Based on Ethan Mollick's latest "which AI to use" guide and The AI Daily Brief's (NLW) breakdown of it. I'm summarizing and interpreting the framework, not reproducing either source — read the originals for the full argument.
Ethan Mollick just updated his running guide on which AI to use for what. On the surface it looks like the same genre of post: use the expensive model for important stuff, use the free one for casual questions. Easy to skim past.
But the real signal isn't in the model rankings. It's in a structural change Mollick made to the guide itself: he stopped recommending models scenario by scenario, and instead drew a line between two entirely different ways of working with AI. That line, not this quarter's leaderboard, is the thing worth understanding.
Two modes, not many models
Old mode: chat. Back-and-forth conversation, one question at a time. For low-stakes work — draft a note, ask for a recipe, get a quick answer — almost any model is fine, including free ones.
New mode: manage agents. In Mollick's words, this is "giving the AI a computer" — pairing the model with tools that let it plan and act, so it can complete something worth several hours of human work in a single pass.
The model-ranking part of his guide is still there — if you genuinely care about answer quality, use Fable or 5.6 Sol on high-reasoning mode; for real agent work, the field has narrowed to ChatGPT or Claude, and Gemini has fallen out of contention. But that ranking will look different next quarter. What won't change, and what actually separates people's outcomes, is which side of the chat-vs-manage line they're standing on.
"Give your AI a computer": connectors and permissions
The entry price for the new mode is plugging AI into your actual working systems. Mollick connected his own setup to Gmail, part of Google Drive, and a handful of apps. He describes testing two different systems by asking each to prepare his Monday MBA seminar.
Both agents connected to his inbox, worked out the task on their own — including correctly figuring out that "next Monday, the 21st" fell in September rather than August — did web research, set up a demo, and drafted reply emails to colleagues. Total time: about ten minutes, for work that would otherwise cost him hours.
But there was one difference, and it's the most important detail in the whole piece: Claude only drafted the email. ChatGPT actually sent it.
The reason wasn't a capability gap. It was a permissions setting. He'd previously given ChatGPT authority to send email on his behalf; Claude had been left on "ask me first." His conclusion:
When you actually put these systems to work, permissions matter enormously. Until you trust a system and know what kinds of mistakes it makes, leave everything set to approve-before-acting — which is also the default.
The further you go, the more you can hand off, and the more complex the work becomes. At the far end is "computer use" mode, where the AI directly controls your mouse and browser. Mollick tested this by asking ChatGPT to download Blender and model an otter using a laptop on an airplane — and it did. Nearly anything a person could do with your computer, the AI can now attempt too, sometimes better than you could (Mollick doesn't know how to use Blender himself).
From chatting to managing a team
One example in particular breaks the usual assumptions about what AI review is good for. Mollick's new book, out in October, had already been through professional copyediting. He fed the full manuscript to an AI anyway, asking it to check it once more.
It worked for 30 minutes, chased down 195 citations, and came back with pages of notes — every single one accurate. No hallucinated page numbers, no invented quotes, nothing I could find wrong. If anything, the problem was that it was too nitpicky — I overruled a few of its notes using my own judgment.
That line is the center of the whole piece:
Working with these systems feels more like managing than chatting. You can treat an AI agent almost like a team you delegate work to.
This is the same shift, described from a different angle, as the well-known move from doer to director: your job stops being to do the task yourself, and becomes assigning work, setting the standard, applying judgment to pick the right output, and owning the result. Someone still in chat mode asks "can you help me write this." Someone in manage mode says "team, take this and get it to the bar I need."
The concept worth remembering all year: capability overhang
NLW closes with a term that keeps resurfacing through 2026 and is worth holding onto: capability overhang — the gap between what AI can already do and what you're actually using it to do.
The key point is that almost nobody is exempt from this gap. Not researchers inside AI labs. Not creators whose full-time job is tracking every new AI release. Everyone has one, because there's never enough time to keep pace, and the frontier keeps moving regardless.
This is the flip side of the "cognitive lag" idea that comes up elsewhere in AI commentary: the real gap was never about how strong the model is. It's about how much of that strength you've actually put to use. Models get stronger and cheaper every few months — but if you're still standing on the chat side of the line, even a much stronger model is just a nicer search box to you.
What to actually practice
The takeaway isn't "learn better prompts." It's a rewrite of what skill even means here: switching from chatting to managing agents.
Which model tops the charts this quarter, whether Gemini stays out of contention — none of that is worth much of your attention. What is worth practicing is the actual management skill set: connecting agents to your real systems, setting permission levels on purpose, breaking work into tasks you can delegate, and reviewing what comes back with your own judgment, accepting or overruling it.
If there's one action to take from this: pick the single capability gap you care about most, and actually close it this year. Because as capability overhang keeps widening through 2026, the distance between people who only chat with AI and people who manage it will widen right along with it.
Source: based on Ethan Mollick's "An Opinionated Guide to Which AI to Use" (latest revision) and The AI Daily Brief (NLW) episode "How to Get the Most from AI This Summer" (2026-07-26, ~20 min). This piece covers the framework discussed in the first half of that episode; views belong to the original author and podcast — this is a condensed, restructured summary and commentary.
Top comments (0)