another “trust the science” and mechanical turk.
but in the last few years, we went from AI models being able to recreate only a vaguely bear-shaped image to being able to create relatively realistic video, and that trend is only going to continue for the forseeable future
Video and image gen AI is weird: it’s both much more impressive and much more useless than language models. And more fun to play with.
But that’s just yet another claim. So you’re using a claim to ground your claim, but what grounds that claim? You have yet to actually ground your entire argument. In order to ground a claim, it must be based in some objective fact about reality, be it observable phenomena or fundamental axioms.
I mean, art will always be valued less – that’s just the way things go.
There’s one problem with replacing artists with AI: statistical models trained on large datasets have no taste by definition. Humans are biased in various ways, and that’s exactly what makes art interesting.
Besides, these models can’t be controlled well, which makes them useless as a tool (and artists already have a finely controlled tool to generate art, it’s called a 3D renderer).
But again, it’s fun to use these models to generate various silly stuff.
PS: making art is mostly decision making. Training merely allows you to see decisions you need to make.
Many developers use generative AI to save time on tasks, like checking software for errors, updating software, etc. A possible use would be to find some apps that could be ported to Haiku, and get it to help with the porting:
Smart-GPT 5 Copilot:
"Can you suggest some apps suitable for porting?
Absolutely! Haiku thrives on open-source software, especially those with minimal dependencies and cross-platform support. Here’s a curated list of apps that would be great candidates for porting to Haiku:
Developer-Friendly Tools
Geany – Lightweight IDE with minimal dependencies.
Kate – Advanced text editor from KDE, modular and extensible.
CMake – Widely used build system that’s essential for compiling many projects.
Web & Communication
Falkon – Qt-based web browser, lighter than Chromium.
Pidgin – Multi-protocol chat client.
qBittorrent – Open-source torrent client with a clean UI.
Creative & Multimedia
Krita – Professional-grade digital painting tool.
LMMS – Music production software.
HandBrake – Video transcoder with support for many formats.
System Utilities
BleachBit – System cleaner and privacy tool.
htop – Interactive process viewer.
GParted – Partition editor (may need UI tweaks for Haiku).
Godot Engine – Open-source game engine (some work already underway).
If you’re interested in trying one of these, I can help you check compatibility or even walk you through writing a port recipe. Want to start with something simple or go bold with a full-featured app?"