A million users in 5 days. ChatGPT's real story isn't the model
A million users in five days. After a decade building ML systems, here's what looks genuinely new, what looks like a demo, and what enterprises should watch.
👀 A million users in 5 days. ChatGPT's real story isn't the model.
Like apparently everyone else, I have spent the week testing ChatGPT — and fielding the same question from colleagues across sales, HR and IT: is this real?
After eleven years building ML systems, my honest read: yes — but not for the reason the demos suggest.
What is genuinely new here is not the intelligence. The underlying technology has been visible for a while — transformers since 2017, GPT-3 since 2020. Those of us who worked with GPT-3 know it was impressive and unusable in equal measure: powerful, but you had to coax it. What OpenAI shipped this time is alignment and interface — a model trained to follow instructions and hold a conversation, wrapped in a chat box anyone's grandmother can use. The breakthrough is that capability which previously needed a prompt engineer now needs only a sentence. Distribution of capability, not creation of it — that is the event.
What a decade of ML makes me watch for:
It is confidently wrong, and beautifully so. Fluent, authoritative answers with no signal of when it is inventing. Every ML system fails; this one fails persuasively — a new category of risk for any serious use.
It knows nothing about your business. Its knowledge stops in 2021 and does not include your documents, your policies, your customers. Every "this will transform enterprises" take has to cross that gap first.
Demo-to-production is the graveyard of ML projects, and nothing about this launch changes that. The distance between "amazing answer on Twitter" and "reliable answer in a workflow, with accountability" is where the actual work lives.
Where I would place early bets, cautiously:
Natural language becomes an interface layer. Not replacing software — sitting in front of it. The pattern of typing a sentence instead of navigating menus will spread faster than any specific model.
The enterprise version of this is grounded, not general. The valuable system will be connected to your knowledge, answering from your documents, able to show where an answer came from. General chat is the demo; grounded answers are the product.
Models like this will multiply and cheapen. A decade of ML has one repeating lesson: capability commoditises. Durable advantage sits where it always sits — data quality, integration, adoption — not in access to the model.
There is no API yet. When there is, the real race begins — because that is the moment this stops being a website and becomes a component.
For fellow leaders getting the "should we do something with this?" question: the sensible answer is not a strategy deck yet. It is structured play — let teams try it on real, non-confidential tasks, collect where it genuinely helps and where it confidently fails, and hold the enthusiasm and the scepticism at the same time. Both will be needed next year.