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 working at a desk in an Indian city at dawn, with a world map behind them lit by lines running out to every continent

September 2, 2026 · 15 min read

AI Startups in India: Why This Wave Is Global

The last Indian startup wave produced local champions because its enabling technology was local: a phone in every pocket created density you had to accumulate city by city. A model endpoint answers the same way from Bengaluru as from San Francisco, which removes the constraint that used to force Indian software companies to sell into India. That is the argument the closing panel at Y Combinator's first Startup School in India made, and the mechanism holds up better than the national-talent claim stapled to it. Here is what actually changed, the one constraint that is still genuinely harder in India, and the gate nobody on stage had time to mention.

A founder standing in front of a single glowing app icon on a phone-shaped frame, with a crowd of ordinary people stretching out behind it

September 1, 2026 · 16 min read

Consumer AI: Why Now, and What It Takes

Y Combinator's case is that every platform shift mints consumer giants, and three years into this one the only new icon on most home screens is ChatGPT. The two conditions it names, intelligence good enough and tokens cheap enough, are real. They are also the easy half. The head of ChatGPT says model improvements drove only about a third of that product's growth, the CEO of Palo Alto Networks says token prices are set by strategy rather than by the cost curve, and the creator of OpenClaw says the hardest problem is not the technology at all. Here is the version of the pitch with the operating reality attached.

A founder at the front door of their product checking two arrivals side by side, a human customer and a software agent, and sorting each into its own lane

August 31, 2026 · 15 min read

Bot Detection When Bots Are Your Customers

Every page ranking for this term sells you a better wall: sharper fingerprinting, smarter challenges, faster blocking. That advice is now half wrong in both directions. Faking a person got cheap enough that a passed check proves very little, and a growing share of the traffic you want is software acting for a real customer, so blocking everything non-human costs you revenue. Y Combinator has called rebuilding the trust layer of the internet one of the most important problems of the next decade. Here is what actually changed at your front door, and the door policy to run instead.

A founder slotting one of several interchangeable model modules into an open coding agent on a workbench

August 30, 2026 · 17 min read

Open Source Coding Agent: OpenCode's Playbook

OpenCode went from roughly 650,000 monthly active users at the start of the year to about 13 million at the end of June, and it did it as an open source alternative to a product Anthropic bundles with a paid subscription. That is not luck. Founder Jay V picked a structural position (the open slot beneath two dominant closed players), got blocked by one of them and grew because of it, and built an economic model where tokens are the customer acquisition cost. Here is what an open source coding agent actually is, what drove the growth, and the five tests for whether the same play is available to you.

A founder choosing between a general purpose agent harness and an agent built deep into one industry's workflow

August 29, 2026 · 14 min read

Vertical AI Agents: What the Model Can't Eat

Google DeepMind's Logan Kilpatrick says the race to build agent harnesses has roughly a 12 month shelf life: the models will absorb the scaffolding and run it natively. The creator of Claude Code says the same thing from the other side of the industry. That leaves one question worth your roadmap: what does the model not eat? Here is the case for vertical AI agents, the four assets that survive a model release, and a three question test for whether yours is actually defensible.

A founder choosing between a chip built for a single response and one built for an agent loop

August 28, 2026 · 14 min read

Inference Chips Are Built for the Wrong Job

Almost every inference chip on the market assumes one request produces one answer. Agents do not work that way: they loop, call tools, branch, backtrack, and hold context across dozens of steps. Y Combinator says current GPUs land at 30 to 40 percent of peak utilization on that kind of work. Here is what the gap actually is, what the Nvidia and Groq deal really was, and what to change in your own stack when you are not the one building silicon.