ExpertsFMBuilding Web Apps with Just English and AI (with Vercel CEO Guillermo Rauch)
ACQ2 by Acquired · February 18, 2025
5 Selected Lessons:
case · 1 heuristic · 1 reasoning_pattern · 1 framework · 1 priority · 1
Deep technical CEO involvement over delegation-by-default
Source
ACQ2 by Acquired — Building Web Apps with Just English and AI (with Vercel CEO Guillermo Rauch) (February 18, 2025) · at 48:36 ▸
When this applies
when a founder-CEO is deciding how much to delegate away versus stay personally close to the operational details as the company scales
The takeaway
Prioritize direct understanding over outsourced reporting: a leader's value is in giving high-quality feedback grounded in firsthand reality, not in offloading comprehension to subordinates.
Why it works
Staying grounded in the product and customer reality (using your own product, talking directly to customers) lets a CEO give high-quality feedback to executives without needing time-consuming briefings, preserving speed and accuracy of decisions even at scale.
One level up
Scale should be achieved by compressing the distance between leadership and ground truth, not by inserting more layers of abstraction between them.
Where it stops holding
This stops working when the company grows so large or diversified that no single person can maintain firsthand grounding across all areas, risking bottlenecking decisions on the CEO or undermining executive autonomy.
Detail
hierarchy:
- Understand the business/product deeply yourself, in every area
- Use that grounded understanding to give high-quality feedback to executives
- Hold executives accountable to numbers/outcomes while still being the 'bridge to reality'
- Delegate execution, not understanding
rationale: He explicitly says 'you can't just say, this area of infrastructure, I have no idea how it works, talk to Johnny' — understanding, not outsourcing labor, is the principle.
In their words
I think CEOs, especially technical CEOs, can and should be involved in every aspect of the business, whether it's closing a deal, whether it's looking at the numbers of the last quarter and tuning sales quotas and sales compensation, whether it's informing future product roadmap, whether it's pricing discussions.
Embedding company 'taste' into an AI product via evals, benchmarks, prompts, and RAG
Source
ACQ2 by Acquired — Building Web Apps with Just English and AI (with Vercel CEO Guillermo Rauch) (February 18, 2025) · at 26:18 ▸
When this applies
when you need to scale a scarce internal expertise or 'taste' across a huge audience without linear headcount growth
The takeaway
To scale expertise that doesn't scale through direct human service, convert your top practitioners' judgment into structured training/evaluation artifacts that an AI system can apply broadly.
Why it works
By encoding the demonstration and preference data of your best internal experts into evals, benchmarks, prompts, and retrieval systems, an AI product can reproduce expert-level judgment at near-zero marginal cost per user.
One level up
Institutional expertise becomes a scalable asset only once it's externalized into reusable, machine-usable form rather than remaining tacit knowledge in individuals' heads.
Where it stops holding
This stops working when the expertise is too tacit/contextual to be captured in evals and prompts, or when the AI model lacks the capability to act on that encoded guidance reliably.
Detail
procedure:
- Identify the scarce internal experts whose judgment customers constantly request (support, training, audits, migrations)
- Capture their preference and demonstration data on what 'good' looks like
- Build evals and benchmarks to measure AI output against that standard
- Use prompts and RAG to steer foundation models toward reproducing that taste
- Deploy as a self-serve product instead of 1:1 consulting
benchmarks: {"adoption":"used heavily for legacy-to-Next.js migrations as an emergent use case"}
rationale: Rauch explicitly could not scale himself to every company asking for support/training, so the team encoded their collective expertise into vZero instead.
In their words
We are imbuing the preference data, the demonstration data of what they think is the best way to build software. So it's our answer to how you actually make support, advice, audits, migrations at scale sustainable and efficient.
Development as a dissolving 'role' becoming a ubiquitous 'skill'
Source
ACQ2 by Acquired — Building Web Apps with Just English and AI (with Vercel CEO Guillermo Rauch) (February 18, 2025) · at 21:40 ▸
When this applies
when assessing how AI will reshape any specialist profession historically gated by scarce human expertise
The takeaway
Identify which part of a profession is 'translation' (mechanically converting one form to another) versus 'taste/direction' (deciding what should exist); expect AI to dominate the former while the latter stays human.
