justinbuilds.mov
Region: US
Monday 17 August 2026 06:50:54 GMT
21950
1062
62
28
Music
Download
Comments
.jose.a.s :
"current models like Qwen" bro I guarantee these were Qwen architecture models trained from a random init
2026-08-19 22:23:08
0
Duamutef :
Too be fair "emergence" rarely happens in people either. Few people ever manage to come up with something entirely new.
2026-08-17 18:24:32
34
adrianlivinginthereal :
So in short , the models are simply as good as their training data and nothing more… who would have known 😅
2026-08-17 15:36:03
62
Bladeey :
that’s a good idea but 10B is still ~nothing, no?
2026-08-17 23:51:08
1
notmottsapplesauce :
emergence? nobody intelligent expects emergence
2026-08-17 15:41:08
8
user174959592287 :
true emergent with MoE requires 26B sparse or 14B dense there’s a paper on this somewhere
2026-08-17 09:13:51
5
__mattex :
please stop moving so much
2026-08-17 20:39:41
4
Ruben_Kelevra :
I mean obviously?! If you put a 5 grader on a college math test, it will also fail, because the concepts are not taught, *before* you reach college. You basicly got no understanding of the principles you need to complete the task, and no way to research it. The right test would be, to give the model access to more stuff and select it's training data based on the task, and see if it complete the task after that.
2026-08-17 12:34:39
12
eyeswideopen :
Interesting!
2026-08-17 09:57:50
1
225/226 :
its to build an ai for the 80%
2026-08-17 18:33:12
3
AlphaTauri :
ok, not emergence, but after giving the 6th grade curriculum and putting it in context could it solve those types of problems?
2026-08-18 06:18:09
0
λ :
Well basic education is very focused on Application rather than understanding. Since the basic axioms aren’t taught, it only makes sense that it cannot reason with them. Unless it somehow reverse engineer them
2026-08-17 17:17:37
1
Matt Stueck on X instagram :
LOVE
2026-08-18 01:43:07
0
Ten Vi :
Learning requires somewhere for new information to become available, enough pre-existing conceptual machinery to interpret it, and enough state/retention for that information to remain usable for the relevant task. It does not inherently require immediate retraining. Where retraining becomes relevant is durability and capability formation, not whether any learning-like acquisition can occur at all.
2026-08-18 00:42:36
1
ate50eggs :
it would be interesting to add smells. each smell is basically one or more tokens worth of chemical information. 2nd grade teachers, back me up on this
2026-08-18 03:48:36
1
James :
‘Yeah bro we just added the Trump education RL it’s working well’
2026-08-17 17:15:54
1
the_untidy_john :
Omg the bots. You deserve better. Great content, btw. 🙏
2026-08-17 09:24:21
0
Jim :
kinda points at education gap between lower grade school conceptual understandings and what is required for high school. we all know the slope is not smooth and curriculum creates cognitive leaps to be made
2026-08-17 18:21:06
0
offline :
thoughts on yuval noah harari?
2026-08-17 07:17:04
0
leonov#1224 :
круто
2026-08-17 07:07:05
0
schultZ :
bro fucking OBVIOUSLY, the max amount of params is 5 billion, its basically useless
2026-08-17 16:42:19
1
ggggyyhht :
They made a model that’s at the intelligence level of the average American, that’s awesome.
2026-08-17 23:16:07
0
War Badger :
1. I’d bet $50 it wrote better English than fucking Claude opus 5, Claudelish which seems to be only capable of lazy code now; 2. 88b paramaters isn’t big enough for novel COT we’ve known this a long time much less emergence; 3. Most really cool behavior is post training now and idk if we heard any of that in what you said but doesn’t sound like it; 4. It’s a confounded study because large models “filtered” the dataset and if you’ve ever done something like this with even the best large models you’ll know how wildly inconsistently it does so even with the most rigorous methodology. It’s a complete dice roll of what exactly it obliterated below the threshold; 5. If the authors used the word emergence relative to a sub 100b parameter model they don’t understand this field at all and this was some shitty grad project.
2026-08-18 09:18:17
0
To see more videos from user @justinbuilds.mov, please go to the Tikwm
homepage.
