Your hands-on guide to 'doing stuff' with AI. From Wharton's Ethan Mollick, this is the most practical guide I've yet seen on getting real work done with AI platforms —including ChatGPT, Bard, Claude 2, Stable Diffusion, Midjourney, and more. It includes tips on prompt-crafting, advice on integrating AI into your workflow, and comparison charts to help you understand where each model excels and flails.

Google's vision for how doctors and AI can team up. In a new paper in Nature Medicine, Google's DeepMind and Google Research propose a system called CoDoC that aims to answer: “When is AI more accurate, and when is a human?” The paper outlines a hypothetical health care scenario where AI attempts to diagnose X-rays but intelligently determines when to defer to a clinician, finding the approach could reduce false positives by 25%.

• Related: While it's nice to imagine humans and AI playing harmoniously together, will the partnership always lead to better results? Maybe not, according to a new NBER working paper. It describes an experiment that found “providing AI predictions [to human radiologists] does not uniformly increase diagnostic quality.”

The White House just secured safety commitments from top AI firms. The commitments — agreed to by companies including Google, Meta, Microsoft, and OpenAI — are voluntary and broadly defined but include “developing robust technical mechanisms to ensure that users know when content is AI generated,” “publicly reporting their AI systems' capabilities, limitations, and areas of appropriate and inappropriate use,” and “investing in cybersecurity and insider threat safeguards.”

Meta just made its AI model LLaMA 2 free. Is that a good thing? “Free” here means truly “free”: You can download the model, run it on your own computer, and even use it in commercial applications. This is a big deal. Other AI models such as ChatGPT and Bard may be free for consumers to use, but the models run remotely on computers operated by the models’ owner. A model you can run yourself offers big advantages (e.g., you don't have to trust anyone else with your data) but also big risks (e.g., spammers and scammers could use state-of-the-art AI to nefarious ends).

ChatGPT goes to Harvard. Writing for Slow Boring, Harvard student Maya Bodnick describes how she (openly) asked professors to grade ChatGPT’s versions of assignments in classes from Microeconomics to Intermediate Spanish. The AI scored a 3.34 GPA, with graders saying its work was “beautifully written” and showed “impressive … attention to detail.” Bodnick was left doubting whether it will be possible— or even desirable — to stop students from using AI in coursework: “Even if colleges can successfully prevent students from using ChatGPT to write their essays, that won't prevent the AI from taking their jobs after graduation,” she writes.

Can you guess what AI can — and can't — do? I loved this “game” that asks you to predict whether ChatGPT can successfully complete a wide range of tasks. It's an immensely practical way to test your AI instincts and get immediate feedback: Can AI tell you the capital of France? Compute the integral of x sin(x) from 0 to 2𝜋? Code a playable game of tic-tac-toe? And if you think you know the answer … are you sure?