With GPT-3, didn't have a way to steer the model's behavior. Now we pick a clusters of issues (e.g. model doubling down on wrong answers — something that used to be a huge problem), build evals, collect data, retrain the model weekly, and deploy. Anyone can test if we've improved
@gdb
-
GPT-4 Alignment Progress: Scaling Beyond Human Judgment
By
–
But aligning GPT-4 is a step, not the destination. Great progress (compare to early complaints about GPT-3), but will need to scale to problems where human can't even judge the outputs. GPT-4 itself already used a lot in our alignment, step towards this:
-
Model Generalization of Helpful AI Assistant Behavior Across Topics
By
–
One example observation — much of model's good behavior across such a broad range of tricky or adversarial topics comes from the model having generalized the human instructors' concept of being a helpful AI assistant. Not something I saw coming!
-
GPT-4 Deployment and Practical AI Alignment Progress
By
–
Deploying GPT-4 subject to adversarial pressures of real world has been a great practice run for practical AI alignment. Just getting started, but encouraged by degree of alignment we've achieved so far (and the engineering process we've been maturing to improve issues).
-
GPT-4 for Just-in-Time UI Generation
By
–
GPT-4 for “just-in-time” UI generation: https://t.co/zF1ljr2xI3
— Greg Brockman (@gdb) 30 mars 2023GPT-4 for “just-in-time” UI generation:
-
Virtual Holographic AI Companion Technology and Applications
By
–
Virtual holographic AI companion: https://t.co/hMpvElP8QO
— Greg Brockman (@gdb) 30 mars 2023Virtual holographic AI companion:
-
OpenAI Mission: Ensuring AGI Benefits Humanity
By
–
Mission of OpenAI is to ensure AGI benefits all of humanity. Key ingredients we think will be important:
-
Sam Altman tour engaging users developers and policymakers
By
–
Hang out with @sama on upcoming tour. Looking to engage with & hear feedback from users/devs/policymakers/anyone else interested in sharing.
-
GPT-4 Strengthens Cybersecurity Defense with Formal Verification
By
–
GPT-4 for security, giving a real power boost for defenders. One encouraging (but still tentative) idea that I've heard over the years is that AI may permanently shift the favor to defenders; for example, by enabling formally verified systems.
-
Evaluating World Model Capabilities Across GPT Versions
By
–
The debate over questions like “does GPT have a world model” continues. No real need to argue about it based on individual anecdotes — better to just write an OpenAI Eval, and check the trendline of performance across different GPTs: https://
github.com/openai/evals