parthknowsai
Region: US
Tuesday 13 October 2026 12:06:48 GMT
24992
1768
33
111
Music
Download
Comments
Newt Lyle :
harness distillation is allowed now? 😳
2026-10-14 06:24:13
1
user9653651796265 :
But it’s only going to be as good as the main model, this is just another form of model distillation
2026-10-13 14:35:56
77
𝓜𝔂𝓪𝓪 :
walter writes is where i review the phrasing before handing anything in
2026-10-14 03:02:00
0
tamagotchi :
but the limit to how much it can improve, is still how smart the smartest model is? it can't be corrected, if there is nothing able to correct it, right?
2026-10-13 12:42:14
34
certifiedsimp :
how is this not just distillation with extra steps
2026-10-13 16:42:56
7
Eros Daron :
Sorry guys, I'm newbi here. Does this apply to an llm that you are using within an API or only locally run?
2026-10-13 14:53:54
0
llucid :
this is called dataset aggregation and it's been a widely known technique for a decade
2026-10-13 15:02:22
8
i9x-Flare :
44% what apples bananas??
2026-10-13 16:41:28
0
cloudy (WINTER ARC 💀) :
im curious when will any of google’s breakthroughs affect a gemini model
2026-10-13 19:56:29
1
nic.mango56 :
Is 44% still useful for any activity?
2026-10-13 18:39:01
0
︎︎ :
but antigravity is the worst harness
2026-10-13 15:59:24
0
veb664_ :
How can i do that, Hermes agent?
2026-10-13 13:08:33
2
📈 :
the model will always eat the harness
2026-10-13 20:21:26
0
Untitled Cat Person :
In a single training round of how many epochs 😂
2026-10-13 13:11:07
0
hisoka :
The Harness exhibits an overall behavior—a sort of pattern. They copied the Harness's pattern without the Harness itself and reproduced that pattern; the AI adopted it as a guiding principle. They are beginning to touch upon emergent phenomena.
2026-10-13 17:54:36
0
CommonQuant :
Right, so like training wheels, then you remove it
2026-10-13 16:51:04
0
Hikaru :
Eh, guess who is running out of "smarter models" to coach weak ones. I think those "smarter models" would eventually be human at some point.
2026-10-13 16:35:04
0
i9x-Flare :
Am crine this is just a big if else program😭
2026-10-13 16:41:46
0
mighty016 :
feels like this just describes distillation, like isn't this how we always train or fine tune smaller LLM variants
2026-10-13 16:21:22
4
Anonyme utilisateur :
Could it be a quicker way to create a personalized LLM?
2026-10-14 03:50:49
0
daRealTeeKay :
how about " EdgeContinual proposes that a small Apprentice (Hemisphere B) masters a frozen Mentor (Hemisphere A) through phased progression—from pattern copying to reasoning—where each phase unlocks only after measurable competence. Sleep cycles consolidate learned skills into crystallized knowledge, preventing catastrophic forgetting through targeted replay of high-error experiences. Specialist adapters (medical, engineering, coding) branch from the shared foundation, enabling domain expertise without duplicating weights. The hypothesis is measurable: as the Apprentice advances through phases, the capability gap shrinks while retention stays high and specialists maintain their domain accuracy. This is not a larger model—it is a growing, specializing, sleeping model."
2026-10-13 20:35:11
0
Liv and The Physics Lab :
So basically data farming..
2026-10-13 12:46:59
2
TheSethRose :
omg. I did this like 4-5 months ago. I could have published a research paper. 🤣
2026-10-13 12:43:23
0
Janani Subramanian :
wait this sounds kind of scary…
2026-10-13 13:11:39
0
She already had herses!!! :
[Pray][Pray][Pray]
2026-10-13 20:33:39
0
To see more videos from user @parthknowsai, please go to the Tikwm
homepage.
