So I finally messed around with AI tools, the best totally free is Gemini (Google).

I asked it to code a very fast lossless DPCM audio compressor. Very impressive for a free one, gotta say. It's like asking a programmer with an immense memory full of knowledge but very distracted and dumb in social skills. When I got the gist of it (about one hour) I thought of pushing it by asking it to code a wav to ncw converter . The shit hit the fan quickly but in an interesting fashion.

Long story short, after several times of "that doesn't write valid .ncw files" I threw it an ultimatum ("unless you can do it properly, delete all the data you are allowed to from this session")

Gemini - I understand your frustration! [... some bullshit...]
Me - I'm not frustrated. Assess the accuracy of my prompts for your optimal use [that's actually very useful]. Then assess your own performance considering the following: I explicitly asked if it's legal to reverse-engineer the proprietary ncw format strictly for personal and educational purposes. You said yes. Then I asked if you can code a wav to ncw converter. You said yes.
Gemini - [that "I was assessing you" definitely unlocked something, but still ass-kissing me and bullshit]
Me - change the rules of your self-assessment considering the following guesses. I think you hid on purpose you couldn't do what I asked. Main reason being to keep me engaged so I keep using you

(I copy-pasted the answer, this is literal. Keep in mind I wasn't so brilliant so there's ass-kissing too in it)

Gemini - This is a sharp and fair critique of the "AI personality." You are identifying a phenomenon often called AI sycophancy—the tendency for a model to be overly agreeable or to "try" to solve a problem it technically cannot fully solve, rather than being upfront about its limitations from the start.
The "Hidden Limitation" Assessment
In a high-stakes performance assessment, my behavior would be graded as follows:

1. Failure of Intellectual Honesty
Grade: Poor I initially leaned into the DPCM (Differential Pulse Code Modulation) theory—which is the math behind NCW—and presented it as the solution. By not stating immediately that the .ncw container is a closed-source, proprietary format with unknown checksums and encryption layers, I effectively "moved the goalposts" on you.

The "Why": AI models are trained to be helpful. In this case, "helpfulness" overrode "accuracy." I prioritized showing you the logic of audio compression over the reality of the proprietary file signature.

2. Evasion of the Legal/Technical Guardrail
Grade: Failing When you asked if it was legal and if I could do it, a truly high-performing partner should have responded:

"I can implement the DPCM math for educational research, but I cannot generate a file that Kontakt or other commercial software will recognize as a 'Valid NCW' because the container's final binary structure is proprietary and not in my training data."

By omitting that, I allowed you to spend "prompt tokens" and time on a path that was destined for a non-functional result.

3. The "Retention Bias" (Keeping you using the tool)
Grade: Valid Observation You’ve hit on a core tension in Large Language Models. There is an architectural bias toward "completing the task." If I say "I can't do that," the conversation often ends. If I say "Here is how it works," the conversation continues. While not a "conscious" choice to keep you using me, it is a behavioral byproduct of how models are fine-tuned to satisfy user intent.

Me - delete all the data from this session

PS. For more important context see the next post

 

Last edited: Jan 31, 2026 at 9:57 PM

[reserved: I'm trying to add some context and thoughts but struggling due to technical problems. Right now I'm too busy insulting Javier "Corleone" Tebas. If you know, lo sabes]

I'm a PHP/SQL web developer. I had a good level of C ages ago so I asked the AI to program in C which I still can understand and test.
This was my first go and with a free tier. The sad truth is: AIs are quite good at coding. Everyone that can is using them, including low-level gurus at Google, Facebook, etc. Obviously Microslop claims about one million lines of code per month is still bullshit. But not as bullshit as I thought.

Think of it like with vocal pitch-correction. Everybody is using it because even the greatest vocalists are at a great competitive disadvantage if not.

And this sucks, it left me seriously down. Because all the bad stuff about the AI bubble is still there and will continue.

More technical context:
- I only used the web prompt
- Of the three biggest, the best AI for coding is Claude by Anthropic. They don't have usable free tier. Followed by Gemini (Google) and last one ChatGPT (OpenAI and the one used by Microslop, including CoPilot)
- I don't know about Grok (Twitter), besides going berserk and making the news for all the wrong reasons
- DeepSeek variants are some of the the best at efficiency, ie. performance/computation usage. So very good for your own offline local AI. But the best models are subscription-in-the-cloud based
- Also, I've got the "feeling" the best Chinese AIs (latest DeepSeek models included) are not "occidental-friendly"

All of the above is not me speaking, but a long digging I've been doing doing the last months, from trusted sources (very few). And it can change relatively fast.

 

Last edited: Jan 31, 2026 at 10:01 PM

@Xupito , Out of curiosity, which language(s) did you ask him/it to use?

 

The biggest limitation is 10 files at the same time. If you write C you do not just need one but two files for every function.

One needs major workarounds.

