Vendor-run launch benchmarks overstate real-world speed and cost gains

Artificial intelligence research, products, and impact across society.

DeepSeek's 552-billion-parameter mixture-of-experts model, released 10 September 2026, and the smallest in a new architecture family with native visual understanding. A Causal Encoder-Decoder design activates just 8 billion parameters for input and 16 billion for output, and cuts KV cache memory to a quarter of the previous generation, which DeepSeek passes on as lower API prices while phasing out DeepSeek-V4-Pro.

Standardized methods for evaluating AI system performance. Enables comparison across models, tasks, and metrics.

A class of AI models introduced by TypeSafe AI that output typed, probabilistic decisions in parallel instead of generating text token by token, designed for fast, structured decision-making inside software systems. Named after Kahneman's fast, intuitive "System 1" thinking.

The practice of measuring and comparing the capabilities of AI models through standardized benchmarks and evaluation protocols.

Statements AI companies make about their products and risks, and whether they serve marketing goals.