DeepSeek V4 and the New Frontier of Efficient AI
DeepSeek's V4 isn't claiming to be the most powerful model in the world. It's claiming to be the most efficient at long context, and backing that up with a detailed technical report on a hybrid attention architecture that achieves a 90% reduction in memory costs at the million-token scale. That architectural bet, not the benchmark numbers, is the real story.
Read full story →Nvidia Sits at $5 Trillion. Its Biggest Threat Is Paying Its Bills.
Nvidia hit an all-time high this week with 81% of the AI chip market. The challengers everyone watches, AMD and Intel, are not the threat that matters. Google's Ironwood TPU and Amazon's Trainium chips are being built by Nvidia's own largest customers, following a playbook Amazon already ran to displace Intel in CPUs.
Read full story →The Scaffold Is Disappearing: What AI Is Taking From Junior Lawyers
AI adoption in law has doubled in a year, and productivity gains are real. But two new analyses raise a question the profession isn't answering: the rote, formative work AI replaces was also how legal judgment was built. Whether review-mode develops the same instincts as production-mode is an open question nobody is measuring.
Read full story →The Impossible Entry
The entry bar for software development now demands breadth and depth simultaneously, then asks newcomers to compete with AI on top of that. This is not a description of an exceptional candidate. It is a description of someone who doesn't exist. The profession has quietly stopped having room for ordinary talented people.
Read opinion →