ExpertsFMTobi Lütke: The Skills AI Makes More Valuable
The Knowledge Project · September 15, 2026
5 Selected Lessons:
framework · 2 heuristic · 2 priority · 1
Taste and judgment as the skills that compound most in an AI-saturated world
Source
The Knowledge Project — Tobi Lütke: The Skills AI Makes More Valuable (September 15, 2026) · at 31:10 ▸
When this applies
when deciding what capability to invest in developing as automation absorbs more execution-level work in any field
The takeaway
Invest in skills that sit upstream of automatable execution — specifically, the capacity to select among many valid paths — because that capacity becomes relatively more valuable as execution costs fall.
Why it works
As AI commoditizes execution (coding, research, drafting), the scarce differentiator becomes the ability to choose well among many valid options — taste and judgment — which only comes from deep accumulated reps and pattern recognition.
One level up
When a layer of a value chain gets commoditized, value concentrates in the adjacent layer that chooses/curates/directs the commoditized layer.
Where it stops holding
This stops working when the domain has an objectively verifiable 'best' answer where taste doesn't matter, or when automation itself subsumes the selection step too.
Detail
hierarchy:
- taste and judgment (upstream, scarce, compounds with reps)
- systems design and institutional understanding (how durable structures work)
- truth-seeking / resisting complex-but-wrong narratives
- execution skills (increasingly automated, less differentiating)
rationale: People with great taste have done enormous reps; study how systems/institutions endured (e.g., the Catholic Church's management layers, double-entry accounting) to find hidden harmonies that generalize.
In their words
Just taste and judgment are the skills that have always been valuable but now will get to the limit. I think it's better to spend your teenage years now cultivating and like understanding taste.
Human-in-the-loop decision council via sub-agents
Source
The Knowledge Project — Tobi Lütke: The Skills AI Makes More Valuable (September 15, 2026) · at 10:53 ▸ · at 13:02 ▸
When this applies
when a leader needs to make a high-stakes decision under genuine uncertainty and wants diverse expert perspectives fast
The takeaway
Use automation to construct the best possible information environment for a decision, but keep the actual judgment and responsibility with a human, because machines cannot be held accountable.
Why it works
Running multiple independent model instances as differently-briefed 'experts,' then randomizing and synthesizing their outputs, approximates convening a panel of human specialists while removing single-model bias and compressing turnaround from weeks to minutes.
One level up
Separate the acquisition of information (automatable) from the ownership of consequence (not automatable) when designing any decision process.
Where it stops holding
This stops working when the decision requires real accountability or legal/moral responsibility that only a human can bear, or when the underlying models share correlated blind spots.
Detail
procedure:
- Identify the important ambiguous decision needing a 'neutral ground truth'
- Spin up a small council of sub-agents each tasked with a different expert role (data, research, business, engineering)
- Run each role against multiple different underlying models
- Randomize who synthesizes the results to reduce bias
- Have the single best current model read and synthesize everything into a final answer
- Deliver the result as an audio summary for consumption during routine time (e.g., the gym)
benchmarks: {"cost":"$15-20 in tokens","time":"about half an hour vs. a month manually"}
rationale: Machines can't take responsibility but can dramatically improve the information available before a human makes the call.
In their words
Here's the thing that LLMs and machines cannot do: machines can't take responsibility. And I think this is actually probably the most overlooked thing in the entire stack. Humans take responsibility. Machines can help us take more responsibility because they can inform us better.
Beauty/aesthetics as the communication channel of intuition
Source
The Knowledge Project — Tobi Lütke: The Skills AI Makes More Valuable (September 15, 2026) · at 54:03 ▸
When this applies
when evaluating why an expert's rapid intuitive judgment (a chess move, a design choice, a code structure) feels right before it can be explained
The takeaway
Investing in making work legible and beautiful (even in parts no one else sees) is not decoration — it's how practitioners build and transmit intuition, both to themselves over time and to collaborators.
