Do you know what GPT-4 can do? A game from Nicholas Carlini: nicholas.carlini.com/writing/llm-fo…

We all need to join in a race for AI safety.  In the coming weeks, Anthropic will share more specific plans concerning cybersecurity, red teaming, and responsible scaling, and we hope others will move forward swiftly as well. whitehouse.gov/briefing-room/…

“I barely studied for this test” kids all grew up to be “please don’t mind but my apartment’s a mess” adults

If powerful AIs could mess with our perceptions and lead us to provide incorrect training signals, we could lose control over the training process without even realizing it. We made the first empirical benchmark for this problem (arxiv.org/abs/2308.15605). 🧵 (1/6)

This is surprisingly general! For example, you could design networks that literally “write” arbitrary info about their training data to memory (ie weights)

Nassim Nicholas Taleb Now to see if BetaBinomialDistribution works.

I think AI control will be a crucial part of safely deploying ASL-4 level models and am excited about people doing followup research in this area!

Thank you for having me MSNBC. It was an honor to discuss a subject I care deeply about.

How can we check LLM outputs in domains where we are not experts? We find that non-expert humans answer questions better after reading debates between expert LLMs. Moreover, human judges are more accurate as experts get more persuasive. 📈 github.com/ucl-dark/llm_d…

Claude 3 Opus is great at following multiple complex instructions. To test it, Erik Schluntz and I had it take on Andrej Karpathy's challenge to transform his 2h13m tokenizer video into a blog post, in ONE prompt, and it just... did it Here are some details:

We’re hiring for the adversarial robustness team Anthropic! As an Alignment subteam, we're making a big effort on red-teaming, test-time monitoring, and adversarial training. If you’re interested in these areas, let us know! (emails in 🧵)

x.com/i/article/1772…

I’m hiring ambitious Research Scientists at Anthropic to measure and prepare for models acting autonomously in the world. This is one of the most novel and difficult capabilities to measure, and critical for safety. Join the Frontier Red Team at Anthropic:

Many have said this already, but worth repeating: this is not correct. We take security seriously and that's why we end-to-end encrypt your messages. They don't get sent to us every night or exported to us. If you do want to backup your messages, you can use your cloud provider

I’m going to be honest, I had no idea what this meant

Enjoying Claude Artifacts? Want to build the next generation of Human-AI Interfaces? We're hiring an ML Lead for Artifacts. This is a unique full-stack role where you'll help co-develop new user interfaces along with the model capabilities which support them. You'll work as

Here's a letter we sent to Governor Newsom about SB 1047. This isn't an endorsement but rather a view of the costs and benefits of the bill. cdn.sanity.io/files/4zrzovbb…

My team at Anthropic is hiring research engineers and scientists. We find out whether AI models possess critical, advanced capabilities and then help the world to prepare. We'd love to hear from you! robertheaton.com/anthropic/

Join us from 9:30am PT tomorrow: clau.de/livestream.

love claudius 🥹