Writeups

Human task duration makes AI capability easier to interpret. Productivity, autonomous research and catastrophic risk depend on more than the work a benchmark captures.

Part 1 follows five research bets—from longer inference to online models—then examines Sora's physical consistency, Groq's utilization economics, and Gemini's million-token context.

Imbue’s approach starts with a recruiting product’s trust problem and asks what it takes to make models dependable enough to delegate work to.

Pushmeet Kohli and Sal Candido on protein structure, biological data, and the work required to make models useful for science and medicine.

Falling model prices weakened OpenPipe’s distillation business. Reinforcement learning offered a new opportunity: improve agents through experience, provided their training environments behave like production.

Narrow tests and leaked repository knowledge make higher scores harder to trust. The next coding evaluations need harder work and credible judgments of quality.

Why Codeium built its own editor—and how developer actions, code retrieval, reversible execution, and enterprise requirements shape the product.

Converting a blog into a podcast requires more than a convincing voice: worthwhile source material, controllable delivery, publishing tools, and a way to fix the result.

A fictional terminal and a browser for imaginary websites turn prompting into exploration—and reveal the gap between plausible state, working software, and understanding.

Useful imperfect models, evaluations drawn from real tasks, and security research that changes deployed systems—from the conversation published August 29, 2024.

Large-scale perturbation experiments give X-Cell data about interventions. The test is whether it can use that data to predict responses in unfamiliar biological contexts.