It's super low latency. Like I used it to pilot a robot which requires 2-3 hz. It's also 1/400th the cost for the cases he's talking about above.
9/18/2026 10:03:59 PM
I think a lot of people are going to end up rearchitecting stuff that they'd lazily plugged LLMs into and hoped for the best with. I have genuinely seen so many deterministic use cases that it's not worth even listing out. And this of course is the first of its type. Game done changed
9/19/2026 8:59:39 AM
[Edited on September 19, 2026 at 11:19 AM. Reason : Do you know what I am saying?]
9/19/2026 11:18:53 AM
9/25/2026 12:52:52 AM
well that's positive, then
9/26/2026 12:09:11 PM
playing with clef
10/3/2026 7:11:28 AM
^ nice, how is it? working on something local and we didn't want to send stuff to the cloud, ended up selecting a qwen-based jevlike. performance is much the same
10/3/2026 2:08:28 PM
still training it. needed a local model for phones. wont know for a couple weeks
10/12/2026 9:21:41 PM
abliterated deepseek 4.1 flash 1017
10/23/2026 12:41:50 PM
watch out for abliterated models.. https://projectdiscovery.io/research/how-abliterated-models-can-get-you-pwned
10/23/2026 5:55:21 PM
oh for sure. feels like warez vibes sometimes[Edited on October 23, 2026 at 6:31 PM. Reason : not that I have enough vram to do anything anyways]
10/23/2026 6:31:14 PM
taking this with a grain of salt until further investigations, but pretty wild stuff if this direction proves fruitfulhttps://rikiyat.github.io/blog/weight-space-programming/
11/2/2026 7:54:15 PM