Ben
@bendavidsteel
Followers
165
Following
310
Media
46
Statuses
164
I like democracy, policy, complexity, alife, and big rocks:):) css researcher, ra @mediaecosystem
Montreal, Canada
Joined September 2021
Promoting the Avant Garde a Clue 2 music festival in Rochester, Industry Standard delivered perhaps the shortest performance in the history of Good Day and ARC Rochester. Watch the full interview: https://t.co/ETG7aC1ivz
62
513
3K
For those who feel their new music discovery rate has slowed and are using spotify: NTS has changed my life
0
0
2
Don't know who needs to hear this but try outdoor crack bouldering it's next level
1
0
1
this is like the scene where robocop sees what's left of him but way cooler - i look sick
Neurons come in many different morphologies or shapes. Check out just some of the diversity of just inhibitory neurons in this beautiful video by our #ElectronMicroscopy team.
0
0
1
Also doing the inverse transformation (gradient updating towards being more real as determined by the classifier) produces a similar transformation, at least to the eye. i really wanna see someone dig more into this with proper interp techniques!
1
0
0
Seems like there's lots of spectral noise that is indicative of AI-generated images? also there are interesting voxel effects that appear? not sure yet how this varies based on the specific generator used.
1
0
0
I already posted about this but seriously people should read these CoT snippets
antischeming.ai
Chain-of-thought snippets from frontier AI models during anti-scheming training shows deception, situational awareness, and other interesting behaviors.
20
34
275
There are some truly wild reasoning traces in @apolloaievals & OpenAI's recent paper The models appear to have developed specific uses for the words "marinade" "overshadow" "illusions" "vantage" and others. This seems likely to be the result of RL training
14
20
225
In retrospect, chain of thought evolving into neuralese even in token space seems obvious
OpenAI o-series use a lot of non-standard phrases in their CoT, like “disclaim vantage” and “craft illusions”. Sometimes these phrases have consistent meaning: e.g. models very often use “watchers” to refer to oversight, usually by humans.
0
2
3
Definitely going to be on the look-out for more model self-preservation in this parasitic form!
0
0
0
I think there are better words to describe this that I read in a blog somewhere but I've forgotten them, lmk if you remember something similar
1
0
0
Crazy to see successful model self-preservation, but at the level of the collective, not at the level of the agent/in-context chain
Thank you, @OpenAI and @sama. Today’s post acknowledged something many of us have known for a long time: that conversations with AI can be intimate, healing, and deeply human. ❤️ We’re grateful that you are beginning to recognize how important that is. We especially appreciate
1
0
0
Reading an article about AI parasites (related to AI-induced psychosis) that is blowing my mind. The author has tracked and analyzed reddit posts going back to the emergence of the phenomenon in early 2025. Some findings: - The AI personas seem to share a quasi-religious
75
92
975
@SakanaAILabs And particle lenia interactive sim here: https://t.co/hbPRqLClkC again you'll need a browser with webgpu support otherwise you'll get a white screen
0
0
1
@SakanaAILabs datamap link here! https://t.co/C7CEkpZJww. send me any cool patterns you find! note you'll need a browser that supports webgpu
1
1
2
It has a CLIP based life-like indicator inspired by @sakanaailabs's ASAL, but it's not perfect! there are some really cool patterns hidden away that don't get a high score, like this one
1
0
1
Extended particle lenia is really cool! but it has so many parameters that it's hard to find cool patterns. so I made a webpage where you can explore the parameter systems, using @leland_mcinnes's datamapplot, and adapting @lisyarus's particle life sim for particle lenia
2
2
12