Daily AI Roundups
Kevin Roose publishes the first 10 pages of Dario Amodei’s 2017 scaling memo. Reflection announces Beam with 501B total parameters and 23B active, but public weights are still promised for later this month. Wikimedia reports unauthorized OpenAI-agent edits and heavy traffic without evidence of system or data compromise. Reka previews the 19B Rho-1 model, while Cognition introduces Dreaming and Git-based cross-session memory. AI21’s Yuval Belfer explains when agents need verification or diverse attempts; Chester Lam benchmarks Nvidia’s Olympus CPU core, and Geoffrey Huntley explains interactive Lisp development. Cohere introduces North 2, AWS brings GLM 5.3 to eligible Bedrock enterprise customers, and OpenAI promises faster incident notification at an Australian hearing. South Korea plans a $3.5B frontier-model programme subject to budget approval. Bloomberg reports DeepSeek is nearing an 80 billion yuan funding round backed by Tencent and CATL.
- AI models
- Agent memory
- AI security
- AI research
- Developer tools
- AI infrastructure
- AI policy
- AI funding
Stories in this roundup
- DeepSeek nears an 80 billion yuan ($12B) funding round backed by Tencent and CATL, Bloomberg reportsBloomberg via Reuters/CNA