Keep pulling the thread on Mike Krieger.
Roughly 90% of Anthropic's code is now written by AI.
Mike Krieger believes it is achievable within 2024 for an AI agent to identify emergent user feedback, propose a solution, and generate a pull request to implement it.
The increased volume of AI-generated code at Anthropic created a bottleneck in the merge queue, forcing the company to completely re-architect it.
Over 70% of pull requests at Anthropic are now generated by its Cloud Code tool.
In the last year, over $1.4 trillion, representing 1.3% of global GDP, was processed through the Stripe platform.
Anthropic's CEO, Dario Amodei, predicted in 2023 that their models would achieve a 90% score on the SWE-bench coding benchmark by the end of 2025.
Anthropic's models improved their score on the SWE-bench coding benchmark from 50% in 2023 to approximately 72% in mid-2024.
Anthropic's Claude Code team has shifted its code review process from line-by-line human review to using another instance of Claude for review, with humans performing higher-level acceptance testing.
Mike Krieger estimates that over 95% of the code for Anthropic's Claude Code product is written by Claude Code itself.
At Anthropic, Mike Krieger found that embedding product teams with research teams to improve the core model provided significantly more leverage than having product teams focus solely on the user experience layer.
Anthropic's product strategy focuses on building features, like Artifacts, that require a deep integration between the product and the underlying model, rather than features that could be built using off-the-shelf APIs.
The decision to shut down the AI-powered news app Artifact was driven by a combination of the deteriorating mobile web experience, the product's lack of natural virality, and the challenges of making strategic pivots with a fully remote team.