AI Thinks, AI Codes... So What's Left for Us?

AI large-model progress moves at a blistering pace these days, yet people seem to have gone numb to the headlines — a bit like the letdown after the moon landing. But there's one interesting new development worth flagging: OpenAI has rolled out its reasoning model, GPT-o1.

OpenAI didn't hype GPT-o1 much, and multi-step reasoning techniques have been around for a while. But OpenAI's knack for packaging this into something ordinary people can actually grasp is something few companies can match, which is why the technology still feels genuinely impressive. The key is that this model can reason through logic — it has a chain of thought. Once you've used it, you can basically sense the model simulating thinking, and it even tells you it's "thinking." A lot of users find it fun to pose a hard problem and watch the model chew on it for a while; stump a large model, and hey, it's a nice little ego boost!

Reasoning Gets Hot Again — Cheaper and Less Error-Prone

Ask the model how it actually reasons, though, and GPT will only give you the generic explanation of how models work in general. OpenAI itself has been very tight-lipped on the subject — beyond saying it relies on reinforcement learning, the company doesn't seem eager to reveal much about the underlying mechanics. What makes the model look like it's thinking is mainly that it starts breaking a task down into steps, which closely mirrors how humans reason.

This kind of "thinking" model isn't actually new. OpenAI research scientist Noam Brown previously built a formidable game-playing AI that stunned the field on its 2019 release — the best poker-playing AI of its time. His early code kept losing, until he discovered that a planning-and-reasoning architecture worked far better than other approaches. In a recent talk, he laid out just how much room there still is to improve reasoning, and how worthwhile it is to keep investing in it. Right now GPT's chain of thought only runs for a few seconds, but Brown believes future multi-step reasoning models will eventually think for hours at a stretch — which starts to feel eerily human.

The cost of training ever-larger models is a fact of life, which is why this not-so-new technical path of improving reasoning has heated back up — its cost-benefit ratio is quite good, and it offers real relief as today's models keep growing bigger and more expensive. The benefits of reasoning ability go well beyond that, too. In traditional generative domains like image generation, it may not move the needle much, but for math, game-playing, coding, and explaining things, it can noticeably outperform other approaches, since these tasks benefit enormously from planning. On top of that, the biggest headache in generative AI right now is "hallucination." Hallucinations have many causes — one is problems in the underlying training data, which is a thorny issue to fix — but a chain of thought can meaningfully cut down on them.

Speaking of thinking, coding is already being reshaped by AI in a big way. There are quite a few popular self-serve coding tools out there now; among general-purpose models, Claude tends to be the crowd favorite. If we're talking specifically about coding, the SaaS tool Cursor has become close to the de facto standard among practitioners. Another one, Replit Agent, also has a large following. Between the two: Replit is better at getting something built and deployed fast, though the output quality isn't as strong; Cursor, on the other hand, genuinely fits into a real production workflow.

My gut feeling is that both tools are genuinely impressive, and my sense is that within another couple of years, most low-level code will be something you can just have written for you with minimal effort.

After translation got eaten away piece by piece by machines, coding was still supposed to be one white-collar job that needed a human. So it was notable that Nvidia CEO Jensen Huang floated the idea last year that "everyone will code," and this year former Google CEO Eric Schmidt has said something similar — it seems to be emerging as a consensus among Silicon Valley's elder statesmen. But if everyone gets to practice this skill, do we still need computer science degrees? For the past few years, the conventional wisdom was to push your kids into computer science if you wanted them to have a future. If everything turns into natural language, are we supposed to tell them to major in Chinese literature or foreign languages instead?

I don't think the immediate impact will be this dramatic, but people in the software industry can't help worrying that if everyone can code just by talking, simpler software will stop being something you can sell in the market at all. Is that alarmist? Not necessarily. Take spreadsheets: they used to be a fairly specialized skill, and now spreadsheet software is something anyone with a computer can pick up and use. Go back further, and didn't the same thing happen when the Venetians invented double-entry bookkeeping?

Don't Panic, Programmers — Let AI Assist Your Coding

But for the fairly long near term, humans writing code with AI assistance will remain the better setup, so nobody's job is about to vanish overnight. It's like how spreadsheets today are still most efficiently built by people who actually understand accounting. But eventually, at some final-form version of this, humans probably won't be needed much at all — or the work will no longer follow logic we can currently understand. Self-driving cars make this easy to picture: at true L5 autonomy, no human driver is needed, and that day will arrive sooner or later regardless.

First translation stopped needing us, then coding starts to follow. What comes after that? Eventually the entire chain of thought — all our reasoning — could be handed off to AI, and one thing seems safe to say: the end result will be better than what we'd produce ourselves. There's a well-worn Chinese phrase, "the more you think about it, the more terrifying it gets," and it genuinely fits this situation (the good news is that by then, we'll have outsourced our thinking and reasoning to AI too, so we probably won't find it "terrifying" anymore — might as well grab some ice cream, binge a show, and relax). Also worth noting: once OpenAI ships something like this, rival products with their own distinctive twists tend to show up within about six months, so the market should soon have plenty more reasoning models to choose from.