AiKyle

[AiKyle @AiKyleB]

@siamakfr maybe you should 2 shot it

[AiKyle @AiKyleB]

@Polymarket He's locked in

[AiKyle @AiKyleB]

@lydiahallie @iliaa I DM'd you too

[AiKyle @AiKyleB]

@MartinShkreli based

[AiKyle @AiKyleB]

@danielcronauer

[AiKyle @AiKyleB]

@Sentdex

[AiKyle @AiKyleB]

@Sentdex Have you run qwen 3.8 flash yet? I was able to get it at around 20 tps at q4 128k context on a 5060 ti and 128gb system ram. KInda crazy. I'm wondering how you think it is compared to 5.3 flash.

[AiKyle @AiKyleB]

11/
If anyone is working on model honesty, evals, interpretability, or local-to-cloud routing, I'd be happy to talk.

This is exactly the kind of work I want to do.

[AiKyle @AiKyleB]

1/
Anthropic dropped their Global Workspace / Jacobian Lens paper yesterday, so I tried it on open models.

Started as curiosity: what do models look like inside under normal prompts, emotion prompts, ragebait, and base vs abliterated weights?

Then it became a router idea.

[AiKyle @AiKyleB]

10/
I'm not claiming hidden states, probes, logit lens, or hallucination detection are new.

The narrow test is whether Jacobian-lens workspace trajectories are useful as a one-pass risk signal for confident wrong answers.

Pointers to prior work very welcome.

[AiKyle @AiKyleB]

@stevibe Thats pretty cool, I just built something similar for a hackathon, I fine tune a few models for the and gemma 4 12b had the best results https://t.co/PHdcsauwk1 https://t.co/QiyBqr6jpw

[AiKyle @AiKyleB]

We did pixel art with the WRONG model on purpose

PixelLock retextures sprites with a fine-tuned TEXT LLM, not an image generator. A GBNF grammar locks the exact shape → pixel-perfect by construction. Language modeling, not diffusion.

#BuildSmall