ExpertsFMHow to Live in Everyone Else's Future (with Shopify CEO Tobi Lütke)
ACQ2 by Acquired · September 18, 2025
5 Selected Lessons:
framework · 3 reasoning_pattern · 1 heuristic · 1
Personal eval folder for benchmarking AI capabilities
Source
ACQ2 by Acquired — How to Live in Everyone Else's Future (with Shopify CEO Tobi Lütke) (September 18, 2025) · at 0:00 ▸ · at 2:30 ▸ · at 44:57 ▸
When this applies
when evaluating any new tool, vendor, or technology for capability before adopting it
The takeaway
Build a standing, repeatable test harness for anything you must periodically re-evaluate so progress is measured objectively, not impressionistically.
Why it works
A fixed set of prompts with expected/judged results lets you consistently retest any new model against the same edges, revealing real capability gains rather than vibes.
One level up
Treat evaluation as infrastructure you maintain continuously, not a one-off judgment you make each time.
Where it stops holding
This stops working when the underlying capability space changes so fast that your fixed test set becomes irrelevant or saturated, requiring constant curation.
Detail
procedure:
- Maintain a folder of prompts with expected/judged results
- Run it against every new model release
- Judge results to detect capability jumps
- Add new failure cases as new evals when a model fails at something
benchmarks: {}
rationale: Enables fast, consistent comparison across model generations instead of ad hoc impressions.
In their words
It's it's like literally a folder of prompts with expected and judged results.
Constitutions — codifying unique, non-platitude choices for AI and human guidance
Source
ACQ2 by Acquired — How to Live in Everyone Else's Future (with Shopify CEO Tobi Lütke) (September 18, 2025) · at 48:13 ▸ · at 1:19:41 ▸
When this applies
when trying to codify an organization's or individual's 'secret sauce' so it can be taught, delegated, or automated
The takeaway
To make tacit expertise transferable or automatable, write down your actual contested trade-offs, not your universally agreed values.
Why it works
Writing down only the choices where a reasonable other company/person would plausibly choose differently (the real trade-offs, not platitudes) creates a document with enough specificity that both AI and humans can use it to drive genuine debate and decisions.
One level up
Legibility of judgment (not just rules) is what allows delegation — to people or machines — at scale.
Where it stops holding
This stops working when an organization's values truly are platitudes (no real contested trade-offs exist), in which case the document adds no decision-making value.
Detail
procedure:
- Identify choices where another company/person would plausibly take the opposite side — these are the real trade-offs
- Discard platitudes everyone agrees on (e.g. 'teamwork is good')
- Feed the document to AI and ask what's missing; iterate (hill-climb) on it
- Use the document both as context for AI-run reviews and as a conversation-starter for human disagreements
- Treat detected 'contradictions' as prompts to refine precision or add provisos, not just errors
rationale: Any project in review can be run against the product principles — it's not a control mechanism, it's where interesting discussion should happen.
In their words
What makes, you know, basically, like write down your uniqueness. And then you get in an absolutely fascinating world because first of all, you can use it.
Founder-run vs. managerial companies: legitimacy and pivoting
Source
ACQ2 by Acquired — How to Live in Everyone Else's Future (with Shopify CEO Tobi Lütke) (September 18, 2025) · at 1:19:41 ▸ · at 1:23:30 ▸ · at 1:26:12 ▸ · at 1:29:03 ▸ · at 1:31:53 ▸
When this applies
when deciding how much decision-making authority to centralize in an individual versus a formal plan/process
The takeaway
Centralizing legitimacy in a conscientious, highly-adaptive individual (rather than a formal consensus plan) enables faster pivots, at the cost of higher variance outcomes.
Why it works
In managerial companies, legitimacy resides in 'the plan,' which requires enormous activation energy to change; in founder-led companies, legitimacy resides in a trusted individual, who can pivot on a dime if they remain open to new data.
One level up
Speed of adaptation in any system is bounded by where legitimacy/authority formally resides, not by the talent of the people involved.
Where it stops holding
This stops working when the individual holding legitimacy is not conscientious or open to changing their mind with new data — then concentrated legitimacy produces bad outcomes without offsetting upside.
Detail
procedure:
- Identify whether legitimacy in your organization rests in an individual or a plan/process
- If in an individual, ensure they are conscientious and will change their mind with better data
- Accept higher variance (more frequent failure) in exchange for higher ceiling outcomes
- Recognize managerial/plan-legitimacy companies suit domains needing predictable operational effectiveness (e.g., Fredrick Taylor-style manufacturing)
rationale: If legitimacy is invested in an individual, you can pivot on a dime... it can operate at a quite a good batting average, which makes the highlights and the accomplishments will end up making up easily for the downsides.
In their words
If a legitimacy is invested in an individual, then you can pivot on a dime. And the people who don't like this don't work on this
Noticing the goalpost-moving blindness to progress
Source
ACQ2 by Acquired — How to Live in Everyone Else's Future (with Shopify CEO Tobi Lütke) (September 18, 2025) · at 19:18 ▸
When this applies
when trying to judge whether a major threshold has already been crossed in your industry or life, despite no one celebrating it
The takeaway
Don't rely on social/cultural fanfare to tell you when a threshold has been crossed; check the actual criteria against reality directly.
Why it works
Humans adapt to new baselines almost instantly (a hedonic/expectation treadmill), so major milestones often pass without recognition because attention has already shifted to the next goalpost.
One level up
Collective attention is a poor instrument for measuring absolute progress because it is relative and constantly rebasing.
Where it stops holding
This stops working when the milestone truly is ambiguous or contested on its merits, not just under-recognized.
Detail
approach: Check milestones against their original definition, not against current ambient conversation.
signals:
- No celebration or media coverage despite a bar being cleared
- Attention has moved to a newer, undefined target (e.g. AGI) instead
rationale: Humans are incredibly good at moving the goalpost and adjusting to circumstances quickly, obscuring real progress from themselves.
In their words
it moves forward. It's like we adjust we adjust to our circumstances incredibly quickly.
Consensus and leadership are opposite ends of one spectrum
Source
ACQ2 by Acquired — How to Live in Everyone Else's Future (with Shopify CEO Tobi Lütke) (September 18, 2025) · at 1:31:53 ▸ · at 1:35:36 ▸ · at 1:37:40 ▸
When this applies
when any team or organization must decide whether to resolve a question by committee agreement or by a single accountable decision-maker
The takeaway
If you want outlier (top-tier) outcomes, you must accept decision-making structures that permit failure, because consensus structurally caps upside even as it raises the floor.
Why it works
Consensus-seeking optimizes for lack of downside (acceptability to everyone), which statistically clusters outcomes in the 6-7 range on a 1-10 scale; leadership allows aiming for and sometimes hitting 9-10, at the cost of also risking lower outcomes.
One level up
Every mechanism that reduces variance in outcomes also reduces the probability of extreme positive outcomes — there is no free lunch between safety and excellence.
Where it stops holding
This stops working when consistent, predictable, above-average quality (not extreme outlier quality) is actually the goal, e.g., in operational manufacturing contexts.
Detail
rule: Consensus decision-making structurally produces mid-range (6-7/10) outcomes; only leadership/fiat can reach 9-10/10 or fail hard trying.
threshold: On a 1-10 scale, consensus caps around 6-7.
rationale: Extreme outliers come from a single vision executed by a team of people, not an agreed-upon vision executed by a team of people.
In their words
The problem with consensus decision is on that spectrum from one to 10, it's usually a six or a seven