Dylan, founder of Assembly AI, says he writes a private document of every expected investor objection with lengthy answers to stay sharp on competitive and market questions.
Dylan says a YC partner argued Assembly AI was competing with a Google product manager navigating internal politics rather than with Google itself.
Dylan says the first Series A-era raise was a mistake: his first post-YC meeting with Sequoia was probably the wrong strategy, since he had no traction or experience.
Assembly AI's founder built speech AI where demand already existed
Assembly AI bet early that deep learning speech models would keep improving with scale, and it offered them through a developer API. Demand was clear before the product existed, so the work was getting the models good enough, with each training run taking one to two weeks. Dylan argues a small focused team can do that against Google.
Takeaways
A small team can compete with Google. Dylan says the best product is not just a function of resources, and more people can slow a project, as The Mythical Man-Month describes. A YC partner told him Assembly AI was competing with a product manager at Google, not Google.
Demand existed before the product did. Dylan says Assembly AI knew market demand existed before it started building, unlike many companies. Adoption matured as the technology improved, and the 2022 generative AI wave accelerated it beyond expectations. Fit now means improving the product for what most customers build with its API.
Deep learning scale justified the original bet. Classical machine learning needed task-specific models and handcrafted features, while deep learning models would keep growing on more data with no ceiling in sight. Dylan's first speech model used 10,000 hours of audio, and the next release uses 12.5m hours on hundreds of TPUs.
Slow model training limited early iteration. During YC each model iteration took one to two weeks to train, which slowed progress against the 10% weekly growth YC expects. Dylan says the company could have moved faster by setting accuracy targets before taking its product to market.
Fundraising is a numbers game of believers. Dylan says raising capital means finding people who believe. He keeps a private document of expected investor objections with long answers. His first meeting after YC, with Sequoia, was probably the wrong strategy without traction.
Take the long view after YC. Asked what he would tell his earlier self, Dylan says to take a longer-term view in the first year after YC, set quarterly goals and avoid chasing quick progress.
From Cisco engineer to Series C
Around 2016 Dylan works as a machine learning engineer at Cisco on natural language processing, gets bored, and quits to start Assembly AI alone.
He applies to Y Combinator's summer 2017 batch 30 days after the deadline as a single founder and gets in. YC invests $125,000, and partner Daniel Gross becomes an early believer.
After YC the company raises about $1m from angels only and stays at three people for a long time.
Two years of building end with the first customers in late 2019. One found the company on Hacker News and monitored TV and radio for brand mentions.
Tap a name to go there · tap a line to see why they are linked
Show 1
On this episode 2
In 2020 Assembly AI raises $5m from angels including Daniel Gross and Nat Friedman, with over $1m in annual recurring revenue and fewer than 10 people.
The Series A follows about a year and a half later, then a Series B and a Series C.