rdos

[flagged]

Recently someone messaged me on Reddit about my post. I replied. They wrote again, I replied again. After a few messages I realized I was talking to an AI agent.

My exact experience. The irony was that we were talking about AI agents

The Tools detail page is wrong, it's the one from last release, November 2025.

almost fell for it

I can't seem to change the colors of the pie chart, other than the predefined themes. But all of those are horrible for a pie chart.

This bug is categorically distinct from hallucinations or missing permission boundaries

I was expecting some kind of explanation for this

Sure, two 3060 can pull usable performance on an usable LLM, but a single one can't (yet).

3x RTX 3060 less tgab the price of a 3090

Interesting, here it is around the same. 200-250€ for a used 12GB 3060 and 600-800 for a used 3090€.

Was any text in the repo NOT written by AI?

14B even at Q4 isn't realistic for coding on a single 12GB RTX 3060. Token speed is too slow. After all they are dense models. You aren't getting a good MoE model under 30B. You can do OCR, STT, TTS really well and for LLMs, good use cases are classification, summarization and extraction with <10B models.

llama.cpp (previously Ollama)

I almost fainted

Interesting. Won't stuff like entity extraction suffer? Especially in multilingual use cases. My worry is that a smaller model might not realize some text is actually a persons name because it is very unusual.

Is it possible for such a small model to outperform gemini 3 or is this a case of benchmarks not showing the reality? I would love to be hopeful, but so far an open source model was never better than a closed one even when benchmarks were showing that.

This is very interesting. Especially the last part where it shows gpt-5.2 and gpt-oss and their very similar and unique outcome of being 90%+ Serious.

I tested this locally and got the same result with gpt-oss 120b. But only on the default 'medium' reasoning effort. When I used 'low' I kept getting more playful responses with emojis and when I used 'high' I kept getting more guessing responses.

I had a lot of fun with this and it provided me with more insight than I would have thought.

Are you saying this from experience?

Yes. I mostly work on Quarkus microservices and use cursor with auto agent mode.

we wouldn't give an AI some vague requirements and ask it to build something we would discuss as a team

seems like a reasonable workflow. It's the polar opposite of what was written in the blog post. That is the usual, easy way people use agents and what I think is the wrong path. May I also ask what language and/or framework you work with where so much context works good enough?

Asking AI to explain code and help me learn how it works means I can pick up new systems significantly quicker.

Summarization is generaly a great task for LLMs

I didn't read the blog yet because I clicked on cat pics and there weren't any!!!

LLM's are good at making stuff from scratch and perfect when you don't have to worry about the codes future. 'Research' can be a great tool. But LLMs are horrible in big codebases and multiple micro services. Also at making decision, never let it make a decision for you. You need to know what's happening and you can't ship straight AI code. It can save time, but it's not a lot and it won't replace anyone.

I don't want to sound rude, but what was your reason to go from scratch instead of joining an already established, open source effort? The likes of Cline, Roo, Continue, ...

nice, what's your approach? Graphs?

this is going straight into my funny folder

It won't load for me right now

Hello, would you add something to this list? I think it's pretty good

Over‑polished prose – flawless grammar, overly formal tone, and excessive wordiness.

Repetitive buzzwords – phrases like “delve into,” “navigate,” “vibrant,” “comprehensive,” etc.

Lack of perspective shifts – AI usually sticks to a single narrative voice; humans naturally mix first, second, and third person.

Excessive em‑dashes – AI tends to over‑use them, breaking flow.

Anodyne, neutral stance – AI avoids strong opinions, trying to please every reader.

Human writing often contains minor errors, idiosyncratic punctuation, and a more nuanced, opinionated voice.

It's not just x, it's y

Hello, I am interested in this topic. What would you say were the tale tale signs of AI generated text for you? Apart from:

- excessive em-dashes

- useless words, verbosity

There would't be a problem if there was transparency and clear boundaries. The future is simply enjoying what you want, but we have to get there past these first steps.

Benchmarks show that open models are equal to SOTA closed ones but own experience and real world use shows the opposite. And I really wish they were closer, I run GPT-OSS 120b as a daily driver

casual workstation flex to kick off the blog

https://www.removepaywall.com/search?url=https://www.ft.com/...

my bad, but you know what I meant

Anthropic has much more funding than that. Most recent one was at $13B at the one before was at $3.5B. Now imagine that GPT recieved $40B in one round!

There is no point in using a low-bandwidth card like the B50 for AI. Attempting to use 2x or 4x cards to load a real model will result in poor performance and low generation speed. If you don’t need a larger model, use a 3060 or 2x 3060, and you’ll get significantly better performance than the B50—so much better that the higher power consumption won’t matter (70W vs. 170W for a single card). Higher VRAM wont make the card 'better for AI'.

In that case you should run a model locally, this one for example: https://huggingface.co/ds4sd/docling-models