In Taiwan, in recent years, we have been talking about digital optimization and digital transformation. My view is that in the AI era, there is no such thing as AI optimization, only AI-native: things that could be optimized digitally in the past have already been handled well by software, and what we should look for is what the past software systems couldn't do. Sam Altman explains this more clearly in this 69-minute interview, saying that today's startups still look like startups from ten years ago, and most people are just 'using more Codex', which doesn't look sufficient. This article first explains my judgment (including a criterion of looking at multiples rather than percentages), then compiles three things from his interview that are truly worth using, with a first step you can take this week.
- People who are already using AI every day but can't explain 'what structural difference my work has from six months ago'
- People who want to introduce AI into their team but can only think of 'accelerating existing processes' as the direction
- Experienced workers who have a lot of experience but can't explain it or teach others
- In the AI era, there is no AI optimization, only AI-native. Things that software has already solved well don't need to be polished again with AI
- Judging whether something should be handed over to AI, look at multiples rather than percentages. Only improving 20% to 50% is probably not AI's strength
- The most valuable thing you have is exactly the 'something you can learn but can't teach' part, and that part now has the ability to be externalized
Watch the full video ↗
I. My view: No AI optimization, only AI-native
In recent years, Taiwan has been talking about digital optimization and digital transformation. My own view is that in the AI era, there is no such thing as AI optimization, only AI-native.
The reason is straightforward: things that could be optimized digitally in the past have already been handled well by software. Now, using AI to optimize those things is mostly just polishing an already solved process a little more. I think what we should do instead is look for things that past software systems couldn't do and use AI to do them.
Already written system. It is pre-recorded. You press the button, and it runs along the predetermined path once. For things with a definite answer and fixed task flow, it has already done well.
On-site immediately generated system. It is generated and changed in real-time. Things that past software could not solve, and that carry uncertainty, are the ones that it is worth trying to use.
This contrast is not something I invented. I heard a speech about this before, and I have always remembered it: all the software and programs we have had in the past are actually already written systems, while AI is an on-site immediately generated system. These are two different things.
I am not trying to compare who is better. What I want to say is that they are applicable in different situations. For things that software has already solved well, it may not even be necessary to use AI to optimize; for things with uncertainty, things that past software could not solve, it is worth trying to use current AI to solve.
A very useful criterion: look at multiples, not percentages
I myself use multiples to determine whether something should be handed over to AI.
If a task only improves by 20% or 50% after using AI, I think it is not really AI's strength. Because if this task is really suitable for AI to do, you will immediately feel it, it is a jump of two to three times, or even dozens of times. At that moment, you will know: this task, this process, is especially suitable for using AI to amplify.
So when you find that your use of AI only makes you "faster", that signal is clear: you might be optimizing something that already has a solution. Writing letters faster, finding data faster, making presentations faster, these are all true and are good things, but such usage won't change the shape of your work in two years.
Sam Altman's interview explains my previous view more thoroughly and in more detail. So I have also summarized his content for you.
- Why Your AI Adoption Isn't Working: From the Three-Layer Theory to the Organization's Wall This article talks about how individuals should judge where to use AI, while the other talks about the two walls that the same judgment would hit in an organization: direction stopping at node acceleration, and efficiency increasing tenfold while salaries remain unchanged.
II. His statement
At the end of the interview, the host asked him: Are there any thinking frameworks you used to rely on that are now wrong?
His answer was roughly: Today's startups still look largely like startups from ten years ago, because existing wisdom tells you to do it that way. Of course, some things are different now, people will say I should hire fewer people and spend more money on tokens. But it should look very different. He has met a few people who are really doing it in completely different ways, but most people's approach is just "using more Codex", which doesn't look sufficient.
▶ 01:09:15The five words 'use more Codex' apply to any AI tool you have.
Use more ChatGPT. Use more Claude. Use more automation.
The shape hasn't changed.
In the same interview, he also mentioned another thing that can serve as context for this judgment: he encountered a startup team of about two weeks old that had completely rebuilt their entire office productivity toolset, designed for a world where AI would directly use documents, presentations, and spreadsheets. He said this was a full year's work not long ago.
▶ 00:00:34A ten-week startup can now do what took a year in the past. The thing itself is not unusual; what's unusual is that most people use this capability to do 'things they originally wanted to do, just slower.'
He himself put it more directly: the current temptation is to use today's agent to pick up easy wins. He completely understands why people choose this, and it's mostly going to succeed, but it will be very competitive. What surprises him is that, with the ground shaking so much, there aren't more people using this new set of tools to do that crazy thing.
▶ 00:02:43The next three paragraphs are three judgments I picked from this interview that I think can be directly applied to personal work.
Three, the first thing: start doing those 'not yet profitable' things now.
Sam mentioned in the middle of the interview what he considers the most important and hardest for people to accept: to genuinely believe that the scaling law will continue. People should plan for things that may be impossible or unprofitable this month; these things might become possible in two or four years. He said the market has not yet fully adapted to this. Similarly, the market has not fully adapted to 'betting on high-growth young founders.' The money is still available. He stated that if he were still giving advice to entrepreneurs, this would be the most important thing he would want people to understand.
