Descriptions:
Matthew Berman examines DeepSeek V4 — the latest flagship release from the Chinese AI lab — arguing it poses a deeper long-term threat to US AI competitiveness than most coverage acknowledges, and connecting its release to escalating government action around AI intellectual property theft.
Berman covers the model’s technical specifications in detail: DeepSeek V4 Pro is a 1.6 trillion parameter mixture-of-experts architecture with 49 billion active parameters and a one million token context window, while V4 Flash offers 284 billion total parameters at significantly lower cost. Both were trained on approximately 33 trillion tokens. Benchmark comparisons against Claude Opus 4.7 and GPT-5.5 on MMLU Pro, GPQA Diamond, and SWE-bench Verified show V4 Pro trailing slightly but competing at the frontier — achieved using export-controlled, lower-tier Nvidia GPUs through algorithmic innovation.
The video gives particular attention to two policy-level developments: Anthropic’s published report alleging that Chinese AI labs conducted industrial-scale distillation attacks against Claude — systematically extracting training signal from the model’s outputs — and a follow-up statement from US government AI director Michael Kratzios announcing action to protect American AI intellectual property. Berman frames the broader tension through Jevons Paradox: as DeepSeek makes frontier AI dramatically cheaper globally, demand and usage accelerate, potentially benefiting Chinese AI adoption even as export controls attempt to limit Chinese compute access.
📺 Source: Matthew Berman · Published April 25, 2026
🏷️ Format: News Analysis