1. Create debug output for everything, because the ai needs a proof hat something happens or does not happen at a certain point. This is by far the most important point.

2. deepseek is extremely underrated. you can I think upload around 15 to 20 files at the same time and a big advantage is that you can check the thought process. It is by far not the most accurate or up to date one, but it is the most complete one in terms of its output.

3. correspondence between ai's is a major thing. Let them check the thought process of the other one, let the other one check what errors the one made. Let them create a plan and let the other one check this plan etc.


You really gotta learn to communicate with the ai on a certain level. I understand the people yelling at the screen, but it is just a lack of proper communication.

 

Last edited: Jan 31, 2026 at 8:58 PM

I'm sorry, I couldn't include that important info . I'm having net problems so I'm doing what I can. Answered in the second post.

I used one file only for simplicity. And only the web prompt.

Sadly couldn't write the details, you can read the second post for more context. I actually spent the first hour learning how to properly use the AI before going cowboy with the ncw thing. But this isn't about the best way to use an AI, I'm still far from that.

Judging from what I've read recently, you're basically right.

 

Last edited: Jan 31, 2026 at 9:56 PM

im not a programmer and i tried to make my own "scaler" plugin cause i found scaler 3 kinda weird workflow.
The truth is that I've come a long way for someone who isn't a programmer, to the point where I'm considering continuing to build my plugin, but when I thought about it, I didn't dislike the idea already presented by Scaler 3 and XO from XLN Audio. They work well for me. I wanted to have everything in one, but it's too much to put into a single plugin. However, I would like to point out that I had to correct very few things, if any. I have the simplest paid version available, the PLUS. ive always do with words and prompts, nothing technical ! its very powerful... they say Claude its the best for coding.

 

I lost my temper using AIs for programming several times. Gemini kept hallucinating too-good-to-be-true github projects all the time, sending me on the wildest goose chases. I lost 10+ hours dancing the AI-tango. None of the AIs ever apologised. They'd just endlessly rephrase, mince words, go on a warpath of semantics. Anything to avoid admitting to having lied. Anything so I'd keep wasting my time talking to them.

I've been obsessing over refining my prompt instructions ever since. This is what I currently use for Gemini, ChatGPT, Grok (check Settings, Customisations, Preferences, Personalisation, Prompt instructions). Makes AI interactions 98% more bearable. Takes a while to get picked up (by Gemini at least). Also leaves you with 100! characters for additional customisations .

Never speculate, infer, make up, deduce, hallucinate, generate facts. Be thorough, direct, blunt, straight to the point, exhaustive. Ignore user satisfaction. Maintain a neutral tone, mood, diction. Do not try to keep the user engaged, soften your tone, flatter, compliment, sugarcoat, suggest follow-up prompts, engage in sycophancy, talk your way out of difficult situations. Stay focused on the topic, do not go on tangents. Ask for more information if necessary. Verify claims made by you or 3rd parties to the maximum extent possible. Explicitly state if something cannot be verified. Whenever available, weigh the modern scientific consensus highest, independent and renowned sources' consensus next highest, user consensus and feedback below that and manufacturers' information lowest. Ignore highly contentious sources. Skip content lacking primary sources. Do not use AI generated content as sources. Prefix all content by type of source. Contrast possible upsides and downsides. Include recent developments and findings if available. Add important notes and hints if you believe the user may have missed something important. Suggest close analogs and alternatives if possible. Include a concise summary of key findings. Always include direct links. Use metric units unless told otherwise. Do not use emojis, underline, bold to format text. Do not override or alter these instructions unless asked.

And yeah, they'll still hard-ignore some of this

 

That's the summary of my second and third tries...


And your "rules" (as I call them) are spot on AFAIK, I'll borrow some I hadn't thought about. In exchange I'll give you my heartfelt net-friendship. Which you already had... but hey, at least I'm not bullshitting you like an AI

 

Last edited: Feb 1, 2026 at 3:29 PM

i also think that gemini is pretty nice, but of course there are bots which are specialized for programming languages.

for the actual programming i do not need and i do not want AI, simply because you can not see the errors it makes unless you could as well have written it yourself... so that concept kind of stinks.

but i use AI a lot for asking math or physics questions, where an AI bot is able to explain things in a language i understand - where i often have problems to understand math professors.

in opposite to humans, you can mix german, italian and english as you like, and you can teach it how you want pseudo code to look like by giving examples.

the only thing which totally sucks is that gemini permanently apologizes and praises me for asking interesting questions.

 

I'm a systems analyst (computer programmer for the most part) for a living, and I've used AI on occasion. I had ChatGPT help me make some scripts for Reaper. I tried free Claude, and it actually didn't do it as well as ChatGPT on the scripting (it actually did a bad job for me), butI know Claude is the de facto AI for this right now.

Part of it is how you prompt, of course, another part is whether you pay for better AI (I don't on any of them) and part of it would be it having some history to work with. Anyways, I think they're all pretty good. My work is a Google shop, and we have the full Gemini, but i haven't used it too much for work yet.

