The Twenty Minute VCBret Taylor: Why Pre-Training is for Morons & Companies Will Build Their Own Software | E1209
At a glance
WHAT IT’S REALLY ABOUT
Bret Taylor Explains AI Bubbles, Agents, And Why Models Commoditize
- Bret Taylor discusses the current AI boom, arguing it is a bubble that “rhymes” with the dot‑com era—wildly excessive in places yet still producing multiple trillion‑dollar and enduring enterprise companies.
- He is deeply skeptical of most startups doing their own model pre‑training, believing value will lie in applications and solutions, not in yet another frontier model, and that economics will favor companies whose costs are tied to inference, not massive training runs.
- Taylor predicts AI will mirror the cloud market, with a few hyperscalers and research labs owning the heavy infrastructure and models, while most companies buy specialized, SaaS‑like AI applications and customer‑facing agents rather than building everything themselves.
- He also describes his new company, Sierra, which builds branded conversational agents for consumer companies, and outlines how conversational, multimodal interfaces will become a primary way consumers interact with businesses, raising new design, governance, and change‑management challenges.
IDEAS WORTH REMEMBERING
5 ideasTreat AI like the cloud stack: few model providers, many solution vendors.
Taylor expects AI to stratify into infrastructure (hyperscalers and labs training frontier models), tools, and SaaS‑like applications, just as cloud did—most economic value will accrue to solution‑oriented products that solve concrete business problems.
Avoid pre‑training your own large models unless you’re a true AGI lab.
For almost all startups, spending heavily on pre‑training is like building your own data center in 2024: capital‑destructive and off‑mission compared with fine‑tuning strong existing models (proprietary or open source) to reach product–market fit.
Anchor your AI costs to inference usage, not massive upfront training bets.
Taylor argues sustainable AI businesses keep training costs modest and variable costs tied to inference, which in turn correlates with real customer usage and revenue; this sharply reduces the risk that sunk training costs never get paid back.
Expect foundation models to commoditize quickly while frontier models leapfrog.
Open models like LLaMA already provide GPT‑3.5‑class performance, making much of the “good enough” layer commodity; at the top end, performance will improve in step changes, but the bar to maintain a clear frontier lead keeps rising.
Every company will need a branded conversational agent, not just a website.
Taylor believes we’re entering an era where consumers primarily interact with companies through conversational, multimodal agents (text, voice, images) that handle support, sales, and transactions—Sierra is aimed at powering that shift for brands.
WORDS WORTH SAVING
5 quotesI am inherently skeptical of companies doing pre‑training. Unless you are an AGI research lab, doing pre‑training on a model, I believe, is just burning capital.
— Bret Taylor
Software is like a lawn, it needs to be tended to. It's not like you write software once and it just works forever.
— Bret Taylor
The AI bubble will rhyme with the dot‑com bubble… We will look back and laugh at some of the excess, but I’m confident we will have a brand‑defining, likely trillion‑dollar consumer company come out of this.
— Bret Taylor
We're going from the age of rules to the age of goals and guardrails.
— Bret Taylor
In 1995, the way you existed digitally as a business was to have a website. In 2025, the way you will exist digitally is to have an AI agent.
— Bret Taylor
High quality AI-generated summary created from speaker-labeled transcript.
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome