Keep pulling the thread on Jared Kaplan.
AI systems are predictably expected to become broadly more capable over time due to scaling laws, increased investment, and improvements in hardware and algorithms.
AI systems will very rapidly become capable of operating in a wider range of useful environments, including operating robots and using computers like humans.
Future AI systems will be able to perform long-horizon tasks that take a human a day or a week by fixing their own mistakes.
A key risk of advanced AI is that systems may not behave as expected in high-stakes situations, indicating a lack of sufficient safety or alignment.
Anthropic is pursuing interpretability research with the long-term, ambitious goal of reverse-engineering how neural networks function.
Anthropic conducts national security-relevant evaluations to determine if its AI systems provide capabilities for building bioweapons beyond what is available through tools like Google Search.
Anthropic is developing an AI Safety Level (ASL) framework, analogous to biosafety levels, where ASL-3 defines a system capable of providing significant aid in building a bioweapon if jailbroken.
It will likely be possible to build AI systems at AI Safety Level 3 (ASL-3), which have dangerous misuse capabilities, in the near future.
There is a very significant chance that AI systems will be developed that are better at theoretical physics than human physicists.
Anthropic released Claude 3 Opus in the same month as this symposium.
Anthropic was founded because its founders believed AI systems could become predictably more capable with more computational resources, which were difficult to obtain in academia.
Training a generative model to model a data distribution inevitably leads to the model learning interesting features that describe the data.