▶ 00:04:22He himself also admitted to making a mistake in this area. When the host asked him when was the last time he realized he wasn't ambitious enough, he said: the investment in computing power was clearly underestimated, and he could have calculated it correctly with the right mental model, but was scared off by things like the financial market. That was clearly a mistake.
▶ 00:48:13People at that level will be blocked by 'not yet profitable,' and ordinary people will even more so.
What this means for you
The key is to avoid betting on a capability that does not exist yet. Instead, we need to find things that are difficult to do now. We should start laying the groundwork for those things that would be smooth if the model were better.
The difference lies in what you're laying the groundwork for. If you invest today in the operation skills of a certain tool, the model update will render it useless. If you invest in writing down your judgments in your mind and organizing them into data that AI can understand, then the stronger the model becomes, the more valuable that batch of data becomes.
Most of what I've done this year is the latter. Rather than waiting for a tool to become stronger, first accumulate the things that can be directly used when the tool becomes stronger.
The first step you can take this week
Take a piece of paper and answer three questions:
- What is currently 'something AI can help with a little but not fully,' so I've put it aside and not done it yet?
- That task couldn't be done because of lacking model capability, or because I didn't share my knowledge?
- If the latter is missing, which part can I hand over today?
Second question's judgment clues: If you have to explain the background to the AI every time, and it only does it right after you explain, then the problem is that you didn't share your knowledge. If you've explained clearly and it still can't do it, then it's a model capability issue.
Those who have the answer to the third question will have something to do this week.
- Models are like supercars, and your knowledge base and workflow are the road you're driving on This section only explains why we need to lay the foundation now. The other article explains what the path itself looks like and the order of laying it.
- Knowledge Operation System Mother Architecture: Designing a System That Continuously Distills Tacit Knowledge into Executable Judgments If you want to know what the final structure of the "judgment of sharing out" will look like, that article is the complete system map.
Four, the second thing: dig out the part that is "learnable but not teachable"
This is the most valuable part of the entire interview.
The host asked him: how did you become good at dealing with this constantly changing chaos? He said it was through practice. Some things, no matter how thoroughly you understand them intellectually, require a lot of practical attempts to be able to handle them emotionally. His conclusion was: this kind of thing can be learned, but can't be taught (learnable, not teachable). He then said the real weakness of young founders is that they haven't accumulated enough career experience yet, and haven't reached the emotional calm of coexisting with chaos, so they will be very painful in the early stages, but eventually will learn it, though at a great cost and with many unnecessary mistakes.
▶ 00:04:55He then explained the method: if you want to learn this kind of thing, ask that person to explain it to you. According to his experience, it has never been useful. What's truly useful is to study him, sit next to him in meetings, observe yourself, and learn yourself. When he wanted to get stronger, he did exactly that, staying close to that very strong person and deeply studying. He added: the other person probably can't teach him through explanation.
▶ 01:03:58He also mentioned the other side of this when talking about design. He said the most he learned from working with Jony Ive was that that person would first thoroughly research the problem, even not letting himself think about solutions too early. Truly good design involves understanding the problem much more than a sudden insight.
▶ 00:58:37What this means for you
The most valuable part you have in your hands is usually this kind of thing that you can't explain clearly. How you look at a client, how you judge something strange in a proposal, how you decide whether to push or wait in a meeting. These things work well for you, but you can't explain them, so you can't take them away or replicate them.
This used to be unsolvable, because it was hard to really have someone around who could sit and watch you work all day.
Now it's possible. AI can be that person sitting next to you, provided you let it see the process, not just the conclusion.
"I need a proposal, please write it for me." This is asking for the conclusion. AI always starts from scratch guessing what you want, and you have to explain the background again every time.
Let it ask you the opposite, and it should ask you at the moment you make a decision. After it asks, you will have an additional set of judgment rules written in your own words.
The first step you can take this week
Next time you make a judgment (pick one that you do quickly and almost without thinking), paste the process to AI, and use this direction to ask:
Do this three times, and you will have something that others can't take away, and the model won't lose it even if it changes.
- What are Claude Skills? Turn your professional workflow into a knowledge asset that AI can repeat. This part talks about how to get the judgment out, and the next article talks about how to turn the extracted judgment into something AI will do every time.
Five, the third thing: distinguishing real trends from fake trends.
After the interview, the host said he learned one of his favorite mental models from Sam, which is distinguishing real trends from fake trends.
The host explained this model as follows: fake trends are situations like VR. In these cases, there is a lot of discussion and some people buy into it, yet they do not truly love it or design their lives around that experience; eventually, the trend is put on the shelf. Using ChatGPT presents a different case. The speaker uses it almost every day. Sometimes he uses it for three hours in a busy workday. Other times, he uses it very little, yet it remains part of his life.
