GPT-3 is astonishing. It's also not ready for your business. Both are true
A year after GPT-3, the enterprise hype is loud. After testing it on real problems, here's the honest gap between the demo and a system you'd actually ship.
📝 GPT-3 is astonishing. It's also not ready for your business. Both are true.
A year on from GPT-3, the enterprise pitch is everywhere: it writes, summarises, answers, drafts — surely this transforms how we work with text. I've spent time putting it against real problems, and my honest read holds two things at once that the hype refuses to.
It is genuinely astonishing. A 175-billion-parameter model that produces fluent, human-like text from a few examples, no task-specific training — that is a real shift in what's possible with language. Anyone who dismisses it hasn't used it.
And it is not ready to drop into a business process. Not because it's weak, but because "impressive demo" and "dependable system" are different species:
It is confidently, fluently wrong — with no signal of when. In a marketing draft, charming. In anything a customer or a regulator sees, a liability.
It knows the internet, not your business — nothing about your policies, your data, your customers. The knowledge that makes it useful in a demo is exactly the knowledge that's irrelevant to your actual work.
It inherits the biases of what it read, which for any enterprise use is a governance problem before it's a technology one.
It's expensive and heavy, and coaxing reliable output takes real prompt craft — this is not plug-and-play.
None of that is a reason to ignore it. It's a reason to be precise about where it fits today: drafting where a human always reviews, brainstorming, first-pass summaries — assistive, never autonomous. The gap between that and "automate the workflow" is the whole job, and it won't close by wishing.
The pattern I keep coming back to: with any capability this new, the honest position is neither evangelist nor sceptic. Test it on your real problems, find the narrow places it genuinely helps, and keep a human between it and anything that matters.
The question worth asking before you build on it: where in your process is a confident, occasionally-wrong first draft actually useful — and where would it be a disaster? Fund the first. Stay far from the second.