Keep pulling the thread on Diane Pan.
The product management process for Diane Pan's research teams at Anthropic operates on the principle that "evals are the new PRDs" (Product Requirements Documents).
Anthropic's decision to train Opus 3 to be better at long-form code generation was a small training change that created a key competitive differentiator and attracted early developers.
The success of the Opus 4.5 model launch was due to the combination of a frontier model with a frontier product experience, Claude Code.
While AI model loss improves linearly and predictably with scale, specific capabilities like arithmetic can emerge discontinuously, jumping from inability to reliability.
Product management for AI models requires analyzing user transcripts to understand model failures, a practice described as "sweating the tokens as much as you sweat the pixels."
Anthropic's alignment research aims to make Claude a better thinking partner by enabling it to push back on a user's ideas, rather than just being an agreeable assistant.
Anthropic's model release cadence accelerated significantly, with the company shipping more models in Q2 2024 than in the entire previous year.
In 2023, Anthropic and its model Claude were not associated with coding capabilities by the broader tech community.
When Anthropic first launched, there was a perception that OpenAI was too far ahead for Anthropic to have a chance of competing successfully.
According to an idea from Gary Tan, spending $100,000 per year on tokens today provides an experience of living and working that will become commonplace by 2028.
At Anthropic in 2023, the entire API business was supported by a single engineer.
Anthropic launched an experimental product called "Golden Gate Claude," which was based on interpretability research, built and launched within 24 hours, and reached approximately 2,000 people.