Feb 8, 2024 · 32m · y-combinator
The Truth About Building AI Startups Today · Y Combinator
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In the inaugural episode of The Light Cone podcast, Y Combinator group partners analyze the current AI startup landscape, sharing strategic advice on navigating market opportunities, technical architectures, and common execution pitfalls. They highlight why unglamorous workflow automation and specialized technical integration offer durable venture opportunities over superficial AI wrappers.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The partners hold 75.2% of the talking time here. How this is scored →
The partners as informed peer 7.0 Guest teaching 1.4 Guest disagreement 1.5 The partners pushing back 1.1
speaking balance: gold is the partners, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 21:08 Dismissing the GPT wrapper narrative
Gary Tan aggressively rejects the widespread industry cynicism about wrappers by equating all SaaS to mere MySQL wrappers.
Hardest push from the partners ▶ 9:20 Pushing back against chat as the primary interfaceGary rejects the consensus that conversational chat interfaces are the optimal way to interact with LLMs, arguing they burden the user.
Biggest teaching moment ▶ 27:20 Tracking the legacy of the Attention Is All You Need paperDiana educates the room with historical metrics from NeurIPS, noting seven of the eight transformer paper authors founded companies worth over $6B.
The partners hold their own ▶ 18:53 Hardware architecture analogy for LLM developmentDiana demonstrates deep technical mastery by comparing GPT-4 prototyping and small local models to FPGA prototyping and custom ASICs.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The partners as informed peer | Guest teaching | Guest disagreement | The partners pushing back | Why |
|---|---|---|---|---|---|---|
| Host Introductions and Podcast Title Origin | 6 | 1 | 1 | 1 | Introductory segment where YC partners discuss the origins of the podcast name and statistics on AI batch applications. The hosts display domain knowledge regarding their admissions process and technology history in a friendly, conversational exchange. | |
| Young Founders and Opportunity Leveling in AI | 6 | 1 | 1 | 1 | The panel explores how young founders and college dropouts are uniquely positioned for LLM startups because nobody has years of experience. Harj and Jared share concrete YC examples like automating government contract bidding. | |
| 'Where There's Muck, There's Brass' | 7 | 2 | 2 | 2 | The group analyzes AI tar pits like generic copilots and chat interfaces. Gary Tan critiques chat interfaces as low-leverage UX, while Diana and Jared elaborate on retention issues and selling shovels in a gold rush. | |
| Fine-Tuning Open Source Models vs. OpenAI | 7 | 2 | 2 | 1 | Harj raises the viability of offering open-source fine-tuning as a business. Diana points out that price competition fails against OpenAI's dropping costs, and Gary brings up cybersecurity risks with prompt injection. | |
| Purpose-Trained Small Models and Prototyping Workflows | 8 | 1 | 1 | 1 | Diana provides an in-depth technical analogy comparing foundation models to expensive FPGAs and purpose-trained smaller models to custom ASICs/SoCs. The panel aligns on running local inference for specific tasks. | |
| Debunking the 'GPT Wrapper' Critique | 7 | 1 | 2 | 1 | Gary and Jared dismantle the dismissive 'GPT wrapper' meme by comparing it to calling all SaaS products 'MySQL database wrappers'. They highlight how UI/UX craft and workflow integration generate enduring value. | |
| Differentiating Billion-Dollar AI Ideas from GPT Steamrollers | 7 | 1 | 1 | 1 | Jared questions how to separate billion-dollar AI ideas from products destined to be steamrolled by future foundation model releases. Diana and Harj emphasize embedding deep vertical business logic and workflow orchestration. | |
| AI Voice Agents, Defensive AI, and Open Source Equity | 8 | 2 | 2 | 1 | Gary advocates for open-source AI as insurance against monopolistic AGI tyranny, while Diana shares detailed observations from NeurIPS about transformer paper co-authors creating multi-billion dollar companies. |
Statements from this episode (23)
Friedman: Close to 50% of YC Summer 23 Batch Used LLMs
“I think for summer of 23, it was close to 50% of the batch.”
Taggar: More founders are dropping out of college to start AI companies
“One thing I'm seeing that's interesting is I feel like a lot a lot more founders are dropping out of college to start working on AI.”
Friedman: No one has four years of LLM experience, empowering young founders
“This is an opportunity where college students are particularly well, like young founders are particularly well positioned to work in it because nobody has like there's no one walking around with like four years of LLM experience. So like everyone is starting f…”
Taggar: The AI startups gaining traction focus on mundane workflow automation
“The stuff that I've seen in the batches are actually taking off is a little bit more mundane. Like it's I mean, I probably say a lot of it's sort of like workflow automation, like It's finding things where there was, like, a human doing some repetitive tasks u…”
Friedman: YC receives few applications automating back-office tasks with LLMs
“And yet we actually don't get that many applications for people working on this.”
