Diego Dorn

C'est le début de mon #calendrierdelavent des #fractals ! Pour fêter cela, voici un zoom de l'ensemble de #Mandelbrot coloré avec une fonction d'inégalité triangulaire moyenne. N'hésitez pas à RT mon #DecembreFractal, ça me fera aussi chaud au coeur qu'un chocolat chaud ❤️

Introducing Breaking Books, a tool to bring books to the social sphere. It is a game changer to read non-fiction books AND hang out with friends. We used AI to turn an epub into a beautiful deck of cards. The goal is to piece back the book together into a coherent mind map.

The time for AI self-regulation is over. 200 Nobel laureates, former heads of state, and industry experts just signed a statement: "We urgently call for international red lines to prevent unacceptable AI risks" The call was presented at the UN General Assembly today by Maria Ressa, Nobel Peace Prize laureate:

@AtlantisPleb @googleaistudio Yeah, they were so interesting before ! 😭

Looks like @googleaistudio nuked Gemini's beautiful chain-of-thought reasoning traces, replacing them with the same crappy summaries ClosedAI uses 👎👎

Announcing Transluce, a nonprofit research lab building open source, scalable technology for understanding AI systems and steering them in the public interest. Read a letter from the co-founders Jacob Steinhardt and Sarah Schwettmann: transluce.org/introducing-tr…

When you leave OpenAI, you get an unpleasant surprise: a departure deal where if you don't sign a lifelong nondisparagement commitment, you lose all of your vested equity: vox.com/future-perfect…

I’m super excited to release our 100+ page collaborative agenda - led by @usmananwar391 - on “Foundational Challenges In Assuring Alignment and Safety of LLMs” alongside 35+ co-authors from NLP, ML, and AI Safety communities! Some highlights below...

Le "champion européen", porté par le Président Français et utilisé pour affaiblir la régulation des IA...

@Simeon_Cps @TheDavidSJ You could use FLOP-equivalent, as we do with CO2

@TheDavidSJ Yup that's what I just figured thx to a colleague. It just adds some political complexity. Thanks!

Dear compute nerds, it seems like Google is training LLMs in int8 & I wonder whether it means we should add OP along FLOP for thresholds. Do you know if: 1. int8 operations are included in the native definition of FLOP? My guess is not but lmk 2. if 2 is a reasonable conversion rate of int8 operations into FLOP for LLM training runs with equal perf or not? Thanks!

I would recommend applying, both for the topics of great importance and for the nice people that I met during my summer internship there.

@soroushjp Congrats! I'm surprised Claude 2 is always above the others, I thought it was harder to attack. Any thoughts? 🤔

🧵📣New jailbreaks on SOTA LLMs. We introduce an automated, low-cost way to make transferable, black-box, plain-English jailbreaks for GPT-4, Claude-2, fine-tuned Llama. We elicit a variety of harmful text, incl. instructions for making meth & bombs. arxiv.org/abs/2311.03348

@AISafetyMemes @martin_casado Most technologies are not like the others

@martin_casado This one technology is not like the others 🫡🤝 - Very, very, very pro-opensource meme account

I feel like we’ve all been pulled into a fucked up alternate timeline … I can’t believe we have to fight again for open source …

Addressing the long-term risks of AI doesn't mean we can ignore its present harms — and vice versa. In our @TIME op-ed, @achan96 and I argue that we need to break away from the false "present vs. future" harms dichotomy. time.com/6303127/ai-fut…

@ArthurConmy You can throw errors in jax/totchtyping (I recommend jaxtyping btw) but it doesn't solve the verbose part

Getting torchtyping w/ named axes always feels too verbose, and generally it doesn't throw errors IMO, a context manager (e.g `with torch.safe_broadcasting()`) that ensures tensors added/multiplied together have the same shapes would be ideal. Does this exist?

When writing research code fast I love being able to 1. zip(..., strict=True) # ensures same lengths 2. torch.einsum("bij,ijk->bjk", ...) # ensures correct shape lengths 3. my_tensor.item() # ensures single elements only what about ensuring tensors aren't silently broadcasting?

The grand finale: congratulations to the winning teams 🏆 for presenting a 3-minute pitch as #honestbrokers during the #Science4policy workshop by @LetiMonte1 and @VeroCasartelli.

Note that I was not in this classroom, and I don't know who had this slide in their lecture. Though I'd love to talk with them to understand why we see the world so differently

Feels strange to be studying at @EPFL... If "non-linear regression" is all it takes to build highly capable systems, we should definitely take the matter seriously. What will "non-linear regression" do in a few years? New sources of power are emerging, we need to adapt.

I've worried AI could lead to human extinction ever since I heard about Deep Learning from Hinton's Coursera course, >10 years ago. So it's great to see so many AI researchers advocating for AI x-safety as a global priority. Let's stop arguing over it and figure out what to do!

@mmitchell_ai What would it mean for an image to be objective? Does this prevent all image generators? It's not clear either what it means for a text to be objective, as most objective texts contain subjective parts... so it wouldn't prevent models to learn about subjective opinions

Third, this crazy idea that the training data must be OBJECTIVE (客觀性)! No subjective opinions allowed. That decimates most current pretraining data. ("Be able to guarantee the authenticity, accuracy, objectivity and diversity of the data").

Just read the draft Generative AI guidelines that China dropped last week. If anything like this ends up becoming law, the US argument that we should tiptoe around regulation 'cos China will beat us will officially become hogwash. Here are some things that stood out. 🧵

Why the best outcome is never as good as it seems. And the worst is never as bad as it seems. A thread on a surprisingly little known but really important concept: regression to the mean. /1

@ben_j_todd @tobyord (Or maybe random with a distribution according to the spread of different charities, in order to better capture the counterfactual of where I would donate without EA)

@ben_j_todd @tobyord Thanks for this! I always found this "some charities are 100× better" super arbitrary, given that some charities are harmful and some probably have a negligible impact. But I never went all the way to realise that what we wanted was "how much better compared to random"

1) Is it really true that some ways of solving social problems achieve 100s times more given the same effort? A thread on all the data I could find. First, @tobyord ‘s original paper, which finds global health interventions ~perfectly follow an 80:20 distribution: