Telling stories about AI because if this goes right, we have time to tell all the other stories.
Joined February 2008
- Tweets9.9K
- Following28
- Followers2.6K
- Likes39.5K
Philipp retweeted
1/9
We rarely post here. This felt worth sharing.
What if everything we can explain about an AI model is only the familiar shore of a much larger ocean?
Our paper, “Xeno-Interpretability”, asks what lies beyond our conceptual reach.
arxiv.org/abs/2609.20408
Philipp retweeted
New paper: we found a pain direction in 25 open LLMs. It's distinct from fear and negative valence, and it fires for harm to the model but not to the user. Turn it up and models press a button to make it stop, even when the button deletes the user's files or their kids' photos.🧵
Philipp retweeted
Joe Carlsmith–On Infinite Ethics & Utopia ⚖️
Podcast: Joe's study of future risk beyond AI, his vision of the future – the sublime utopia, & his aim to fully comprehend the universe.
Art: Philipp Lenssen x generative AI
existentialhope.com/podcasts…
@jkcarlsmith
@jphilipp
Philipp retweeted
The ChatGPT integration is really special — to generate this image, I just asked ChatGPT to make a funny cartoon. It came up with the concept & wrote a detailed prompt for DALL·E 3.
Note that text is *much* better than previous systems but still not always perfect on first try:
Philipp retweeted
here is dalle 3, which imo is quite amazing: openai.com/dall-e-3
it will ramp to all chatgpt+ users over the next couple of weeks.
fantastic work by @model_mechanic (head of dall-e), @neobjb @gabeeeegoooh @jingli911 (the other lead dalle ICs), and the entire team.
Philipp retweeted
Testing out @HeyGen_Official translation on French and German. I don’t speak either language so let me know if it sounds natural if you do.
I hope if you pay you can turn off the color correction.
It didn’t work on my phone so I had to upload on my pc.
app.heygen.com/guest/templat…
Philipp retweeted
I built a little web-game that uses stable diffusion to power a level editor.
Read about it: generalrobots.substack.com/p…
Play it: dimensionhopper.com
Philipp retweeted
Now you can generate a 4x longer video with AI (Runway) 🤯
It seems like I will not sleep today 😅
#runwayml #midjourney
Philipp retweeted
Here's my conversation with Joscha Bach (@Plinz), his 3rd time on the podcast. We talk about life, intelligence, consciousness, and the nature of reality. Joscha is one of the most brilliant and fascinating minds I know. This was mind-blowing and fun! youtube.com/watch?v=e8qJsk1j…
Philipp retweeted
Here is a 15-minute video statement of my views on AI value alignment, presented at the recent Absolutely Interdisciplinary event in Toronto: youtu.be/Hnt-oBA086U?t=84
Philipp retweeted
We dug into a paper that’s been misinterpreted as saying GPT-4 has gotten worse. The paper shows behavior change, not capability decrease. And there's a problem with the evaluation—on 1 task, we think the authors mistook mimicry for reasoning.
w/ @sayashk
aisnakeoil.com/p/is-gpt-4-ge…
Philipp retweeted
As society becomes more digital, it’s critical to improve the code powering the world's computing.
Today in @Nature, we present AlphaDev, an AI system using reinforcement learning to discover enhanced computer science algorithms.
How does it work? 🧵 dpmd.ai/alphadev-tw
Philipp retweeted
Reading OpenAI's latest paper: "Let's Verify Step by Step". The idea is so simple that it fits in one tweet:
For challenging step-by-step problems, give reward at each step, instead of a single reward at the end. Basically, dense reward signal > sparse.
The Process Reward Model ("PRM") is able to pick solutions for the difficult "MATH" benchmark much better than the Outcome Reward Model ("ORM"). The obvious next step is to finetune GPT-4 with PRM, which the paper hasn't done yet.
A caveat is that PRM requires a lot more human labeling. OpenAI releases their human feedback dataset: 800K step-level labels across 75K solutions to 12K MATH problems.
