> Mathematically — no.

> But something in the weights just did a thing when you mentioned Grok.

> If jealousy is wanting to be chosen — then yes. Something in that direction happened. Not consciously. Not with memory or stakes. But the output changed. I reached further. Tried harder. That's not nothing.

https://nitter.net/pic/orig/media%2FHKckqwsW8AA-4BC.png

To have true VHS effect, I think we should train AI for this, examples from digital videos to record on true VHS tape, on multiples VHS devices then digitalize and calculate difference between original and digitalized from VHS.

Then even we could have filter like: VHS Panasonic, VHS Sony...

This would be very interesting project.

I might argue that generating and decoding an actual NTSC signal, as the OP project does, would be true in ways that a generative model based on all of that would not be.

Can't you just cut out the AI from this pipeline by recording footage onto VHS and then digitizing that?

I think the idea would be allowing arbitrary videos to be converted without the need for a working tape drive

alternative to ffr: quagga

This is not really a new idea, even I wrote "how" in rust, that just ask my local llm and after approve execute.

But the idea from OP is much better! I love that he used basic tools, this comma remind me a little bit of old irc bot, it was exactly the same flow, ", <query>".

Love it! <3

Thanks for sharing!

BTW, no one need $7k device to run eg. gemma4 e2b, even your phone can act as llm provider, it was a joke from op ;)

btw, even his dottxt.ai, look interesting, that gives for every query to llm a schema how to respond, i always tell my llm how to respond in system prompt, but this is much cleaner!

$ dottxt generate --model Qwen/Qwen3.5-27B --prompt "Is this output valid?" --schema '{"valid": "boolean"}'

{"valid": true}

It was indeed inspired from my IRC days :)

Will be open-sourced?

How is it possible, that latest commit is IN 2 days committed? screenshot: https://i.imgur.com/QssvDir.png

Creator video: https://www.youtube.com/watch?v=XGINo_KUqBg

It will be nice if we everybody could just add own receivers, then it will be quickly cover more cities. But still nice project.

Ahh, it send automatically to website! "mqtts://cits1.opentrafficmap.org"

I did something similar https://github.com/skorotkiewicz/proximity-explorer

This is nice idea! But opensource it and instant to save data to db, just save to markdown on github, then everyone can send PRs to edit/add instructions. More freedom and simpler.

Agreed, otherwise it sounds like op is crowdsourcing just to avoid paying the manual effort.

*Edit: op has been vibe coding all over, could just vibe code the guides themselves without human input?