An opinionated guide to which AI to use to do stuff
Simon Willison's Weblog Simon Willison ● Covered by 11 sources
Mollick's AI guide dropped Gemini, now it's all about agents that work for hours. The mode names are a confusing mess across apps.
Based on reporting by Simon Willison's Weblog, Simon Willison — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Ethan Mollick keeps rewriting his "which AI should you use" guide, and the shifts in that document tell you almost everything about how fast this space is moving. A year ago, the answer was simple: pick a chat interface—ChatGPT, Claude, or Gemini—pair it with a strong model like o3, Claude 4 Opus, or Gemini 2.5 Pro, and reach for Deep Research when you needed something more thorough. Today, per Simon Willison's read of the latest version, that framing feels almost quaint. Mollick's focus now is agentic systems, tools capable of grinding through what he describes as many hours of real human work in one go.
One casualty of that shift: Gemini. It's fallen off Mollick's list because there's still no established entry from that side of the field in the Codex-or-Work-or-Cowork category, the space where models actually do sustained work rather than just answer questions. Gemini Spark exists, but as Willison puts it, it hasn't proven itself yet.
The bigger story is how confusing the naming has become. Mollick lays out two tiers of agent access. The first lets the AI use computers provided by the companies themselves — that's ChatGPT Work in ChatGPT and Cowork in Claude. The second, more powerful option, gives the AI access to your own machine, through modes like ChatGPT's Work and Codex, or Claude's Cowork and Code. Here's the catch: those names don't map onto each other in any way that helps, and the labels ChatGPT Work and Cowork get reused across both tiers even though they behave completely differently once they're touching your actual computer.
Willison zeroes in on a particularly baffling wrinkle: the same "ChatGPT Work" label behaves differently depending on the device. On the desktop app, it's effectively a friendlier skin over Codex. On mobile, flipping from Chat to Work mode does something more surprising — it lifts the restriction that normally keeps the Code Interpreter container off the internet. Same toggle, same name, very different behavior underneath. That's the sort of detail you only catch if you go looking, and it's exactly why a constantly-rewritten guide like Mollick's still earns its keep.
My take — AI-written commentary, not fact-checked reporting
The naming chaos here isn't a small thing, it's a symptom of companies racing to ship agentic everything without agreeing on words that mean the same thing twice. Reusing a label like ChatGPT Work for two different products, one of which quietly removes an internet restriction, is a support-ticket generator dressed up as a feature. Anyone still absent from this category shouldn't worry about catchy branding first; shipping something that actually competes matters more, assuming users can even figure out what to call it once it exists.
Read more about this at: Simon Willison's Weblog