Kye

Which is to say, tuning fights with prompts less as the models and interfaces improve.

I'm not talking about using a default mode LLM with LinkedIn Standard Obsequious Bullshit as a conversational imperative that emerges from simple prompts interacting with the heaviest weights. It pushes back because I told it to and it has redirects around common LLM failure modes, and modes unique to how I use them. That's in a set of instructions I've had a bunch of different models tear apart so I could put it back together better.

I treat it and describe it as a language coprocessor, not a buddy. The instructions are the kernel I boot it with.

For writing as thinking with trouble starting from scratch, LLMs are the most important technology to emerge in my lifetime. Microblogging filled that gap in a way, but it had too many downsides.

My handy heuristic for headlines like this: Is it a scary new trend that means something or did other factors suppressing its natural emergence decline? Or is it a matter of observation?

A recent real-world example was the detection of two different objects entering the solar system. The naive speculation was "they came on the same plane, so they must be alien!" But the reality is more mundane: the new detection method that found them, while flexible, started by looking at that plane. So of course both objects it detected were on that plane.

>> "How would you organize these LLM quirks, ontologically speaking? I have this notion that the better path is to identify what kinds of things are emerging and prompt to do those things better; accept it as something LLMs are going to do and treat it as something to improve on instead of something to eliminate."

The output is a bit better on blind prompting with applying the results. Here's the gist:

1. Compression artifacts — the model encoding structure implicitly

2. Attention-economy mimicry — the model trained on engagement-optimized writing

3. False epistemic confidence — the model performing knowledge it doesn't have

4. Affective prosthetics — the model simulating emotional register it can't inhabit

5. Mechanical coherence substitutes — the model managing the problem of continuity

Spot corrections are too spotty. Going higher levels with these kinds of problems seems to work better.

It's neat that some websites communicate on AP, but every single one compounds the problems AT attempts to solve. You talk about software playing nice with Mastodon while AT is converging on standard lexicons and Cisco is investing in changing the nature of the relay for the better:

https://www.ietf.org/archive/id/draft-nandakumar-atproto-ato...

https://bsky.app/profile/ipv.sx/post/3mg4h4tncmk2l

The usual pattern is like that Alex Norris comic:

"you do not fit in here"

"okay we will make our own place"

"why are you excluding us"

"oh no"

Except here they can stay connected with the broader Atmosphere but engage on their own terms.

The toggle by "Your past chats with Gemini"

Squishy brain heuristics can't last long enough to matter in this environment. Personally, I created a Claude skill to run this query (with some refinements) and check it against an article I suspect of being lazy AI writing. If it's good AI-supported writing, I probably won't suspect it, and I won't care if it is.

The people trying to fool you with lazy writing run the same list on their outputs to have the LLM fix it.

"Here's my response written in a stylized way that will appeal to highly technical readers. Is there anything else I can help you with?"

Interesting piece though.

The specific way it applies to your specific situation, if it exists, either hasn't been found or hasn't made its way to you. It really is early days.

There might be some insight in there as to why they invested knowing lock-in wasn't on the table.

https://ufos.microcosm.blue/collection/?nsid=sh.tangled.repo

It's not like GitHub where all the important stuff is locked up on GitHub. Everything here is readable and portable.