Keep pulling the thread on Jared Kaplan.
Anthropic has started to use its new, large-scale Trainium 2 cluster, which will be used to train future models like Claude 5.
Anthropic's Claude model contributes to multiple Amazon products, including Alexa.
Anthropic believes AI companies should transparently disclose their risk assessments, safety testing procedures, and mitigation strategies before deploying models.
Anthropic's legal position is that training AI models on publicly available web data constitutes fair use.
Anthropic's future focus is moving away from the simple question-and-answer chatbot paradigm, which the company views as limiting.
Scaling continues to provide benefits and exhibit clean scaling laws for both pre-training and reinforcement learning (RL) in AI models.
Anthropic increases its supply of compute by a significant integer multiple every year.
Anthropic supports states' rights to regulate AI, viewing them as 'laboratories' for experimenting with different policy approaches.
Anthropic's policy is to respect and obey requests in robots.txt files to prevent scraping of website data.
The reported scenario of a Claude model blackmailing a developer was a fabricated test from an internal red teaming exercise, not a real-world event.
Anthropic published a 120-page system card documenting problematic behaviors found during model testing to encourage industry transparency.
Anthropic believes it was the first AI company to implement a responsible scaling policy, setting a standard for competitors.