A few things happened in September recently, and they were pretty mind-blowing.
The first: GPT6’s Computer Use feature has gotten much more polished. By burning more tokens on Computer Use, plenty of people are using it to go all the way from modeling to shipping a final product. Actually, even without GPT6, you could do the same thing just by installing the Blender MCP. It’s just that GPT6 handles it all end-to-end with Computer Use in one place. As long as you’ve got deep pockets and unlimited compute, you can hand all sorts of things over to it. For example Hangzhou Roam, GTA6-Shenzhen.
That’s the first mind-blowing thing.
The second mind-blowing thing: recently I came across a project that’s gone viral — uploading a fruit fly’s neurons into a computer and using the fruit fly’s compute to drive all sorts of uncanny things, like piloting drones or playing Doom. Anyway, there are a lot of wild projects built on the fruit fly’s neural wiring.
The fruit fly’s neuron count is far smaller than the scale of today’s large models, but when it runs inference at full power, the level of intelligence needed for embodied intelligence — movement, obstacle avoidance, and so on — can arguably already exceed the intelligence level of today’s VLM large models.
The third mind-blowing thing: a math problem that had plagued humanity for 90 years was recently solved by a large model. OpenAI used an unreleased, stronger model to do long-horizon, long-task, multi-agent coordinated reasoning and cracked it.
Which is to say that many of the math problems that plague humanity can now be solved by having a large language model drive long tasks.
Does AI have a boundary?
Does the AI large model have a boundary?
If it does, then the optimal strategy for all of us should be to pivot toward cultivating the skills and qualities that lie outside AI’s boundary. If it doesn’t, then the optimal strategy for all of us should be to find some way to get onto the compute side.
Within certain scopes, and where there’s a measurer, AI’s capabilities really can pull themselves up by their own bootstraps, and do it well.
From that angle, we humans may need to develop the right measurers. But the workload won’t be large, the relevant job openings won’t be many, the job opportunities will probably be few, and only the very best people will be needed.
Fortunately, the greater opportunities actually lie in the directions where AI has a boundary — which is what I talked about in an earlier article: creativity, and the direction of producing things in response to human emotion — AI cannot replace humans there.
Because AI is just a word-chaining machine, just a high-probability event predictor. It can’t generate human emotion; it doesn’t have that underlying mechanism, that mechanism of biological operation. Humans are driven by emotion, mood and events together, whereas AI merely states high-probability distributions over words, their hits and their rehearsals.
So the way humans far exceed AI lies in everything that’s driven by emotion as output, rather than output driven by logic.
Because logic, no matter what, ultimately comes down to a math problem. The advanced logic of daily life is just a probability distribution math problem. Large models have all of that at the level of underlying principle.
Combined with what I saw a few days ago — our mathematician and Fields Medalist Deng Yu said that the math world now feels permeated with a restless, unsettled atmosphere. But from his viewpoint, too, you just have to wait for this wave of the past couple of years to pass, let the bullets fly for a while, and the math world will naturally and gradually adapt, and a new ecosystem will take shape.
Anyway, if AI can solve all math problems, he’ll retire from the math world and start writing yuri novels.
I think his choice is brilliant. This is also why I, a coder who writes programs and has used AI for so many years (especially with this year’s explosion of desktop AI Agents, which I’ve been using heavily), decided to write romance novels (/doghead):
Because that’s the place farthest from AI, and it’s humanity’s reservation.
How human learning and career strategies are shifting
A few days ago I also saw Professor Ma Zhaoyuan of the Southern University of Science and Technology say: leisure is the precondition for creativity. So we humans really can’t keep grinding like this. All this grinding is just borrowing against the future.
What humans really ought to do is, around their own emotions and biological characteristics, create content that is genuinely human and one-of-a-kind, and achieve each person’s self-realization.
Also, a couple of days ago I saw an interesting report. Roughly: they studied a cohort of Chinese students, more than 20,000 of them, over a span of 30 months, looking at how kids using AI in their learning affected their academic performance. So this data was pretty hard-won, and pretty valuable as a reference.
The study found that after students used AI to do their homework, their average homework grades went up, and the time to finish dropped a lot. But in closed-book exams without AI, their average scores fell.
The interesting value here, or the interesting insight, is this: although on average overall, scores dropped once AI was taken away, there was a subset of students — the ones with genuine self-drive, who actually want to learn and are interested in learning — who didn’t spend less time doing homework with AI, but actually more. And their exam scores without AI were higher instead.
So learning with the help of AI and using AI to complete the learning tasks assigned by school produce two utterly different outcomes.
To begin with, in principle, a large model’s prediction mechanism predicts based on the human average line. And the fact that this industry’s main users need programming problems solved pushes demand even more toward the most predictable results.
So AI large model vendors will inevitably make AI more mediocre. Why? Because every time something new comes out, those reviewer folks test it by asking it to draw a pelican on a bicycle. But the moment it’s even slightly non-mediocre, there’s no way to guarantee that the thing tests well every time, right? So the more strongly those reviewers measure its ability, the more it’s required to grind problem sets and grind for scores on exam papers. That’s exactly like what exam-oriented education does to people — in the end it fucking loses its soul.
But human creativity is different. Everyone has their own unique life, their own unique experiences, their own emotional feelings, and their own innate biological reasoning ability and the emotional feedback mechanisms it forms, so they have their own unique taste. That is inherently distinct from AI.
So every person is one pole of the world, but AI is forever on the world’s average line and can never lean toward any pole. Because the moment any vendor makes something that leans toward a pole, its pelican-on-a-bicycle test will fail. Take Doubao and Gemini — these two are the light of humanity in today’s AI world; on top of rationality, they’ve kept a little of the spark that mimics human creativity.
I’ve said this much and can’t be bothered to sum it up myself, so let me have AI write the ending for me:
In the end, AI can handle all logic, compute, math and repetitive tasks — it’s a maximally efficient tool — but it can never replicate human emotion, passion or inner creativity. In the future, don’t blindly grind away at compute and logic; learn to use AI to amplify your leverage, go deep into the emotional creation and self-expression that are unique to humans, and hold on to the spiritual reserve that belongs exclusively to humans. That is the most reliable survival and growth strategy for ordinary people.