▶ 01:07:39The host then asked Sam: how do you now judge whether something inside the company is a real trend or a fake trend? Sam's answer is the same principle: whether there is something real, deep, and lasting. He said this framework was developed at YC, where you get a lot of data, and if you are willing to spend time analyzing and understanding, you can really see a lot.
▶ 01:08:11What this means for you
This criterion can be directly used to check every AI tool you introduce.
The only question to ask is: has it changed the way I arrange my day? Whether it's powerful, whether others are using it, is all outside this question.
This criterion's advantage is that it doesn't look at the spectacle, only at the records. The excitement during the trial period is easily mistaken for demand, and checking the usage records usually reveals the truth. By the way, the same tool being false for you and true for others is completely non-conflicting. This criterion is inherently about people, not objects.
This criterion is part of the same group as the multiple criterion I mentioned earlier, just looking at it from a different angle. It looks at continuity: whether this thing stays in your life. I look at the magnitude: whether this thing made you feel a several-fold jump. Only those that meet both criteria are truly integrated into your work.
The first step you can take this week
List the AI tools you are currently using. For each one, answer three questions:
- How many times did I actually open it in the past 14 days? (Check the records, not by memory)
- If it disappears tomorrow, which one thing would be directly blocked?
- Did it help me with that task, making it 20% faster, or significantly faster?
If you can't answer the second question, it means it hasn't entered your work structure. You can keep it, but don't invest more time in it, and don't recommend it when others ask you. If the third question is only 20%, go back to the first paragraph: that task might have already had a solution, and you're just polishing it a bit more.
- When everyone has AI, the real difference is judgment This section provides criteria for screening tools, while the previous article talks about what remains of you after the tools are screened.
Six, other worth-remembering quotes from the interview
These sections are separate from the previous three things. I think they will be useful to keep.
When something is being done, kill off other good things. He said the truly difficult part is not cutting off projects that can't be done, which is actually easier. The difficulty lies in when something starts operating very well, deciding to kill off other good things to let the best one become better. OpenAI has done this a few times: when GPT-3 started to show promise, they shut down their excited robot project; recently, when coding agent started to run well, they pulled back investment in Sora and the browser. He clarified that Sora would have been very successful, but the reason for pulling back was that the computational power and human resources were more important for coding agent.
▶ 00:55:52You have to ask for it. The host asked him about the last time he got something others thought was impossible, and he gave Codex as an example: at that time, they were behind Claude Code in coding tools, and the market consensus was that it was almost impossible to catch up in a category where others already had momentum. But they judged this matter to be important enough, so they asked a team to do it. That team achieved a result he described as very rare in business history. He said if they hadn't asked the team to do it by saying, 'We have an extremely important but extremely difficult task,' this wouldn't have happened.
▶ 00:52:03An organization's speed is mostly determined by who you put in the leadership position. He said various management rhythms and management techniques do have an impact, but mainly it's about people. And most of the time, the company's senior executives should come from within, not from outside.
▶ 01:04:30Strengthen your strengths, don't fix your weaknesses. He said that the obsession of 'I want to improve something I completely can't do and will never be good at' is a big pit.
▶ 01:02:53The opposite of bad experience is no experience. He said that most people think the opposite of bad experience is good experience, so they always want the good one. His idea is: the opposite of bad experience is actually no experience, and in not too far future, you will enter the state of no experience. Once you understand this, you can feel grateful for bad days. He quoted Naval Ravikant's words: If life had a fast-forward button, your life would end.
▶ 00:07:07The last part has nothing to do with work methods, but I still want to include it. When asked about the most painful thing in the past 12 months, he said it was having a child and working hard at the same time, which was very torturous. He felt that he was already a present dad, doing almost nothing except work and spending time with family, but still felt he missed out on too much of this one-time phase.
▶ 01:05:33Seven, going back to the first sentence.
If you now use AI and it only makes things 'much faster', these three things can be considered as three entry points:
Start laying out those things that will happen in two years but need to be done now, and put knowledge out first.
Use the way of being questioned to turn those unclear judgments into your own words written as rules.
The only criterion is: does it change the way you arrange your day?
The common point of these three things is: they won't make your output this week faster, but they will make the things you accumulate over the year different.
Sam Altman's words are worth reading again: most people's approach is to use more Codex, which doesn't look enough.
The issue is that new capabilities are being placed inside old structures. Returning to the first sentence, the focus should move from AI optimization to AI-native solutions. Look beyond "which stage can be 20% faster" toward work that past software could not do and that produces a several-fold leap when you try it.
Sam Altman - How to Start a Startup (Relentless, 69 minutes)↗
After completing that prompt, welcome to share the results.
If you tried that prompt in the fourth section, after completing it, welcome to share it in the LINE community for discussion. The shape of your 'judgment rules' usually shows more about how your work operates than any methodological article.
The community will also share the skill packages and free lecture information I usually use.
LINE Community: Discuss knowledge base, skill package, and AI workflow together.
Join LINE community ↗Skill package download: Hand over the process to AI for repeated execution.
See skill package →