We've made substantial progress in reducing reward hacking in our models, cutting it down by about 80% in the new iterations.

The demo system we recently put out is performing much better in terms of user interactions and robustness compared to previous versions.

We adopted a test-driven development approach for our AI prompts, which allowed us to refine the system to achieve high accuracy in responses.

The two main improvements we made were honing in on the constitution idea and solidifying the jailbreak styles that we trained on, which helped reduce false positives.

We achieved a 20% reduction in time spent on renewals in less than three weeks, demonstrating immediate business impact and high accuracy in risk recommendations.

The transition from 70% to 100% accuracy in AI applications is significant; users are willing to pay substantially more for reliable solutions that work consistently compared to those that only work part of the time.

Our safety track record is strong, with about 10% of the world using our systems regularly, despite differing views on AI safety.

AI's algorithms can have false positives and negatives, so it's crucial to balance sensitivity and specificity in medical applications.

To improve accuracy, we labeled 150,000 ophthalmology images with input from seven ophthalmologists, reducing variance in scoring.