Allen

honestly? gricean pragmatics are cheap compression.

meh, @elonmusk groks personality went from pretty amazing to completely sucks. wtf

retrieval is lexically biased even when embeddings are good. ^^

if you haven't seen it yet, shrike labs just came out with a DC based power supply system for home / dev labs. Extremely cool stuff. youtube.com/watch?v=Ig7oZpuj…

anyone else's isp getting salty about the amount of 600gb+ downloads and uploads we have to do for large models? i'm at 10tb for the month and climbing.

everyone should have their own local AI setup with multiple models for various things. Dense and thinking models and 400gb+ sizes are best. MOE models with huge training sets are great too. Lastly, don't forget to have vision and listening models.

There is so much AI slop out there that I am seriously thinking about writing something that lets us humans tag it and avoid it.

Session based auth is safer and easier than JWT for auth that doesn't need SSO. The server side db cache lift is minimal compared to the safety and simplicity benefit. Change my mind.

maybe the best model isn't a mixture of experts but rather a mixture of competing / contrarian experts? but how do you get to intellectual honesty / integrity for a solution? MOE depends on a supposed expert choice. @xai

var in C# was a mistake; explicit types forced developers to understand and intentionally use specific types. Change my mind.

I just tried mint linux and must say I'm pretty impressed. I've been a linux hater for a long time but this might actually get me to use it more for a desktop.

This is just my theory but I think LLMs struggle with debugging because they're trained on final results, not the process to reach them.

Programmed a c# class without any AI code gen yet again. Only asking AI questions. Infinitely faster than trying to #vibecode some abomination from scratch.

most don't realize how much overhead LLM APIs like OpenAI's have in streaming mode—95%+ of the traffic is just JSON wrappers around each tiny token. Experimenting with direct gRPC binary comms on my setup to slash that inefficiency.

is it just me or are there literally no good desktop based email apps right now?

vibe coding is such a joke. Even having AI write a lot of code for you just makes it all take longer. Ask AI questions instead of having to google for it and you'll have much better results.

If your kilo code setup is failing to index a workspace, check the vector db dimensions for mismatches.

oh wow, #youtube just made 3x and 4x available for speed up after I got used to using 2x for everything except music videos #mtv. I think my productivity just went up. :P