@parthknowsai: Attention residuals a new research paper showing how neural nets could pull only the important information forward #Tech #ai #fyp #learnontiktok #news
I do this on my own, especially with Claude and at the end of a session or at the end of the day. I tell Claude to summarize everything into a PDF and word document which I then downloaded and save. And I tell it to save that into long-term memory. This has been very effective as I’ve been working on a project for weeks that has not lost any memory.
2026-03-20 13:24:10
27
Dan Neuman 🇨🇦💚🇪🇺🇲🇽 :
For comparison, our cerebral cortex has 6 layers.
2026-03-20 19:29:16
13
zu :
Ur vids are very helpful to me bro, it inspired me to create lots of llm engineering related things
2026-03-20 12:40:49
28
Giorgi Ba :
residual NN is old concept
2026-03-20 20:15:24
6
Quran Every Minute :
AI is finishing the 'last mile' of work-related skills, but the scaling funnel is narrowing. We’re hitting a wall where more data ≠ better reasoning. The next pivot isn't more compute; it’s mimicking the biological 'stack'—how the human brain piles data, processes emotional noise, and reasons through complex, messy inputs. We’re moving from mimicking output to mimicking the biological process
2026-03-20 12:56:25
14
tiktok :
This means the end for model expert reduction tricks like Cerebra’s’ “REAP”
2026-03-31 15:35:00
3
Matthew Fishman :
We need the same thing for context engineering
2026-03-22 15:38:13
3
Melodio Brando :
2026-03-20 13:11:17
9
Tamroc :
How long do we have to for this? Can we train existing ones with a prompt?
2026-03-23 06:23:14
1
Mike Marrotte :
Been thinking about this for 3 years… kinda surprised it’s taken this long to get here
2026-03-21 16:27:08
3
Geekonomicon :
How do they decide what's important? Is that hard coded, or is it learned behaviour from prior training data? Either way it's only as good as the filter implemented.
2026-03-20 20:30:57
1
glenrun🇿🇼 :
yesterday I saw a article saying that Nvidia managed to solve issue of memory on llms they used the same technic that used codecs where music files are compressed without loosing quality they are saying a normal computer will able to run 70billuon parameter without an issue please check it out
2026-03-20 14:18:43
5
Rob2079🇨🇦 :
Ooooh that's extremely interesting. going to go read that paper. You always find some good stuff!
2026-03-20 13:06:31
3
BowTiedCyber :
This is actually something I’ve been working on. I’ve been making custom layers for my bots with code wrappers.
2026-03-22 00:16:59
1
Daft Citizen :
so they keep stacking 2d Layers instead of Intercompressional reverse refinement into 4D - 3x3x3 topological Mobius Stacks? No wonder our data Centers Drink more Water than my mother in law.....
2026-03-22 06:25:09
2
Vittorio Zamparella :
Machine learning had big breatroughs, but now they are back at tinkering by hand every little 1% of increased performance. They are scraping the bottom.
2026-03-22 12:57:39
1
user2633740932859 :
The memory is the second thing to go...
2026-03-20 21:08:24
1
user4517038289519 :
brilliant
2026-03-20 12:32:55
0
A̶D̶A̶M̶ :
is this an ad?
2026-04-19 03:02:34
0
Dr_EmmettBrown NO FOLLOWERS :
Ir you remove a hidden layer, doesn‘t that disconnect the inputs from the outputs?
2026-03-23 13:13:52
0
bartabee :
the state of the art models are obviously way ahead of this.
2026-04-02 00:32:24
0
666bbb.0 :
Architecture 🔥
2026-03-27 23:54:53
0
TrueSPAC Web3 Crypto :
That was a basic and practical approach.
2026-04-02 16:00:31
0
H :
So basically LSTM? 😂
2026-03-26 22:16:36
0
aka.mtv :
thanks
2026-03-20 12:50:19
0
To see more videos from user @parthknowsai, please go to the Tikwm
homepage.