Taste and aesthetics
In the AI era, every programmer has turned into the exact kind of team leader they used to hate:
- Six months without writing a line of code, coding ability badly rusted, still calls himself “a technical guy”
- In meetings, takes all the work the frontline devs did and calls it his own achievement, and even gets someone else to write up his own report
- Barks orders with no fucking clue, and when something breaks, dumps it on the people below for not executing properly
- Doesn’t know the actual state of the project, can only push empty demands like “did you add unit tests”, “performance still needs optimizing”, “we should get the docs written down”
- Basically can’t read a PR, can only wait for someone else to review it and then thumbs-up along
- When a bug shows up, mindlessly forwards it for someone else to look into, and his own job is forwarding the messages back and forth
- Also complains all the time that “this year’s new grads are getting worse and worse” by 象牙山劉能
That’s a joke, but I do think taste really is getting more important in the AI era. Everyone from here on is going to be that kind of tech lead, so how do you keep from being the lead who drags the team down and become the one who takes the team a level up? It comes down to pointing clearly at the right direction. But what the right direction actually is, that’s exactly where tacit knowledge comes in, or in other words the feeling you get when something is off. Call it “taste”.
Plenty of articles have talked about taste, but only the third one links it to tacit knowledge, and even there he doesn’t go especially deep into how this increasingly important skill is actually built up. To me the word “taste” isn’t some black box you can’t touch or can’t learn. Like any tacit knowledge, taste can be picked up through expert training methods such as naturalistic decision making.
But I’d rather draw the analogy between taste and aesthetics, because both are hard to put into words and both decide direction at the moment it matters. When I look at a piece of code and feel it’s wrong, and have a view on how it should change to be better, that’s the same as looking at a design, feeling it’s off, and knowing where it should go. Taste and aesthetics both mark out a heuristic judgment about direction. Based on what’s in front of me right then, I notice the “bad smells” immediately, find them hard to swallow, and know they have to change. So I think building taste is like building aesthetic sense, and there’s a clear route to be found. The simplest one is to look at how art trains aesthetic sensibility.
Getting to good taste is simple in method, exactly the way naturalistic decision making shows it. Exposure to a large volume of material with expert judgment attached to it is the best road for learning an expert’s taste and carrying it over. Watch, and immerse yourself in catching the signals in an expert’s decisions. What kind of “smell” are they picking up when they decide something is “bad”? More than catching the expert’s logic, taste is a heuristic intuition, so catching the signals that “set off the intuition” is what counts.
So I think education at the applied level is going to have to take building this “taste literacy” seriously, because as LLMs get stronger, spending a lot of time learning and mastering one specific “operational skill” becomes very poor value. An LLM picks it up without needing to be trained at all, so if I pour that much time into a skill whose marginal return is close to zero, isn’t that just wasted time? The other way around, building “judgment” becomes the part that matters most in applied fields. And building that “judgment” is more or less the same thing as building “taste”.
So the old applied training system has to change, or it will only fall further behind the times. But for knowledge that isn’t in an applied field, I still think accumulation is critical, because accumulation is part of how the literacy gets reached, and it’s the road you have to walk to build the harder-to-copy kind of “taste”. In applied work I can judge whether something is right because a large body of engineering experience handed me many signals with feedback attached. But for theoretical research knowledge, especially the networked, “priming” kind of innovative knowledge, that “decision with a feedback signal attached” scenario just isn’t there. Theoretical “priming” demands “accumulation”. Only when the accumulation is enough can a knowledge network get built, and only then does association across that network become possible. So this part still takes time as I see it, and can’t be rushed.
There’s an opposing position, of course, which says wouldn’t an LLM be even better at building that kind of knowledge network? Look at mathematics now, where scientists are nearly like engineers, making judgments in the key areas and getting effective research out of it. Yes, that’s true. But the answer is already contained in the question. Where does “the scientist’s judgment” come from? As I said, it’s exactly the result of accumulated knowledge. If a mathematician hadn’t been soaking in his own theoretical field day after day, how would he notice the LLM went the wrong way, or notice the “bad smells” in the progress of a theory? So soaking in knowledge is still critical, and for theoretical research in the LLM era it only gets more so, because you have to tell what’s garbage from what’s gold. Only the right direction lets us dredge something useful out of the LLM’s enormous corpus.
That said, I don’t think LLMs will completely replace the position of human experts with deep knowledge. To make a bad comparison, it’s like machine-cut noodles against hand-cut noodles. If you’re just eating noodles, the cheapest is obviously the one to get, but if what you want is “noodles a person cut”, then the handmade one is clearly better. So there will certainly be a few niches that still need human experts, but like the high cost of hand-cut noodles, this is bound to stay a market minority, and only in some heavily regulated industries like finance or healthcare will anyone choose to carry that cost in exchange for lower compliance risk. For most consumers in the market, they mostly don’t care who cut the noodles. I just want to eat the noodles, and it only has to taste good.
So an outcome-oriented market points at LLM substitution at scale. The demand for low cost together with speed and efficiency is itself a reverse selection pressure that weeds out the organizations that still can’t work with LLMs. When your competitors deliver higher-quality results in less time and you’re still holding to the old ways, you owe the client an explanation. Why aren’t you using AI? Unless your client’s preference really does land on that “special” requirement of being done by hand, an efficient LLM is going to be the basic standard. And I don’t think this trend reverses.
Take copywriting. It used to be a well-paid profession that took years of practice to master. Even after e-commerce and ad tech drove a surge in demand, that changed slowly as more and more professionals poured into the market. Today, LLMs have already destroyed the livelihood of the overwhelming majority of those professionals. The reason is simple. Most of the demand comes from small companies that need copy, and copy generated by ChatGPT covers what they need perfectly well. Some copywriters do still get hired to prompt, review and send the copy, but demand isn’t infinite, and there’s no way to hire everyone to do that work. One copywriter now does the work of ten, and demand is fixed. Supply going up tenfold doesn’t make demand go up tenfold. The best copywriters still have a market, of course, but they’re about 1%. The other 99% are scrapping over the leftovers. UX writing used to look like a career with a future. Now every one of those people I know has been laid off. Even large companies let them go. You can generate text labels with ChatGPT and it’s good enough 90% of the time, so hiring ten professionals makes no sense anymore. Cut nine, keep one. by Replies to comments on my “LLMs are eroding my career” post | the human in the loop
If the models keep improving in the same direction, we all end up facing the same fate.
How AI Is Transforming Work at Anthropic \ Anthropic How to Be a 30x AI Engineer with a Taste What Do Engineers Mean When We Say “Taste”? LLMs are eroding my software engineering career and I don’t know what to do | the human in the loop
Translation note. I wrote this in Chinese. This English version is an LLM translation, so the wording is not mine even though the thinking is. Original: 品味和审美.