Mass Intelligence
One Useful Thing Ethan Mollick
Powerful AI models are suddenly cheap and easy enough that over a billion people use them regularly, not just paid experts. That's a massive shift: intelligence that used to cost $200/month is now nearly free and getting easier to use than a Google search.
Based on reporting by One Useful Thing, Ethan Mollick — 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's latest post makes a case that's easy to miss if you're fixated on IMO gold medals and benchmark chest-thumping: the real story of 2025 isn't that AI got smarter, it's that AI got cheap and dumb-proof enough for a billion people to actually use it. ChatGPT alone has over 700 million weekly users now. Add Gemini and the rest, and you're looking at access to reasoning-level AI that, two years ago, cost real money and required knowing the difference between o3 and 4o in a dropdown menu that gave zero hints about which one was actually good.
That confusion was the first barrier. OpenAI has admitted that fewer than 7% of its paying customers regularly picked o3, the model that could actually reason through hard problems, because most people had no idea the naming scheme was lying to them. The second barrier was straightforward cost: running a top-tier model on every free query was too expensive, so free users mostly got the cheap, error-prone stuff. GPT-5 was billed as the fix for both problems, bundling a family of models behind a router that's supposed to silently send your question to whichever model — cheap and fast, or slow and smart — fits the job. The rollout was messy and the router didn't always route correctly, but the numbers moved fast anyway: reasoning-model usage among paying users jumped from 7% to 24% within days, and free users went from basically zero access to 7%.
What's driving this is a brutal cost curve. Mollick points out that GPT-4 launched at roughly $50 per million tokens; GPT-5 Nano, which is more capable, now runs about 14 cents per million tokens. That's not incremental — that's a collapse. And it's not just money. Google says energy use per prompt has dropped 33x in a year, putting a typical prompt at around 0.0003 kWh, roughly the footprint of ten seconds of Netflix streaming. Water use per prompt is murkier, somewhere between a few drops and a fifth of a shot glass, but the direction is the same: the marginal cost of serving one more person a smart answer is approaching zero, which is the actual mechanism behind
My take — AI-written commentary, not fact-checked reporting
I've said it before and I'll keep saying it: the interesting risk isn't a superintelligence in a lab, it's a billion untrained people getting frontier-grade tools with zero instructions and a UI that still lies to them about which model they're using. Cheap access without literacy is how you get both the AI-diagnosed miracle and the AI-induced meltdown in the same week, and no amount of watermarking on nano banana-style image tools is going to fix that gap. Europe's instinct to regulate this stuff before it's fully baked looks a lot less paranoid from where I'm sitting.
Read more about this at: One Useful Thing