跪拜 Guibai
← Back to the summary

Future Companies Need Only Two Kinds of People: Goal-Setters and Hard-Core Geeks

It's been a while since I wrote on Juejin. I've had some deeper thoughts recently, so I'll write them down.

What is intelligence

With large models where they are today, I believe we engineers no longer agonize over whether they truly possess intelligence or are just parrots echoing compressed data. Once you've seen what large models can do, all we think about is how to put them to use. Leave the philosophical questions to the philosophers to ponder at leisure. It's like Schrödinger's cat: many interpretations try to explain where the cat is in its superimposed dead-and-alive state, but pragmatic scholars just want to harness quantum mechanics for humanity.

I sometimes think this is truly the best era for programmers. We always knew the importance of documentation and unit tests, but few programmers actually executed on them consistently, due to human laziness, project deadlines, or skill differences. Now everyone can quickly produce massive amounts of documentation and unit tests. These things aren't lacking anymore; they've become overwhelming.

This is also the worst era for programmers. As the first frontier to be conquered by large models, we've witnessed AI's power sooner than others. Even if we practiced on ten thousand Cherry keyboards, we could never match AI's terrifying code output speed, and it's fast, accurate, and good.

It's impossible not to feel anxious. We both enjoy AI-assisted coding and worry about our future livelihoods. I went through that process too, but I've developed my own reflections.

Imagine a hundred years ago, humanity's stars shone brightly. We could rattle off the names of the geniuses who drove human progress, from Newton to Einstein, from Edison to Turing. Now people ask: why no such progressive figures emerge in this era? Clearly, our society continues to advance, but the names we know well have become Apple, Nvidia, Xiaomi, Tencent. People increasingly realize that an aggregated, effective human organization can achieve more lasting life (compared to a short human life). That's the answer.

If we decompose an organization, we find that people are both the most important and the least important.

I'm not talking nonsense. By "most important," I mean that any work item in an organization must ultimately rest on a person. By "least important," I mean that an organization fears irreplaceable people, because that becomes a constraint.

Organizations need people, which is easy to understand from a software engineering perspective. An organization needs an abstract person, or rather, an intelligence. It doesn't care whether the person is called Zhang San or Li Si; it's best to provide a universal alias, which is exactly what "flower names" (花名) are. This person connects to the organization's document repositories, codebases, OKRs, KPIs, and so on, thus providing intelligence aligned with the organization's direction. If that person is replaced one day, the organization's direction doesn't change.

Let's now look at large models. Can they provide the same function?

Let's break down the work of an abstract person in an organization. I divide it into three parts: harness, loop, and graphs. Harness refers to the codebase, permissions, and document library provided to this person. Loop is the daily check-in; otherwise the organization won't pay them. Graphs means the person understands the organization's hierarchy and the connections between people, because human connections are key to getting things done.

If you follow cutting-edge engineers like Potato Sister (@poteto) or Karpathy, you'll often hear these words. Earlier I thought they were teaching us how to use AI better. Now I see they're grasping a more fundamental question: if the intelligence an organization needs can be provided by AI, does the organization still need humans? I believe different people have different views; feel free to discuss in the comments.

New forms of organization

I was previously pessimistic about this, believing humans would ultimately lose, and that on Boss Zhipin (Chinese job app) there might appear a job seeker named Zhang San, whose university is OpenAI, who can do everything: frontend, backend, operations, and who doesn't need leave or have a family. The scariest part: he costs only $200 a month.

How can we compete? Should we just let it happen and wait for that day?

When I reasoned backward from this concept, I reached the opposite conclusion, so I think that day will never come. My thought experiment: if all humans in an organization were replaced by large models, what would happen? Obviously, it would stagnate because it lacks goals.

This stems from a fundamental difference between large models and humans. Humans are bound by Maslow's hierarchy of needs, with various internal needs, and therefore have goals. Whether the goal is grand, like Musk's mission to put human footprints on Mars, or small, like working to have food tomorrow. Large models have no such needs, so they have no motivation to set goals.

Thus, humans in an organization will shift from providing intelligence to providing goals. There's a funny term recently: "wish-style programming," which illustrates this well. But it's too general; actually, proposing a reliable, reasonable goal is not easier than executing an existing one.

Going further, I think future organizations will have people playing two roles:

  1. Deep-tech geeks

These people will tackle the hardest, most intractable problems in the organization. AI is just their assistant, not yet worthy of replacing them. Their other abilities might be zero, like being unkempt or speaking oddly, but their specific capability has been stretched to the top.

  1. Goal-setting bosses

These people deeply understand harness, loop, and graphs and can implement them. They can also propose goals, let their agent team do the work, and they handle acceptance and thinking.

I believe the direction of organizational change is toward a human-and-large-model model. Our generation will suffer somewhat, because some people will inevitably go through a transition process, stumbling for a long time, and eventually awaken to these two new forms of occupation in future society.

Of course, maybe my reasoning is completely wrong, and this is just me spreading anxiety. But I'm not guessing without basis. While I write this, my agents are still working for me. I don't give them prompts; I just review their daily reports, like a boss. I give them three things: harness, loop, graphs, and my goal.

If you want to hear my humble opinion on how to build this agent loop, feel free to like and comment to give me motivation.

Organizational change is not instantaneous; organizations themselves must endure the pain of transformation. I don't like how some media hype up mass layoffs and immediate unemployment. Trust me, you can still find a job, but if your intelligence is partially replaced by AI, your salary will also be partially replaced by AI.

Organizations need time to adapt to new structures, but salaries never lie, and they are the fastest indicator of effect.

All-knowing intelligence?

One final thought: suppose AI keeps developing and eventually produces an omniscient entity that knows everything and can hand-craft a new quantum mechanics. Where do we humans go?

Although this is also a philosophical question, I want to explore it deeply. Perhaps humans cannot create this omniscience.

There's a view that human-made AI will never surpass humans; we need AI to create AI, the so-called ASI and AGI. Because current AI still has a problem: their weights are fixed. After training, the weights are frozen into a checkpoint. They cannot reconstruct neural connections like a human brain, so they cannot evolve.

Thus, there are two hopes: one is to have old AI train a new AI, reconfiguring its weights; the other is to let AI modify its own neural network, that is, be in constant training.

Here's a question: can all intelligences in the world be merged? For example, a person who is an expert in natural sciences, also good at art, also expert in organizational management, and has high emotional intelligence. I think that's impossible; there may be some antinomies.

For instance, rigorous intelligence and divergent art are opposite. You can't mix them into one thing without producing a freak. This seems visible in GPT and Claude models. Some say GPT models are good at fixing bugs and handling logic puzzles, but their design ability is a mess. Others say Claude models handle backend code poorly, but their aesthetic sense is top-notch.

So perhaps Opus 5.5 is a benchmark for large models, and maybe there won't be a better model (I'm just guessing).

But from another perspective, how fast has humanity developed in just a few decades? From a global view, human intelligence has reached a considerable height. Perhaps after merging with large models, all of humanity's intelligence will expand to every corner and every phenomenon of the universe. Isn't that a new kind of omniscient intelligence?