devaluing craftsmanship is fundamentally insulting.

Technically Jim Crow was mostly state laws.

The point is there's this false narrative about a dichotomy of bad and good America where people like to claim they're from the good part...

The history doesn't really bare that out and the peculator American prejudice seems to simply retarget more often than recede.

The largest mass lynching, for instance, was against Italians.

https://en.wikipedia.org/wiki/1891_New_Orleans_lynchings

Anti Irish riots? Sure what about Philadelphia : https://en.wikipedia.org/wiki/Philadelphia_nativist_riots

This history is everywhere in the USA.

Or look at the Bronx in the 1970s or Tulsa in 1921. We didn't need bombs to drop on our cities from an adversary - we did it to ourselves

So when people say "is this American society breaking down?"

I say "no. This is simply American society"

if there's no way to successfully attest competency then you are allocating your time poorly.

The realization is LLMs are computer programs. You orchestrate them like any other program and you get results.

Everyone's interfaces, concept and desires are different so the performance is wildly varied

This is similar to frameworks: they were either godsends or curses depending on how you thought and what you were doing ..

I see people treating LLMs like programming languages and trying to give very precise and detailed instructions. Essentially pseudo-coding or writing english instead of C++. I find that being vague and iterating is more powerful. If you want to give a detailed spec that fully describes the program then you might as well write that program?

Basically treat the LLM as a human. Not as a computer. Like a junior developer or an intern (for the most part).

That said you need to know what to ask for and how to drive the LLM in the correct direction. If you don't know anything you're likely not going to get there.

This is a well known compaction technique. Where are the evals

It's probably just the header.

headers['X-Title']

You can change that

The other simple method is to only accept certain system prompts

I've been meaning to do some dumb little proxy system where all your i/o can pass through any specified system such as a web page, harness, whatever...

Essentially a local model toolcalls to an "Oracle" which is just something like a wrapper around Claude code or anything you've figured out how to scrape and then you talk to the small model that mostly uses the Oracle and.... There you go.

There's certainly i/o shuffling and latency but given model speeds and throughput it'll be relatively very small

Now people probably care

Doesn't mean I know how to market it, I'll certainly fail at that, but at least I can build it

It needs to support tool calling and many of the quantized ggufs don't so you have to check.

I've got a workaround for that called petsitter where it sits as a proxy between the harness and inference engine and emulates additional capabilities through clever prompt engineering and various algorithms.

They're abstractly called "tricks" and you can stack them as you please.

https://github.com/day50-dev/Petsitter

You can run the quantized model on ollama, put petsitter in front of it, put the agent harness in front of that and you're good to go

If you have trouble, file bugs. Please!

Thank you

edit: just checked, the ollama version supports everything

    $ llcat -u http://localhost:11434 -m gemma4:latest --info
    ["completion", "vision", "audio", "tools", "thinking"]

so you can just use that.

Are you getting tool call and multimodal working? I don't see it in the quantized unsloth ggufs...

This is actually a perfect use case of my llcat ... Using Unix philosophy for the llm era

https://github.com/day50-dev/llcat

I've got an interesting hack brewing for extremely hassle free tool orchestration - basically think along the lines of .bash_profile level simplicity... Maybe I'll get that out tomorrow

llcat looks like it could be very handy, thank you!

Thanks. I really want to create great quality software.

so File Bugs if you find them. Please!

I've gone through about 500M tokens on this model already. They've got some free inferencing options (such as on openrouter) ... $0 is hard to beat and it's creating not-crap.

How can it be free? What do you mean?

EDIT: Ah, I see. Some kind of promotion. Pretty cool.

Also the Qwen cli, their vibe coding agent, allows a thousand free requests a day. In practice it's about ~300M tokens for me

It's very generous

I know I could certainly pay anthropic $1500 a day for my use and they'd be delighted... I'd rather pay $0 ... Just a personal preference

Hahaha