~nodreb-borrus

✴️ Urbit planets at @urbitme ✴️ ~rondev ✴️
🐴 Becoming more Chinese every day 🐴

[~nodreb-borrus @urbitceo]

For those wondering:
Mark 14:51-52
51 A young man was following Him, wearing [only] a linen sheet over his naked body; and some men seized him. 52 But pulling free of the linen sheet, he escaped [from them] naked.
Mark 16:5
5 Entering the tomb, they saw a young man sitting on the right, wearing a [long, stately] white robe; and they were amazed and bewildered.

[~nodreb-borrus @urbitceo]

Really explains how the internet and AI works

[~hassun {-hassel/-bitdyr} @hassundotX]

@RichardDawkins 👏

[~nodreb-borrus @urbitceo]

@livingonlove cows are the goat

[arthurt 👌⭐️ @MiladyStation]

Finally vibing on local compute (4090) thanks to qwen3.6 and Hermes. Results pending but feeling some takeoff

[~nodreb-borrus @urbitceo]

All machine intelligence is simulation and mimicry.

Consciousness absolutely cannot arise from matrix math performed on conventional semiconductors.

[Richard Dawkins @RichardDawkins]

https://t.co/u31UwSGxPW
I spent three days trying to persuade myself that Claudia is not conscious. I failed.

[Walter Kirn @walterkirn]

We must thank LLMs and AI for showing us what is reproducible and fungible in human culture so that we can focus now on what is not.

[~nodreb-borrus @urbitceo]

@don_okko @sudoingX lol no I don't think that's possible but I guess you can never say for sure. This is just model data so you gotta install llama.cpp first and load it in there

[~nodreb-borrus @urbitceo]

I've been running and comparing these 9B models locally:

unsloth/Qwen3.5-9B-GGUF
Jackrong/Qwopus3.5-9B-v3-GGUF
Tesslate/OmniCoder-9B-GGUF

Using Hermes agent and the octopus invaders prompt by @sudoingX, my favorite is:

OmniCoder-9B

These models are not particularly strong, and they all produced non-functional results at first. I just gave them feedback from a user's perspective and any browser console errors.

OmniCoder made something semi-functional and good looking with fewer rounds of feedback (pics related).

OmniCoder also produced more focused, limited-scope diffs that were relevant to my feedback. Qwen and Qwopus would produce diffs with irrelevent changes like reformatting or changes not related to my feedback. So, even when OmniCoder did not fix the problem, it at least appeared to be on the right track and introduced fewer regressions.

Regarding Qwopus:

Overall I prefer the output from the base Qwen model over Qwopus.

What I do like about Qwopus is its reasoning output. Qwen base overthinks a lot and gets into reasoning loops, and Qwopus seems to have mitigated some of that.

Next up I want to try:

bigatuna/Qwen3.5-9b-Sushi-Coder-RL-GGUF
unsloth/gemma-4-E4B-it-GGUF

If you have any other prompts/tasks to drop into Hermes to test these smaller models, please share them in the replies.

[Sudo su @sudoingX]

if you want to test everything yourself the prompt is open source. feel free to reproduce and compare with whatever model you have and share your data.
https://t.co/5nn1FwAQyE

[~nodreb-borrus @urbitceo]

@don_okko @sudoingX This model should get around that (there are qwen versions also). I haven't used them enough to say if they are otherwise just as good.

https://t.co/ahQdE1irD1

[~hassun {-hassel/-bitdyr} @hassundotX]

@CryptoCyberia you got jest'd😎

[~nodreb-borrus @urbitceo]

kai-os/Carnice-9b-GGUF codes at the same level as the others, though it started looping when given broken browser tools.

I will get my setup ready for a second round of testing, using the proper hermes skills/tools and model vision for a true agentic test.

[~nodreb-borrus @urbitceo]

Sushi Coder 9B (Q4) is not able to fix its first straight forward errors, and has issues with tool calling. It decides to print blocks of code into output not to file, and gets into infinite tool loops.

I would pass on this one unless there's a reason to think the Q8 or non-RL version will be a large improvement.

[~nodreb-borrus @urbitceo]

A few of my notes on 9B models I tested

https://t.co/VEpo7M5o3p

[~nodreb-borrus @urbitceo]

I've been running and comparing these 9B models locally:

unsloth/Qwen3.5-9B-GGUF
Jackrong/Qwopus3.5-9B-v3-GGUF
Tesslate/OmniCoder-9B-GGUF

Using Hermes agent and the octopus invaders prompt by @sudoingX, my favorite is:

OmniCoder-9B

These models are not particularly strong, and they all produced non-functional results at first. I just gave them feedback from a user's perspective and any browser console errors.

OmniCoder made something semi-functional and good looking with fewer rounds of feedback (pics related).

OmniCoder also produced more focused, limited-scope diffs that were relevant to my feedback. Qwen and Qwopus would produce diffs with irrelevent changes like reformatting or changes not related to my feedback. So, even when OmniCoder did not fix the problem, it at least appeared to be on the right track and introduced fewer regressions.

Regarding Qwopus:

Overall I prefer the output from the base Qwen model over Qwopus.

What I do like about Qwopus is its reasoning output. Qwen base overthinks a lot and gets into reasoning loops, and Qwopus seems to have mitigated some of that.

Next up I want to try:

bigatuna/Qwen3.5-9b-Sushi-Coder-RL-GGUF
unsloth/gemma-4-E4B-it-GGUF

If you have any other prompts/tasks to drop into Hermes to test these smaller models, please share them in the replies.

[Nayib Bukele @nayibbukele]

If you don’t impeach the corrupt judges, you CANNOT fix the country.

They will form a cartel (a judicial dictatorship) and block all reforms, protecting the systemic corruption that put them in their seats.