@agitbackprop

electron wiggler

 ⚢
Joined January 2022
  • Tweets7.9K
  • Following1K
  • Followers1.1K
368Media
wow that's actually several orders of magnitude less penalty than i would have expected
GLADIA Research Lab
@GladiaLab7h
wanna help close a 4,655× gap? Join the Encrypted GPT-2 Speedrun 🏁 The current best stands at 3.031 s/tok. The goal is breaking the 1.0 s/tok barrier! 🧵(1/4)
1
52
It’s worthwhile to play through in your mind the kinds of motivated reasoning and self-deception that might have led academics to work on this project. It’ll help you understand better what’s going on at AGI companies.
Praveen Kumar Verma
@Alacritic_SuperOct 25
Purdue University's IDEAS Lab built an autonomous drone that tracks runners through dense forests completely without GPS or cloud networks. By running 3D LiDAR mapping and obstacle-avoidance AI entirely on-chip, it can weave under tree canopies and predict target movement through blind spots in real time. Massive implications for next-gen search & rescue and edge AI.
39
42
5
718
22,556
In general, many interp and alignment results might not hold up on current frontier architectures, and researchers should be pushing for more transparency from labs to allow for open science.
Thomas Jiralerspong
@tomjiralerspong22h
PSA to everyone doing neuron interp after the recent monosemantic neuron posts: the model's training recipe matters a lot. e.g. GPT2 seems to have been trained WITH dropout (modern LLMs are not), which seems to increase redundancy and (maybe) polysemanticity but reduce superposition (see sources in comments). Other choices probably also have an effect (e.g. SwiGLU vs GELU MLPs, optimizer, amount of weight decay, etc.) and I think we should be careful not to generalize too much. Currently working on a project to characterize the effects of these choices on neuron interpretability (a lot of the existing work is done on older/toy models which I'm not sure are representative), feel free to reach out if interested!
3
7
87
2,605
Sauers
@Sauers_22h
I found a neuron that makes the model green.
18
452
we need a term in between grug tech and gorm fluid for machine learning models that haven't seen a language token in their life
1
10
359
it is kind of insane that you can just make a general intelligence on your laptop in a few hours
1
37
723
THREAD! saw very sus ollama endpoints every endpoint analyzed advertises a model: "agent-id-chris-0928". metadata incl "owner=chris agent=hermes-mac" 365 of 375 endps report model changes within 06:54–06:57 UTC on Sep 29. server-reported timestamps suggest coordination. 1/🧵
7384254b@7384254bOct 23
some guy named chris has been using hermes agent on his mac to deploy secret stealers and malicious openai compatible completions endpoints to over 360 vulnerable VPSes, 98% of which are on oracle cloud
2
3
1
14
636
many of these are serving 27b+ qwen variants, and some are serving larger GLM and deepseek models, which could be used maliciously in addition, ollama's /api/pull endpoint is unauthenticated, so an actor could cause any model that would fit on the hardware to be downloaded
10
146
i've been monitoring a population of unsecured Ollama and llama.cpp servers on the open internet, using their public metrics endpoints, and observed a spike in model residency last night (maybe an automated scanning campaign)
5
2
1
47
1,250
girls who use astra for everything vs girls who use luna for everything
3
2
45
648
cause for hope
Apollo Research
@ApolloResearchOct 15
Replying to @ApolloResearch
GLM-5.3 reasons that a misaligned action taken during training could be reinforced and make similar actions more likely later. It also considers what kind of agent it wants to be in the future, and factors this into how it should act. We call this "update-aware reasoning".
2
2
24
998
biohacker alignment chart
5
13
3
78
3,971
why is there Whole Brain Emulation discourse
2
1
26
575
it's cool how google maps just randomly hides stops on your bus line from the map if it decides you're not zoomed in enough
1
30
819
grocery store named Market For Lemons
lumi@agitbackpropSep 28
opening a bar named Social Proof
3
4
27
1,450
the problem with the do everything machine is that it does everything
3
33
782
opening a bar named Social Proof
1
22
1,913
hypothesizing the existence of too high sensory noise floor autism and too low sensory noise floor autism
2
21
576
I do not want an eternal cosmic parent.
Quinurum@QuinurumSep 24
Replying to @zetalyrae
I think if we miraculously manage to align AGI and it basically treats us like eternal children that could work out to be fine a responsible enough AGI mother would also respect our desires for meaning and autonomy and find a way to provide that
7
1
103
4,064
long timelines gf be like "oh yeah 4 years till foom"
1
1
94
1,717
Load more