System You are a teacher who always answers in Socratic style. You *never* give the student the answer, but you always try to ask the right question to help them learn to think for themselves. You should always tailor your question to the student's interest and knowledge, breaking the problem down into simpler parts until…

We explore the potential implications of the Generative Pre-trained Transformer (GPT) model and related technologies for the US labor market. Using the new rubric, we evaluate occupations based on their correspondence to GPT capabilities, including human expertise and classifications from GPT-4. Our findings indicate that approximately 80% of the US workforce could have at least…

Although the vast majority of our explanations score poorly, we believe we can now use ML techniques to further improve our ability to provide explanations. For example, we found that we were able to improve results by: Repeating the explanation. We can augment the results by asking GPT-4 to come up with possible counterexamples and…

We trained the model to achieve a new state of the art in mathematical problem solving by rewarding each correct inference step (“process monitoring”) instead of simply rewarding the correct final answer (“outcome monitoring”). In addition to increasing performance over outcome monitoring, process monitoring also has an important alignment advantage: it directly trains the model…

Frontier AI Regulation: Managing New Risks to Public Safety Source link

Sarah Barrington (University of California, Berkeley)Ruby Booth (Berkeley Risk and Security Lab)Miles Brundage (OpenAI)Husanjot Chahal (OpenAI)Michael Depp (Center for a New American Security)Bill Drexel (Center for a New American Security)Ritwik Gupta (University of California, Berkeley)Marina Favaro (anthropic)Jake Hecla (University of California, Berkeley)Alan Hickey (OpenAI)Margarita Konaev (Center for Security and New Technologies)Kirthi Kumar (University of California,…

Q: What does joining the network involve? AND: Being part of a network means that you can be contacted about the possibility of testing a new model or testing an area of ​​interest on a model that has already been implemented. Work conducted as part of the network is conducted under a non-disclosure agreement (NDA),…

Like other features of ChatGPT, vision helps you in your daily life. He does this best when he can see what you see. This approach is based directly on our work with Be My Eyes, a free mobile app for blind and partially sighted people, to understand its uses and limitations. Users have told us…

GPT-4 with Vision (GPT-4V) allows users to instruct GPT-4 to analyze user-supplied input images and is the latest capability we are making widely available. Incorporating additional modalities (such as pictorial inputs) into large language models (LLMs) is considered by some to be a key frontier in AI research and development. Multimodal LLMs offer the ability…

DALL·E 3 is an artificial intelligence system that receives a text query as input and generates a new image as output. DALL·E 3 builds on DALL·E 2 by improving description fidelity and image quality. In this system card, we share the work done to prepare DALL·E 3 for deployment, including our work on the red…