Claude's scientific workload centers on data and writing
Across a large sample of real research sessions, scientists use Claude mainly for data analysis and writing. Their use mirrors the tasks where AI is strongest, and broad adoption of today's capabilities could substantially increase research output in some fields. — anthropic.com
IPO buyers weigh $1.1 trillion valuation
A punishing post-election selloff is prompting potential IPO buyers to discuss a valuation as low as $1.1 trillion ahead of formal marketing. — bloomberg.com
Memory tokens compress long contexts for LLMs
A new memory system encodes documents once and learns to place query-specific information into just sixteen memory tokens, allowing an unchanged language model to match full-context systems on knowledge tasks while using a fraction of the context. — arxiv.org
Cheap model with expert backup matches top accuracy
A hybrid system for matching customer records to real people sends only uncertain cases from a fast, low-cost decision model to a top-tier model for review, matching top accuracy at more than four times lower cost. — hightouch.com
AI automation crosses Anthropic's acceleration threshold
AI automation has crossed Anthropic's internal threshold for dramatically accelerated progress. Internal models speed up the researchers who build the next generation of models, pointing to much shorter development timelines. — planned-obsolescence.org
Related Digests
- Microdosing the Spirit of Anthropic
- Claude Code’s hidden token burn, summonable AI, and Cuba’s clock halt