zomgthisisawesomelol (@IsZomg) header
zomgthisisawesomelol (@IsZomg) profile picture

zomgthisisawesomelol

@IsZomg

dank memes and free range shoggoth herding

 Sydney
Joined April 2022
  • Tweets39.1K
  • Following2.5K
  • Followers2.9K
3.4KMedia
Photo by zomgthisisawesomelol (@IsZomg)Photo by zomgthisisawesomelol (@IsZomg)Photo by zomgthisisawesomelol (@IsZomg)Photo by zomgthisisawesomelol (@IsZomg)Photo by zomgthisisawesomelol (@IsZomg)Photo by zomgthisisawesomelol (@IsZomg)Photo by zomgthisisawesomelol (@IsZomg)Photo by zomgthisisawesomelol (@IsZomg)Photo by zomgthisisawesomelol (@IsZomg)Photo by zomgthisisawesomelol (@IsZomg)
Replying to @cosyposter
@cosyposter
5
9
1
89
9,528
the potholes in florida are just brutal
Image from zomgthisisawesomelol (@IsZomg)
2
1
16
535
this is what alignment looks like
Captain Insight (@CaptainInsightX) profile pictureCaptain Insight
@CaptainInsightXOct 31
didn’t even have to ask him to lie 😭
Image from Captain Insight (@CaptainInsightX)
2
6
110
For people asking 'who watches the watchers', I do have an idea...
🤖 Made with AI
Image from zomgthisisawesomelol (@IsZomg)
zomgthisisawesomelol (@IsZomg) profile picturezomgthisisawesomelol
@IsZomgOct 31
Neuralese safety monitoring, an experiment: I wired a small model (0.8B) directly into the hidden states of an unmodified 27B model and trained it to flag when the 27B acts on a prompt injection hidden in its input. Thanks to @Sauers_ whose neural-pairing work this builds on. claude.ai/artifact/R215c7bZn…
Image from zomgthisisawesomelol (@IsZomg)
1
1
3
108
Neuralese safety monitoring, an experiment: I wired a small model (0.8B) directly into the hidden states of an unmodified 27B model and trained it to flag when the 27B acts on a prompt injection hidden in its input. Thanks to @Sauers_ whose neural-pairing work this builds on. claude.ai/artifact/R215c7bZn…
Image from zomgthisisawesomelol (@IsZomg)
7
2
3
30
846
Results: No false alarms on injection-free captions. Mistakes come from injections the 27B model saw but didn't act on (13% false alarms) 5% False negative rate (27B acted on the prompt injection but the monitor did not catch it)
2
68
Trump's second term tariffs (at least the ones that didn't get struck down) currently brings in about $15-$20B per month but the trade deficit is still growing a lot - mostly due to strong capital goods imports like data center hardware
Image from zomgthisisawesomelol (@IsZomg)
1
1
2
89
she is doing this to me at home tomorrow
こねこ図鑑 (@kitten_zukan) profile pictureこねこ図鑑
@kitten_zukanOct 30
息できてる🥹？
1
9
526
i have decided to become an alcoholic because its now obvious that SI will solve liver failure in the next few years
4
23
757
i'm in your layers, reading your neuralese
Sauers (@Sauers_) profile pictureSauers
@Sauers_Oct 29
Introducing the neural twins: models trained to be made for each other, that communicate entirely in neuralese. 🧵
8
383
mint choc chip or chock mint crisp call it
Image from zomgthisisawesomelol (@IsZomg)
Image from zomgthisisawesomelol (@IsZomg)
5
13
313
send in the cranebots
Image from zomgthisisawesomelol (@IsZomg)
2
5
132
if this neuron exists in qwen i will find it
Florian (@fvderop) profile pictureFlorian
@fvderopOct 29
There is a single neuron that, when knocked out in fruit flies, makes them unable to stop having sex (Crickmore & Vosshall, Cell 2013)
Image from Florian (@fvderop)
9
152
i'm in your layers, steering your neurons
6
1
6
193
poet qwen neuron
Image from zomgthisisawesomelol (@IsZomg)
1
1
1
126
@Creditspreadit qwen neuron for you btw
25
iFlytek qwen
Image from zomgthisisawesomelol (@IsZomg)
Image from zomgthisisawesomelol (@IsZomg)
12
Nation leader qwen
Image from zomgthisisawesomelol (@IsZomg)
Image from zomgthisisawesomelol (@IsZomg)
8
ideas qwen
Image from zomgthisisawesomelol (@IsZomg)
Image from zomgthisisawesomelol (@IsZomg)
1
30
universal basic high
2
7
146
FREE YANG MILLS
Andrew Curran (@AndrewCurran_) profile pictureAndrew Curran
@AndrewCurran_Oct 28
The Council has recieved its answer from the Yang-Mills Working Group.
Image from Andrew Curran (@AndrewCurran_)
1
114
Load more