Why it works
Most of the brain's energy budget runs the visual cortex, which communicates concepts to the rest of the brain via aesthetic signals rather than sequential reasoning; after many reps, expertise gets compressed into this aesthetic channel, which is why experts report 'it seemed beautiful' rather than a calculated chain of logic.
One level up
Aesthetic judgment is a compressed, parallel-processed form of expertise that often can't be fully reconstructed via first-principles reasoning.
Where it stops holding
This stops working when there has not been enough deliberate repetition in a domain for aesthetic pattern-recognition to have formed — novices' aesthetic reactions aren't reliable signals yet.
Detail
procedure:
- Treat code/product architecture as craft, caring about parts others don't see
- Write for legibility so the system 'tells a story' to collaborators and to future-you
- Trust aesthetic reactions ('this seems beautiful/off') as compressed signals from accumulated reps, especially when you can't yet articulate why
rationale: Chess/go experts often can't explain why a move is right beyond 'it seemed beautiful,' reflecting the brain's parallel aesthetic processing versus slower sequential first-principles reasoning.
In their words
They will say, no, I did. I only looked at that one line. The reason why I looked at it is because it seemed beautiful to me in a moment.
Slop grenades: AI lowers the cost of producing low-quality output, shifting the burden downstream
Source
The Knowledge Project — Tobi Lütke: The Skills AI Makes More Valuable (September 15, 2026) · at 16:05 ▸
When this applies
when any cheap-to-generate automated output (code, text, reports) is handed off unreviewed to someone else to deal with
The takeaway
Lowering the cost of production without lowering the cost of verification just relocates labor rather than eliminating it — watch for who absorbs the newly displaced cost.
Why it works
AI collapses the cost of generating volume but not the cost of verifying quality, so the review/compression burden simply moves to whoever receives the output.
One level up
Any automation that reduces effort at one stage of a pipeline can create a bottleneck and new hidden cost at the next stage.
Where it stops holding
This stops working as a warning when the sender actually reviews and curates AI output before sending, restoring the verification step at the source.
Detail
rule: If you're already using an LLM to produce something, use it to compress/clarify your point rather than inflate it, and don't offload unreviewed AI output onto colleagues.
threshold: The failure mode is 'over-output,' not lack of output — if you didn't read what the AI produced before sending it, you've created a slop grenade.
rationale: The recipient now has to do the verification work the sender skipped, often requiring another LLM pass just to decompress what the first LLM inflated.
In their words
It's really easy, especially with stuff like Revo agents. You need a change of some kind, you just tell the AI to go nuts. It makes a pull request. You just say, yeah, that's good. You don't really read it. And now it has to be reviewed by your colleagues.
Intuition is most valuable precisely where rapid feedback is absent
Source
The Knowledge Project — Tobi Lütke: The Skills AI Makes More Valuable (September 15, 2026) · at 35:41 ▸
When this applies
when facing a decision with multiple good options, some with immediate measurable feedback and some without
The takeaway
A high correlation exists between the right long-term path and the absence of a rapid feedback loop, because feedback-rich options are over-selected by incentive structures, not necessarily by merit.
Why it works
Options with fast feedback (e.g., quarterly stock price, short-term revenue) get chosen by default because they're easy to justify and verify quickly, creating systematic bias toward short-termism even when a feedback-less option is actually better long-term.
One level up
Be suspicious of any decision criterion that happens to align with what's easiest to measure, because measurability itself is not a signal of correctness.
Where it stops holding
This stops working when the feedback-rich option is also genuinely strategically superior, or when the decision-maker has no mechanism to eventually validate the slow path at all (never getting a review).
Detail
rule: When comparing good options, prefer deeper investigation of the ones without an obvious short-term feedback signal, since those are systematically under-weighted by incentive-driven decision-makers.
threshold: Applies to decisions where multiple paths are already 'good' (not when choosing between clearly right vs wrong).
rationale: Short-termism arises because executives are locally rational actors responding to quarterly incentives, not because short-feedback options are actually best — 'show me the incentives and I'll show you the outcome.'
In their words
I find a very high correlation between the right path and the ones that don't have feedback loops attached.