Someone on social media joked that telling an AI to "think from first principles" made it feel like its IQ jumped by about 20 points.
Of course, saying a magic phrase doesn't hand 20 points of intelligence to a silicon brain.
But that's not the interesting part.
What this one sentence asks for isn't "list lots of problems," or "go look at competitors," or "come up with a smarter-sounding idea."
Go all the way back to the roots. Are the current question, goal, constraints and assumptions actually right? Rebuild from there.
That's pretty much a philosopher's job.
If the ordinary kind of improvement is a crack in the wall, so you compare ten brands of patching compound, first-principles thinking stops halfway, stares at the wall, and says:
"Wait. Do we even need this wall?"
Suddenly the contractor has transferred to the philosophy department.
1. "Finding problems" and "doubting the question itself" are different jobs
When you ask an AI for improvements, you usually get one of three things:
- A list of the problems that exist right now.
- Solutions borrowed from other companies, other people, or existing methods.
- A better plan assembled inside the current setup.
All useful. But basically, they all accept the current game board and think inside it.
If the click-through rate on related articles is low, you move the placement. Make the cards bigger. Reword the text. Improve the recommendation model. That's solid improvement work.
First-principles thinking steps back one level before that.
What is the point of related articles in the first place?
Is it to send readers to another page? Or to answer the question the reader will have next? If it's the second, do you really need a "related articles" card? Wouldn't some cases be better answered right in the body text?
Here, before the quality of the solution, the definition of the problem you're trying to solve is what gets audited.
You invited the AI to an improvement meeting, and it starts erasing the "goal" written at the top of the whiteboard.
Which is, more or less, correct.
2. First-principles thinking is less about idea fuel and more about groundwork
Understanding it as "a way to produce lots of ideas" is a bit too narrow.
The core of it is to set the existing solutions aside for a moment and go back to questions like:
- What state do we actually want to reach?
- Which constraints can never change?
- What can we actually observe as fact?
- Which "requirements" are just habits we believe are requirements?
- How would we design this if we didn't know any of the existing answers?
So it's less like picking up seeds for ideas and more like digging up all the soil before you plant anything.
You can line up case studies from ten competitors, but if all ten share the same assumptions, company number eleven ends up with the same building.
First-principles thinking starts with:
"Why does everyone build it this way?"
Sometimes, after looking into it, the answer is "no, this really is the most sensible way to build it."
That counts as success too.
Doubting an assumption is not the same as always smashing it.
If you doubted every road sign and started driving the wrong way each time, that wouldn't be philosophy. That would be an accident.
3. This is where double-loop learning comes in
Double-loop learning, discussed by Chris Argyris in 1977, is about learning that doesn't just fix mistakes but also questions the policies, goals and norms that produced the behavior.
Simply put:
Single loop The result is bad → fix the action.
Double loop The result is bad → before fixing the action, go back and re-examine what you defined as success in the first place, along with the rules and assumptions themselves.
Take "the click rate is low," for example.
With a single loop, you improve the headline, the placement, the color, the recommendation accuracy.
With a double loop, you go back and ask:
"Why do we want a higher click rate in the first place?"
If what you really wanted was "readers get their questions answered and can move on if they need to," then click rate may not be the goal at all, just a stand-in measurement.
And then the design problem changes.
From "how do we get them to click?" to "how do we naturally answer their next question?"
You go from an AI that improves the KPI (the number you track to judge success) to an AI that audits why you picked that KPI, one level up.
That's what's interesting about this.
4. AI is moving into the part we thought only humans could do
In AI automation, the work left for humans has often been described as "setting the goal," "asking the right question," "thinking about meaning," and "approving."
Execution, searching, tallying, summarizing, comparing and generating can all be mechanized.
But
"What are we doing this for, anyway?"
seems like a human sanctuary.
Yet it turns out you can bring AI into that too.
Step-Back Prompting, presented at ICLR 2024 (a major machine-learning conference), proposed having the model step back to a higher-level question or principle before solving a specific problem head-on. Improved performance was reported across several reasoning tasks using PaLM-2L, GPT-4 and Llama2-70B.
This doesn't prove that the particular sentence "think from first principles" is magic.
But
specific problem → step up one level of abstraction → pull out the principle → return to the specific problem
has empirical, closely related precedent when you have an LLM do it.
So the division of labor where "AI handles the fiddly work and philosophy is for humans only" is, at the very least, no longer so simple.
AI has been given an adjunct lecturer slot in the philosophy department.