Why it works
AI excels at tasks that are essentially translation — converting one representation of intent into another (language to language, intent to code) — so professions built around that translation function get absorbed into AI tools, while human taste and problem framing remain valuable.
One level up
Durable human value in an AI-disrupted field concentrates in judgment and intent-setting, not execution mechanics.
Where it stops holding
This stops working when the professional task requires physical-world grounding, novel judgment calls, or trust/accountability structures that can't be delegated to an AI system.
Detail
approach: Use the translator-profession analogy to forecast which parts of a skilled role get commoditized by AI versus which parts (taste, creativity, direction) remain human-differentiated.
signals:
- AI already outperforms trained professionals on mechanical execution
- the remaining differentiator is creative direction and problem selection, not technical execution
rationale: Google Translate example: transformer architecture arose from translation and now that capability no longer requires human specialization; Rauch expects software development to follow the same path.
In their words
My take is that development will go from being a role, like translator was a role, to a skill that AIs will increasingly dominate.
Start infrastructure-first, then narrow to a focused niche
Source
ACQ2 by Acquired — Building Web Apps with Just English and AI (with Vercel CEO Guillermo Rauch) (February 18, 2025) · at 11:18 ▸ · at 33:45 ▸
When this applies
when your product tries to serve everyone and traction isn't materializing
The takeaway
If initial generality isn't working, narrow to the one niche where you have personal expertise and conviction; let customer demand — not your own ambition — drive expansion afterward.
Why it works
A platform too general dilutes differentiation and fails to build a deep, defensible expertise anyone values; narrowing to one specialty lets the company build true excellence that attracts a committed user base, which can later pull the company into adjacent categories.
One level up
Dominate one category completely before expanding scope; let the market pull you into adjacencies rather than pushing broadly from day one.
Where it stops holding
This stops working when the niche you pick is too small to sustain a company, or when customers need the broader platform immediately and won't wait for you to expand.
Detail
rule: Be excellent at one thing first ('dominate a niche'); only expand once customers are actively pulling more product out of you.
threshold: Vercel waited roughly 6-7 years tightly coupled to Next.js/front-end before expanding into back-end/database/AI token processing services, in response to explicit repeated customer requests.
rationale: The mistake of trying to deploy 'everything under the sun' early on led to unfocused traction; narrowing to front-end deployment created a defensible leadership position that could later be expanded.
In their words
My number one priority is to be excellent at one thing. I want to be world class at one thing. People call it dominating a niche.
Betting on dynamic/cloud rendering against the static framework trend
Source
ACQ2 by Acquired — Building Web Apps with Just English and AI (with Vercel CEO Guillermo Rauch) (February 18, 2025) · at 8:28 ▸ · at 11:18 ▸
When this applies
when the market/community consensus trend diverges from how the proven leaders in the space actually win
The takeaway
When you notice the crowd and investor capital chasing an easier-to-build approach that contradicts how the actual category winners operate, consider betting on the harder path that mirrors the winners — even if it takes years to pay off.
Why it works
Static, device-rendered approaches were easier to build and fund short-term, but didn't match how dynamic, highly personalized leaders (Google, Facebook) actually created engagement and retention; betting on the harder, more aligned-with-winners approach paid off as mobile grew and offloading compute to the cloud became more valuable.
One level up
Durable advantage often comes from a long, hard-to-copy bet rather than riding the popular, easy-to-fund trend.
Where it stops holding
This stops working when the 'crowd' trend is actually correct and the incumbents' model doesn't generalize to your market, or when your harder bet takes too long to pay off before capital runs out.
Detail
situation: The React community and investors were chasing static-site frameworks; Rauch found this puzzling because the most successful consumer products (Gmail, Facebook News Feed) were highly dynamic and cloud-driven.
decision: He quietly built Next.js focused on dynamic rendering in the cloud rather than static, device-side rendering, despite it being technically much harder to scale securely and multi-tenant.
outcome: Next.js became the dominant framework (1.3M monthly active developers) and the 'market spoke' in favor of dynamic, end-to-end frameworks over static, low-level libraries.
lesson: Betting against a popular but misaligned trend, in line with how proven winners actually operate, can create a long-term moat even though it's harder to execute.
In their words
Why were people building static frameworks for React? It puzzled me because investors were chasing a lot of those frameworks. I was like, Why are you funding that? And I quietly kept building this Next. Js thing that was focused on dynamic.