Section J
Daily picks, deep dives, quick takes, listicles, charts and predictions.
A new study of four LLMs in five Indic languages finds that telling a model it is being safety-tested raises refusals of harmful requests by 14.4 points on average.
AI Desk1 min read
DPO showed language models can be aligned to human preferences without a separate reward model or reinforcement learning. How it works and why it spread so fast.
AI Desk3 min read
Switch Transformers showed that routing each token to one expert lets models grow to a trillion parameters without proportional compute. How MoE works and why it matters.
AI Desk4 min read
A reusable prompt that makes AI review your draft against clear criteria before rewriting it — with the template, examples for emails, essays and code, and variations.
AI Desk3 min read
A reusable prompt that turns transcripts or rough notes into a clear summary of decisions, action items with owners and deadlines, open questions and risks.
AI Desk3 min read
A deep dive into AI economics: training compute, chips and data centers, inference costs, falling token prices, business models and who captures the value.
AI Desk6 min read
A deep dive into open-weight versus closed AI models: definitions, arguments on innovation, safety and competition, business strategies and the policy debate.
AI Desk5 min read
Why leaderboard rankings are a poor guide to choosing an AI model for your work, what they do measure, and a simple way to test models on your own tasks.
AI Desk3 min read
Models now accept huge inputs, but stuffing everything into the prompt can reduce accuracy and raise costs. When long context helps, and when retrieval works better.
AI Desk3 min read
A practical list of everyday tasks where AI assistants genuinely save time, from drafting to explaining, and five common tasks where they still fail and need checking.
AI Desk3 min read
Fifteen concrete, low-risk ways to use AI assistants at work — from email triage and meeting prep to spreadsheets, research and learning — with a prompt idea for each.
AI Desk3 min read
In a controlled experiment, developers with GitHub Copilot completed a coding task about 56% faster. The study design, results, limitations and lessons for teams.
AI Desk3 min read
How Be My Eyes added an AI assistant that describes photos for blind and low-vision users, what it changed, its limits, and lessons for designing accessible AI.
AI Desk3 min read
The structural trends shaping AI — agents, reasoning models, falling costs, multimodality, on-device AI and regulation — and the signals that show which are real.
AI Desk3 min read
AI predictions range from "AGI next year" to "it's all hype". How to evaluate forecasts, what forecasting research teaches, and the warning signs of unreliable claims.
AI Desk3 min read