AI does it all, so why hasn't company productivity moved? What's missing isn't the model but people who have tried it

Same model, different results

In the fourth week of September, Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Sol came out on the same day, two hours apart (according to Seungjoon Choi in the episode). A model is the core of an AI that takes a question and produces text, code or images, and these two are each company's main models. Choi observed that such models now arrive about every 70 days on average.

In the episode, Choi said he built a role-playing game with Opus 5.5 in a single day while talking it through with his child — about two hours of play to the ending. Writing the first request took 10 minutes, and the first version, with a village and the neighboring map, took 20 minutes to generate. He used no image or music files at all; the character art and background music were all made in code. Code is the set of instructions a computer understands.

Even with tools like this, people keep asking why company productivity hasn't risen to match. Roh's answer is short. Where many people take AI seriously and do something with it, change happens fast; where people have decided not to, nothing you show them works. So it is not a problem with AI but with the people using it.

Another observation followed in the episode. People who tried AI in many ways and noticed early how much it can do are moving ahead, and over one to two years of watching, even those who weren't are gradually coming over to heavy AI use. The change comes, but it reaches each person at a different time.

Three sticking points — what a four-week class showed

Choi said that in a four-week class he recently finished, students produced interesting work, but the sticking points were clear. They come down to three.

First, not knowing what is possible. He said opening up the ceiling mattered. When people first research how far AI can go today and try it themselves, they find things to imitate or adapt. Without that, they use even the best model beside them only as much as they always did.

Second, stopping too early. He said you sometimes shouldn't hit the brakes too soon. There are stretches that repeating 'do it for me' won't get you through, and stretches you only cross by focusing and pushing hard. If you stop because the first result is below expectations, you never cross them.

Third, not knowing the words to ask with. His examples were Three.js and MediaPipe. Three.js is a toolkit for drawing 3D scenes inside a web browser, and MediaPipe is a Google tool that reads hand and face movements from camera video. If you don't know these names, you can't put them in your request even when the AI could do the job. An expert in a field knows its concepts, key figures and history and can draw on them when needed; his worry is that beginners skip that process entirely.

Roh replied that this observation would apply equally to companies and business, not just to art.

You see it only by doing it

Choi introduced the term 'epistemic action', which came up in a conversation with a model. In Tetris, to know where a block fits you have to rotate it. Some things don't show up by predicting in your head; you see them only by doing.

Roh said that people who have tried it and people who haven't are now separated like oil and water. He added that after a one-week business trip, the world had moved so fast in the meantime that his head felt detached. The shorter the gap between new models, the more it costs to put off trying them yourself.

Three things a team can try this week

Each one answers one sticking point. Before buying a new tool or scheduling company-wide training, these are sized so that a single team lead can do them this week.

First, look at three things others have built, together. Find three demos that other people built with models released this month and watch them as a team in a 30-minute meeting. The goal is not admiration but writing one line each on 'what would change if our work could go this far'. You have to know the ceiling before your requests grow.

Second, ask the AI for the terms you need first. Write down what you want to build in one sentence and ask first which tools and terms you need to know to build it. When the AI brings up names you didn't know, you can put them into your next request. The reverse holds for your own field: the technical terms of your work are what you teach the AI. Already knowing them is the strength of people who do the work.

Third, keep refining one task to the end. Even if the first result is poor, don't give up right away; point out specifically what you don't like and ask for the same task again more than five times. Only people who push like this know where it breaks through and where it gets stuck.

I am not a developer. I want to build the following with AI.

[One sentence on what you want to build. Example: a screen that gathers weekly order volume by client and shows it as a table and chart]

Tell me only the 10 tools and terms I need to know to build this.
For each one, add a one-line explanation and an example sentence I can use when I ask you again with that term.

Basis and limits of this piece

This piece is based on the conversation in AI Frontier Korea EP 116 (recorded September 27, 2026, with Chester Roh, Seungjoon Choi and Jonghyun Park). It reflects the three hosts' observations and opinions, not statistics measuring company productivity. Quotes are rendered for meaning from the captions.

So this piece does not prove why productivity isn't rising. It lays out what to try first in a situation where people holding the same tools get different results. 'You see it only by doing it' applies to this piece too.