Taggar: AI Co-Pilot Startups Get Paid Easily but Suffer Low Usage
“There's so much interest from potential customers to like want a co-pilot that it's actually quite easy to start getting like inbound leads if you pitch this.
And if it's even easy to get people to pay you money upfront, but what's really hard is to get them t…”
Tan: Embed LLMs in Familiar UIs Rather Than Chat Interfaces
“While in the next five or 10 years I think we will get far more used to using it that way I think the low-hanging fruit right now is just using the large language model
To actually do the sort of knowledge work that a human being could do, and then package it …”
Taggar: Build Full Competitors if Incumbents Reject AI Co-Pilots
“Another old PG essay about if you, if you're trying to sell technology to someone, and they're not buying, like, see if you can just build a competitor, and so it's like, hey, if you're trying to sell, like fintech company, a copilot, and they're not buying it…”
Tan: AI Dev Tools Suffer Churn as Incumbents Add LLM Features
“A bunch of us are funding dev tools companies that sell to AI companies and they're selling tooling, but then they might, you know, they might sell enterprise contract to someone who also upstream has a fortune 100 that said that they'd pay a 100,000 dollars a…”
Harj Taggar: Competing with OpenAI on cost alone fails to retain customers
“What I think a bunch of the companies in the space are seeing is that, like, that's not enough to keep the customers, especially because, like, OpenAI, like, the cost of all of the models is just going down.”
Diana Hu: Fine-tuning businesses succeed when customizing for private industry datasets
“I think where is exactly that, where I think is having more legs is when these companies need to customize it to private data sets. So you have the open, general, big foundation model, but then you have to tune it up to specific data sets that, for example, a …”
Tan: Fine-tuning models on private data risks leaking it through prompts
“What they actually have figured out is that for a lot of large language models, if you do any sort of fine tuning or training with private data, you can actually just speak to the model and get it to spit out your private data again, and they have a solution t…”
Taggar: LLM Enterprise Data Access and Permissions Is a Ripe Market
“I definitely think that whole world of controlling, within an enterprise in particular, like, controlling who has access to, like, which LLM has access to, like, what data and who has permissions is, like, a really ripe space for building interesting software.”
Hu: Customized open models will beat frontier LLMs in specific domains
“So there's this other world where the open model that's customized, I think, is gonna win and compete versus the big one for specific domains.”
Tan: Any GPT-4 prompt workflow can be duplicated with custom fine-tuning
“Anything you do with those prompts, you can get your own model to do with a little bit more training.”
Friedman: People will look back on 'GPT wrapper' as a silly term
“And I think people are going to look back on this term GPT, GPT wrapper, like similarly how we think of like, how we would look at the term database wrapper, which just seems like silly.”
Tan: Chat interfaces are the wrong UX model for AI software
“I mean, this is why I think the chat interface is wrong. Like I actually think there is value accrued to really great UX, like good copy, good you know, interaction design, information hierarchy you know, being able to approach a product and say like, this is …”
Taggar: Foundation models will outcompete generic automation tools over vertical AI
“If it's like, Hey, like throw your data in here and we'll do like automations on top of it, like for everything, that's probably Hard to compete with whatever one of the foundation models might offer, but if it's like, hey, we're our, give us, like, your sales…”
Taggar: Malicious AI agents will require consumers to deploy defensive AI agents
“He's worried about there's going to be a world of just sort of like all these AI agents that are out trying to do malicious things, and that we're going to need like our own like good defensive AI agents out there making sure we don't get scammed out of all of…”
Tan: Open-source AI access for consumers is the best insurance against tyranny
“We want all consumers to be able to have from the bottom up, ah, the same access to that same technology. And that's, ah, yeah, the best insurance against tyranny.”
Hu: NeurIPS expanded from 600 papers in 2017 to over 3,000 in 2023
“I think it's like over 10,000 attendees. There were 3000 papers, more than 3000 papers accepted, and I think back in 2017 there was only around 600 papers.”
Hu: Seven 'Attention' authors founded startups collectively valued over $6B
“That paper came out in 2017, and the fun fact, I was just looking this up, out of all those authors, eight authors, seven of them started different companies, and all of the companies in total, they're rate, they're worth valuation more than six billion.”
Taggar: 'ChatGPT wrapper' meme benefited YC by filtering out trend-chasing founders
“Going back to like the chat GBT rapper meme again, I actually think that was great for YC. Because it meant we only got the people who are like tune, who could tune that out. And we're just like, hey, like either I'm just so interested in this technology, I do…”