Blog

Practical AI guides for founders

Short, useful reads on growing your startup with AI, safely. From the team behind AI Operating System for Startups. See also AI startups from Y Combinator.

A founder at a whiteboard answering a column of hard questions about an idea, with the keyboard pushed aside and untouched

October 20, 2026 · 15 min read

How to Validate a Startup Idea Before You Build

A working prototype used to be evidence. Now a coding agent will build a convincing version of any idea you describe, including the bad ones, which means "we shipped it" has stopped filtering anything. The only expensive step left is deciding whether the thing deserves to exist. Garry Tan, president and CEO of Y Combinator, built that decision into a Claude Code skill: six forcing questions, modeled on YC partner office hours, that run before the agent writes a line. Inside: what each question is actually testing and the red flag that means you have not answered it, which questions matter at your stage, the gate working on a real idea where it separated pain from consequence and reframed the idea into an expansion Tan guessed might be worth ten times more, why a gate that never kills anything is not a gate, and the one input no agent can supply for you.

A founder at a planning board tracing the path from a chip, to the memory stack beside it, to the delivery schedule that decides when compute arrives

October 19, 2026 · 12 min read

AI Export Controls Decide the Compute You Can Buy

Export controls read like someone else's problem until you notice they are upstream of your inference bill. The rules decide which chips and which memory exist in which markets, which decides what your cloud can buy, which decides what you pay per token and how long you wait for capacity. Aravind Srinivas, co-founder and CEO of Perplexity, gave 20VC the least comfortable read available on them: short term they worked, long term they may have built the competitor they were meant to slow down. Inside: what the rules actually restrict, including the high bandwidth memory rule most coverage skips; why a restricted stack is a different stack rather than a slower one; why power and memory, not GPUs, are the constraints you will feel first; where chip policy quietly becomes a security and compliance question about where your tokens go; and the planning calls to make now rather than when the rules change again.

A founder at a console directing a live generated scene as it plays, steering it mid-stream instead of waiting for a render to finish

October 18, 2026 · 21 min read

Real-Time AI Video Changes What You Can Build

Generative video crossed a threshold: it now renders faster than a person can watch it, which turns a render you wait for into a session you can interrupt. fal post-trained MiniMax's open-weights H3 model into H3 Max, cut a five-second 768p clip from 53.20 seconds to 2.78, and three days later an indie maker shipped a product on it that 37,000 people watched within a day. The founder lessons are not about video. Inside: token market fit, the demand test that asks whether one person can productively spend 10k a month; why growth now tracks the compute you can get rather than the demand you can create; the four stacked optimizations that produced an order of magnitude; why the post-training pipeline is the asset and the model is the result; what studios actually buy, which is small point solutions rather than magic; and why half the moat in a regulated category is legal rather than technical.

A founder laying a bridge span across a gap between a small glowing machine and a wall of old filing cabinets behind locked doors

October 17, 2026 · 19 min read

Enterprise AI Adoption Is a Workflow Problem

Coding agents went from novelty to default in about a year and everything else in the enterprise is still crawling, which is the most useful fact about enterprise AI adoption right now because the reason is not model quality. Aaron Levie, co-founder and CEO of Box, argues the gap between what a model can do and what a workflow actually needs is vast, that closing it is a lot of software rather than a lot of research, and that the slowness is exactly where the application layer's value comes from. Inside: the likeness test that predicts which workflows automate first, why there is no give-us-your-GitHub for knowledge work, the accuracy threshold that decides whether a process gets automated at all, the permission wall that kills pilots nobody demos, and the two-sided mandate for anyone running a system of record.

A founder at a workbench choosing one tool from an ordered rack of six, with a small architectural model on the bench beside a much larger one

October 16, 2026 · 18 min read

LLM Post-Training: When to Own Your Model

LLM post-training is everything you do to a model after base training so it behaves the way your product needs, and the technical explainers all stop before the only question a founder has: is this mine to do, in what order, and what does the wrong order cost. Lin Qiao, co-founder and CEO of Fireworks AI and previously the leader of PyTorch at Meta, lays out the ladder she sees working (prompt, retrieval, supervised fine-tuning, preference tuning, reinforcement learning), the symptom that justifies each rung, when distillation takes over, why your product team and not your ML team should judge training data, and the single gate that decides timing: product-market fit comes first, because before it the data you would train on encodes guesses. Plus the two failure modes that quietly burn the budget, and what owning weights does to your security posture.

A founder at a single dashboard weighing paid channels while stacks of ad creative variations feed into one funnel

October 15, 2026 · 19 min read

Customer Acquisition Strategy for AI Startups

A customer acquisition strategy is the written answer to three questions: which channels you buy attention on, which single event you pay for, and how you will know the money worked. Matt Swulinski has built that engine three times, at Superhuman, at Wispr Flow, and now at viktor.com, and his version is deliberately narrow: Meta, Google and lifecycle carry you to the first one to ten million in ARR, measurement comes before the first dollar of spend, and a hundred thousand dollar Meta budget needs 400 to 500 new creatives a month to avoid plateauing. Here is the full playbook, with the parts worth arguing about marked as such, plus the fully loaded CAC math most AI startups are getting wrong.