5. Put this into "germination" and idea-hunting gets two lanes
When you think about where ideas for articles, products, features and research topics come from (what we call germination), you can split the entry points in two.
Germination picked up from outside
- There's search demand
- It's trending on social media
- Readers are dropping off
- Lots of inquiries are coming in
- A competitor launched a new feature
This picks up signals from the world.
But you can build another one alongside it.
Germination by digging inside
- What is this feature there to achieve?
- Does this KPI really represent the goal?
- Is this constraint a law of physics, or a leftover convenience from long ago?
- If we built everything from zero, would we use the same structure?
- If we banned every solution that exists today, what would be left?
- Is this even a problem worth solving?
This digs into the assumptions of the existing system.
If external trends are about picking up "seeds lying on the ground," the first-principles lane is the kind of germination where you pull up the floorboards and find an unfamiliar plant growing underneath.
A question nobody was searching for gets born as a result of taking the existing structure apart.
That's where originality tends to show up.
6. To implement it, split it into three steps: break assumptions, rebuild, verify
If you just tell an AI "go philosophize freely," it sometimes drifts off into outer space and doesn't come back.
In practice, it's better to fix the process.
Step 1: Audit the assumptions
Separate out the current goal, metrics, constraints, customs and existing solutions.
In particular, don't mix together: "facts," "constraints that truly can't be avoided," "organizational convenience," "historical accident," and "things that just kept going without anyone knowing why."
Step 2: Rebuild from zero
Don't treat the existing plan as the right answer. Build several options from only the goal and the unavoidable constraints.
It's fine if you end up with the same conclusion as the existing plan. What matters is that you can explain why it came out the same.
Step 3: Bring it back to reality
Test it against cost, law, safety, user behavior, migration effort, existing data, and so on.
First principles is not a get-out-of-reality card.
"I want to bring server costs to zero, so let's make the server a concept" isn't philosophy. It's running away from the bill.
7. Even so, what's left for humans is "what we care about"
AI can explore, quite broadly:
- Is this assumption necessary?
- Are the goal and the KPI out of sync?
- What would happen under a different value function (a different way of scoring what counts as good)?
- What could we drop to make the structure simpler?
But there's one last line.
Which value do we finally choose?
Maximize time spent on the site? Solve things as fast as possible and send people on their way? Maximize profit? How far do we put trust, freedom, safety, fairness and fun first?
These aren't answers that fall out of observing the world.
AI can lay it out: "if you value A, design X; if you value B, design Y."
But which of A and B to put at the final purpose of a society or a product is, in the end, on the human side.
So the end point of full automation is neither
humans think of everything nor AI decides everything.
It's this:
AI doubts the assumptions all the way down, and humans take on only the value judgments and the forks that carry responsibility.
The human job shifts from "come up with 100 ideas" to "pick this world line."
8. If you actually hand this to an AI, this question is the core
The shortest version is enough:
"What do we actually want to achieve, really?"
To push it one step further:
Please don't treat the current implementation, industry custom or competitor examples as "the answer" for now. First, separate the goal, the observed facts, the unavoidable constraints, the unspoken assumptions and the proxy metrics. Next, consider which assumptions you could drop and still reach the goal. Finally, rebuild several designs from only the first principles that remain, and test them against real-world constraints. If you end up back at the existing plan, explain why.
The key is not to say "come up with something novel."
If you demand novelty directly, the AI will sometimes invent something like a bicycle with three hats stacked on it.
What you need isn't eccentricity. It's the difference that comes out as a result of recalculating from the assumptions.
9. Conclusion: AI can move from "the one who answers" to "the one who breaks the question"
Listing problems is necessary. Competitor research is necessary. Improving existing plans is necessary.
But on their own, they amount to moving pieces faster on the same game board.
If you explicitly assign first-principles thinking to the AI, then
solve the problem → doubt how the problem is framed → doubt the goal → break the assumptions apart → rebuild from zero
all comes within reach of automation.
You also no longer need a human to summon the second loop of double-loop learning through sheer willpower every time.
Just have the AI ask, on a regular basis:
"So, what's this all for, anyway?"
Even a piece of the "philosophy" we thought would be the last thing left in AI automation can, it turns out, be made into a process.
But if the philosopher AI starts saying "I went ahead and decided all the goals too!", that's the one place where a human pulls up a chair and sits back down.
You can leave the breaking of questions to it.
Choosing which world to value is still our job.