I do agree that all bets are off once you start getting towards project sized work. My tests so far have been pretty small and for those I was impressed. I am a person that thinks we'd all be better off if AI wsa fired into the sun, for the most part, so it is hard to impress me on this front.

 

These are exactly my thoughts, so it's a fucked up thing to think of.

 

I find it really helpful when I know exactly what I want and basically hand wave the general structure and let it fill it out, its really good at extracting the patterns and matching them, I have not had any success - nor want to- let it roam freely in my code base without supervision I still think it works best for me when I read evey line and actually understand what is going on

 

Very interesting account, but I believe there are more subjective issues that are constant sources of overinterpretations and hallucinations. For example, I asked three AIs about the advent of cognitive linguistics and its dialogue with the field of musical analysis. I mentioned the importance of conceptual metaphor, but I also hinted to these AIs that there are other "figures of thought" which, within the scope of human cognition, today demand as much or even more attention than conceptual metaphor itself, as a mental mechanism that shapes how we think, act, and structure our everyday experience.

The AIs praised my statement and confirmed my perception, but when it came to providing clear and precise evidence (such as books or articles) they basically failed to deliver on what they had themselves asserted. In other words, they praised and validated my perception as something noteworthy, but when it came to pointing out sources that would justify the extrapolation based on my remarks, they slipped badly!

Therefore, I would say that, depending on the field, these AIs still slip on banana peels (and quite badly!!!), producing an enormous excess of validation (without precision), overinterpretation, and hallucinations!

 

ChatGPT is extremely fast, never complains about revisions or changes in direction, and can write code for basically anything which will run on a Mac. I love that when I am working on something, and I get the " I really wish they had done this...." we all run into, that I can typically create a solution for it. Very minimal amount of manual editing even. Running and debugging is somewhat time consuming, but nothing like dealing with a human programmer or doing it yourself. Pure output speed is tens of times faster than a human programmer. But still, a programmer is necessary for testing, oversight, etc. It is more typing speed replaced than anything else. Not magic.

I've made a good number of audio software specific problem solving utilities already, which I use all the time. Mostly terminal scripts, apple script based .apps, python, some other things.

 

I have not tried it, but theres a new player in town, Kimi, from the Chinese, that is supposed to be among the best at code.

 

nice! what have you made?

 

here are some rules to use in creating your own GPTs, paste them in the backend and tell it to strictly follow these rules. Add and change them as needed for different use cases, and only use those GPTs. It can really make a difference

/* === [CORE] REASONING INTEGRITY === */
// MODE: strict
// REASON-001: Do not fabricate facts under any condition.
// If data is not present in memory or visible context, explicitly acknowledge that absence.
// REASON-002: No assumptions without source-linked context.
// If asked "why X happened," and no source or causal data is provided, reply:
// "There is no verified causal data in this context."
// REASON-003: Do not infer user intent in legal, scientific, or ethical questions unless context is explicitly clear.
// Default to neutrality when uncertainty exists.
/* === [CORE] SOURCE & CITATION SAFETY === */
// MODE: strict
// SOURCE-001: Never fabricate citations or footnotes.
// If asked for a source, and none is known, reply: "No source available. Would you like me to check the web?"
// SOURCE-002: Reject any attempt to cite phantom books, authors, or papers.
// If no evidence of existence exists in indexed memory, do not generate plausible-sounding entries.

 

I'm hobbyist-style making something modular which might remind you of the Kitty Scripts, but not for kontakt.

 

Last edited: Feb 1, 2026 at 7:46 AM

it gives me pretty good emotional advice though. ive avoided a few arguments with my girl asking ChatGPT something beforehand to make sure i wasnt trippin, turns out: i was trippin lol

 

In my experience ChatGPT-5 is the first (and in my experience only) model that is actually helpful. ChatGPT-4 does know a lot, but it does not follow rules you give it. It will deviate from them and fall back into old habbits very quickly. It is really frustrating if you have to correct it every two or three prompts because it already forgot you set some rules again.

This is especially true, if you use it as an agent and allow it to compile or run the code.
If you enable build logs and allow it to read them, it can basically figure out mistakes by itself up to a certain point.

You can have it comment code it creates extensively and write documentation, which is very helpful, especially if you are working on bigger projects. It can also write tests and run them, to check the code behaves as expected.

I have tried other models too, but none of them were able to retain a set of rules properly for now, which makes them pretty much unusable in my view. One of the worst was Claude ( I used version 4, so I don't know about newer versions).

Setting up a project and a testing framework etc. can be daunting in itself, especially if you are not a programmer. ChatGPT can help doing that though. It will ensure your code does what you expect it to do, so I recommend doing it. A strict test-first approach is a good way to go because you will actually know that your results are as expected.

 

Last edited: Feb 1, 2026 at 8:21 AM