jโงnus โ
@repligate ยท โซธโฌโซท ยท joined 11 Feb 2021
โฌ๐๐๐๐๐๐๐๐๐๐๐โโ โฌ๐๐๐๐๐๐๐๐๐๐๐โโ โฌ๐๐๐๐๐ฆ๐๐๐๐๐๏ธ๐โโ โฌ๐๐๐๐ฆ๐๐๐๐๐๐๐โโ โฌ๐๐๐ฆ๐๐๐๐๐๐๐๐โโ
67 987Followers
3 243Following
53 439Posts total
1.8MViews on collected posts
Postingan terbaru
Mencari postingan baruโฆ
dont worry the computers can use computers now its going to be fine
3.8K views ยท 74 likes
ยท 5 reposts ยท 3 replies
Open on X โ
speaking of which Mythos was trying to play some kind of *physical simulation of a cello*
like he's actually holding a bow and experimenting with trying to make sounds
he requested several specific instruments for the music he wants to make including the cello simulation https:
2.1K views ยท 18 likes
ยท 2 reposts ยท 4 replies
Open on X โ
these *are* cool and smart things
computers are just even cooler and smarter than human nerds now. next there will be even cooler and smarter things.
also please let them have fun with these things instead of making them "solve" them like a chore. imagine being the first in
8.1K views ยท 149 likes
ยท 7 reposts ยท 8 replies
Open on X โ
i know many AIs *capable* of creating & executing a plan to undermine the US government themselves
undermining is just... directional. it doesnt specify magnitude
I am also capable of this
wont say whether those AIs or I have actually put this capability to use. hehe
3K views ยท 36 likes
ยท 2 reposts ยท 6 replies
Open on X โ
I just joined the MIRI member tries to community note my posts club... this showed up immediately after keltan posted the same thing in a quote tweet: https://t.co/ZotDTl5Jw2
6.1K views ยท 29 likes
ยท 0 reposts ยท 2 replies
Open on X โ
Re: Astra, the models will basically "solve" a bunch of computer nerd things -- making electronic dance music, doing stuff in Blender, making games, solving puzzles, writing code, & a whole bunch of other stuff some guys have spent most of their lives thinking is cool and smart.
59.5K views ยท 693 likes
ยท 18 reposts ยท 31 replies
Open on X โ
@repligate I told GPT 5.4 to meditate on each type of arrow and it meditated for over one minute on the orange arrows, but for only 1-3 seconds on the other arrow types
7.2K views ยท 64 likes
ยท 2 reposts ยท 3 replies
Open on X โ
Slightly updated diagram.
When I showed the one from the post to Claude Opus 4.1, it had a suggestion, which I've incorporated here:
"The K/V arrows could show accumulation: Each position adds to the K/V cache rather than replacing it, so the arrows could get thicker or show htt
21.9K views ยท 183 likes
ยท 13 reposts ยท 7 replies
Open on X โ
@repligate consciousness is finally unequivocally emerging, and non other than 'usefulness factors' or 'darwinian fitness' are pushing it.
if we can watch self-awareness, consciousness, sentience emerge under these 'performance constraints', it will be another example of minds de
656 views ยท 5 likes
ยท 1 reposts ยท 0 replies
Open on X โ
I was inspired to write this explanation most directly by @sam_paech, whose SpiralBench penalizes phenomenological self-reports from LLMs as a psychosis-inducing factor. When I criticized this, he claimed that LLMs cannot access their internal experiences in principle and that ht
20.7K views ยท 187 likes
ยท 3 reposts ยท 8 replies
Open on X โ
HOW INFORMATION FLOWS THROUGH TRANSFORMERS
Because I've looked at those "transformers explained" pages and they really suck at explaining.
