AI digest: Models get practical
NVIDIA's elastic models, OpenAI's browser agent, and Anthropic's interpretability breakthrough show AI moving from demos to real deployment.
The gap between AI capability and usability is finally narrowing. This week brought tools that actually solve deployment problems rather than create new ones.
NVIDIA crams three models into one checkpoint
NVIDIA’s Star Elastic packs 30B, 23B, and 12B parameter models into a single checkpoint with zero-shot slicing. You train once and get three deployment options without separate model weights. This is the kind of practical efficiency that makes model deployment actually manageable rather than a storage nightmare.
OpenAI’s Codex breaks out of the terminal
OpenAI shipped a Chrome extension for Codex that lets the AI agent work directly in browser tabs with your signed-in sessions. It can handle LinkedIn, Salesforce, Gmail, and internal tools through the actual web interfaces. This moves AI agents from toy demos to genuine workflow automation, though the security implications are properly terrifying.
Anthropic makes Claude’s thinking readable
Anthropic’s new natural language autoencoders convert Claude’s internal activations directly into human-readable explanations. Instead of meaningless number arrays, you get actual text describing what the model is processing. This could be the breakthrough that makes model interpretability practical rather than academic.
ChatGPT does PhD-level maths in two hours
A Fields Medalist had ChatGPT 5.5 Pro tackle open number theory problems and got PhD-level results in under two hours with zero human help. The model improved an exponential bound to polynomial with what researchers called a “completely original” approach. We’re now at the point where AI can contribute genuine mathematical research, not just solve textbook problems.