Reflection AI released **Beam**, a 501-billion-parameter open model, designed to complete tasks using fewer tokens than comparable models, reducing per-task costs. The company trained Beam with reinforcement learning, penalizing unnecessary token usage while ensuring task completion. Beam outperforms competitors like DeepSeek’s V4.1 Flash in token efficiency, according to benchmarking firm Artificial Analysis. While Chinese models dominate performance rankings, U.S. firms like Pinterest, DoorDash, and Coinbase have adopted them for cost savings. Reflection positions Beam as a lean alternative, targeting enterprise use cases like coding and agent workflows.
Alibaba Group Holding Ltd. unveiled a new generation of smart glasses priced at $450, expanding its push into consumer electronics. The move deepens the company’s foray into gadgets.
Dario Amodei argues that frontier AI labs should deliberately slow the pace of model capability improvements. The article frames this proposal against Beijing’s speed-first AI safety regime and describes AI safety as an urgent debate.
Finland's data center tally tops €67 billion as the country's boom gathers pace. Several new projects have been unveiled since Alphabet's Google made an announcement.
