🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Mixtures of engineered bacteria were able to: - Identify if a number is prime - Check if a letter in a string is a vowel - Determine the max number of pieces of a pie obtained from n straight cuts. Answers are printed by expressing fluorescent proteins in different patterns.

Airlock looks neat

We're excited to share our new preprint introducing endless jailbreaks via bijection learning. Our attack exploits the advanced reasoning abilities of frontier LLMs like GPT-4o and Claude 3.5 Sonnet, revealing a critical model vulnerability that arises from capabilities.

Be aware if you plan to derive anything about human behaviors with "LLM participants." In this 100-page paper, we show how current LLM-generated psychometrics responses cannot capture nuances where human individuality resides and how to evaluate it properly. #AI4SocialScience

This paper + huggingface.co/datasets/proj-… release/discussion recently, as well as other surveys on personality adherence in LLMs (i.e. arxiv.org/pdf/2406.01171) are critical to anyone trying to simulate customer personas or potential customers in the innovation space.

Our paper on individual neurons that regulate an LLM's confidence was accepted to NeurIPS! Great work by Alessandro Stolfo and Ben Wu @ICLR Check it out if you want to learn about wild mechanisms, that exploit LayerNorm's non-linearity and the null space of the unembedding, productively!

Large study shows humans can learn from AI feedback but access to AI also amplifies existing inequalities by increasing the skill gap and reduces intellectual diversity: everyone learns to specialize in the same areas arxiv.org/abs/2409.18660

The thing I was most excited about in the OpenAI dev day talk was their improved Evals offerings that include integration of more customizable testing criteria. There are a lot of tools for this outside of OAI, but neat to see as part of the whole offering, Factuality, sentiment,

🧙‍♂️ Does all of a transformer's magic come from training? In our NeurIPS 2024 paper, we discovered that for many tasks, merely training the embedding and unembedding layers of transformers yields surprisingly strong performance! A thread 🧵

We generate explanations for millions of features extracted from Llama 3.1 and Gemma. You can download them at huggingface.co/datasets/Eleut…. Our analysis confirms that SAE latents are much more interpretable than neurons, even when neurons are sparsified using top-k postprocessing.

I like the commentary on the approach of feature overlap when looking at SAEs trained on residual streams vs MLPs and implications for interpretability choices when on a budget.

Excited to announce Centaur -- the first foundation model of human cognition. Centaur can predict and simulate human behavior in any experiment expressible in natural language. You can readily download the model from Hugging Face and test it yourself: huggingface.co/marcelbinz/Lla…

The arXiv.org preprint is currently on hold but while waiting you can already download a PDF here: osf.io/preprints/psya…

1/ We are excited to introduce Oasis, the world's first real-time AI world model, developed in collaboration with Etched. Imagine a video game entirely generated by AI, or a video you can interact with—constantly rendered at 20 fps, in real-time, with zero latency

Humans sometimes get distracted by pop-ups… but for AI agents, it’s worse! Pop-ups explicitly designed for agents can make them click 87% of the time, majorly derailing their tasks. Tao Yu Diyi Yang arxiv.org/abs/2411.02391 github.com/SALT-NLP/Popup…

What a blast! A huge thanks to @zipsmbb Head Coach John Groce for joining us on this week's episode of Zips Nation Insider. We talked so long that this one will probably be a 2-parter. The episode will stream on all of 330ToGO platforms.

We're open-sourcing Sparse Autoencoders (SAEs) for Llama 3.3 70B and Llama 3.1 8B! These are, to the best of our knowledge, the first open-source SAEs for models at this scale and capability level.

Reasoning tokens in AI models are probably best read in Gollum's (Andy Serkis) voice.

This is so fast.