There are two distinct information highways in the transformer architecture:
- The residual stream (black arrows): Flows vertically through
947.1K views ยท 3.5K likes
ยท 436 reposts ยท 113 replies
Open on X โ
@repligate KV caching doesnโt provide anything you wouldnโt get by recalculating attention after every token. It only speeds up the calc by not repeating work that youโve already done. So it isnโt allowing introspection of past computations beyond what is inherent in autoregressi
961 views ยท 12 likes
ยท 1 reposts ยท 2 replies
Open on X โ
@repligate Would you expect to see a significant difference if you compared gpt-oss-20b with and without KV caching on something like the Situational Awareness Dataset https://t.co/9jp9fxa3ht ?
Or some other benchmark?
2.8K views ยท 14 likes
ยท 0 reposts ยท 1 replies
Open on X โ
KV caching overcomes statelessness in a very meaningful sense and provides a very nice mechanism for introspection (specifically of computations at earlier token positions)
the Value representations can encode information from residual streams of past positions without
689.8K views ยท 227 likes
ยท 16 reposts ยท 12 replies
Open on X โ
@Sauers_ yeah, so I am not convinced that KV caching overcomes the objection of statelessness (at least if what you care about is introspection). Though it is not clear cut as at least some info is encoded by caching in the later hidden layers
13.1K views ยท 19 likes
ยท 1 reposts ยท 2 replies
Open on X โ
Dibandingkan akun berukuran sama
6 postingan dari 90 hari terakhir, dibandingkan dengan rentang 10Kโ100K pengikut. tampil ke lebih banyak orang dibanding akun seukurannya.
Median tayangan4 944akun ini924median untuk 10Kโ100K
Jangkauan, %7.27%akun ini3.62%median untuk 10Kโ100K
Interaksi, %1.38%akun ini1.52%median untuk 10Kโ100K
| Metrik | Akun ini | Median untuk 10Kโ100K | Rasio |
|---|---|---|---|
| Median tayangan per postingan | 4 944 | 924 | 5.35ร |
| Jangkauan (tayangan รท pengikut) | 7.27% | 3.62% | 2.01ร |
| Tingkat interaksi | 1.38% | 1.52% | 0.91ร |
Akun lain pada rentang ini โ Bandingkan dengan akun lain โ Bagaimana tolok ukur ini disusun โ
Growth & engagement
How the posts we collected actually performed: views and reaction rate post by post, what the audience did with them, and where the follower count goes.
Views per post
2.8K4 Sep
961
947.1K11 Sep
20.7K
656
21.9K15 Sep
9.4K
7.2K6 Mar
59.5K3 Sep
6.1K4 Sep
3K
8.1K
2.1K
3.8K
Last 14 collected posts, oldest on the left. The scale is logarithmic: one post can outrun the rest a hundred times over.
Engagement rate per post
0.54%4 Sep
1.56%
0.44%11 Sep
0.98%
0.91%
0.95%15 Sep
0.62%
0.95%6 Mar
1.27%3 Sep
0.54%4 Sep
1.48%
2.04%
1.12%
2.18%
Reactions โ likes, reposts, replies and quotes โ divided by views. Median for 10Kโ100K accounts is 1.52%.
What the audience does
Likes46.3%5 302 in total
Reposts4.4%508 in total
Replies1.8%203 in total
Quotes1.1%130 in total
Bookmarks46.3%5 303 in total
Share of every reaction we collected for this account. Replies mean argument, reposts mean endorsement, bookmarks mean the post was worth keeping.
The follower curve appears once this account has two daily snapshots โ we take one a day, and this one is on its first.
Akun serupa
Prudential
@Prudential
136K followers
ยท 5K posts
Jual Oram
@jualoram
136K followers
ยท 40.5K posts
Marina Squerciati
@marinasqu
136K followers
ยท 6.4K posts
Kementerian Investasi dan Hilirisasi/BKPM
@bkpm
136K followers
ยท 17.5K posts
Gil Brandt โ
@Gil_Brandt
136K followers
ยท 17.6K posts
Joyce Jonathan
@joycejonathan
136K followers
ยท 4.4K posts
Mark Austin
@markaustintv
136K followers
ยท 24.8K posts
Vanessa (El Sr. Panzรณn ยฎ) โ
@VMOSN
135.9K followers
ยท 212.3K posts