Aug 16, 2023 · 46m · mad
Reinventing Search with AI: Richard Socher on Building You.com & the Future of Google
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 this episode of The MAD Podcast, You.com CEO Richard Socher discusses building an AI-first web search engine, implementing Retrieval Augmented Generation (RAG), investing in early-stage AI startups via AIX Ventures, and predicting the future growth trajectory of generative AI.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 20.4% of the talking time here. How this is scored →
Matt as informed peer 3.2 Guest teaching 4.5 Guest disagreement 1.1 Matt pushing back 0.7
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 42:17 Reframing AI doom and infinite exponential claims
Richard firmly rejects media hype around self-conscious AI, existential threats, and unending exponential progress, arguing that physical limits and human consumption create natural S-curves.
Hardest push from Matt ▶ 20:47 Challenging the non-ad strategyMatt questions Richard on avoiding advertisements when Google built a multi-billion dollar enterprise on search ads, pushing Richard to justify You.com's business model.
Biggest teaching moment ▶ 19:07 Explaining why small teams shouldn't train models from scratchRichard corrects common founder misconceptions about training custom LLMs, explaining that without hundreds of millions of feedback data points, fine-tuning old base models quickly loses to rapidly advancing API models.
Matt holds his own ▶ 38:59 Framing the three-layer AI value stackMatt demonstrates clear sector expertise by framing the generative AI market into a distinct three-tier architecture of foundation models, middleware, and application layers.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| Richard Socher's Background and Founding You.com | 2 | 3 | 1 | 0 | Matt opens with a warm, factually detailed introduction of Richard's academic credentials and startup background. Richard responds with a extended narrative of his transition from Stanford researcher to MetaMind and Salesforce, culminating in the founding of You.com. | |
| Overview of You.com's Product Suite and Features | 3 | 4 | 1 | 1 | Matt lists You.com's product family and prompts Richard to explain how they interact. Richard explains user intent modeling and contrasts search engine behaviors with chat-first architectures. | |
| Retrieval Augmented Generation (RAG) and Hallucination Control | 4 | 5 | 0 | 1 | Matt asks technical questions regarding how You.com detects hallucinations and connects live data to LLMs. Richard educates the host on Retrieval Augmented Generation (RAG) and intent classification. | |
| You.com's LLM Architecture and Fine-Tuning | 4 | 4 | 1 | 1 | Matt asks about YouBot's architecture, underlying foundation models, and dataset curation for copyright-free images. Richard details their hybrid pipeline combining fine-tuned open-source models with external partner APIs. | |
| Insights and Advice on Training and Deploying LLMs | 2 | 6 | 2 | 0 | Matt asks for general advice on fine-tuning and deploying LLMs. Richard firmly warns smaller teams against training custom LLMs from scratch, noting that fast-moving foundation models require hundreds of millions of feedback interactions to beat standard APIs. | |
| Business Model, Monetization, and Advertising Strategy | 4 | 4 | 1 | 2 | Matt presses Richard on why You.com eschews advertising despite search being a multi-billion dollar ad market. Richard candidly admits users dislike paying subscriptions even more than ads, explaining their pivot toward high-quality privacy-focused ads. | |
| Target User Personas and Community Growth Strategies | 2 | 3 | 0 | 0 | Matt inquires about You.com's user personas and community strategy. Richard outlines their focus on developers, students, and creative users alongside regional growth. | |
| AIX Ventures Portfolio and Investment Thesis | 3 | 5 | 0 | 1 | Matt asks about AIX Ventures' investment thesis and probes specifically on how machine learning is used to reduce cow methane emissions. Richard explains adaptive digestive modeling in cows and protein sequence generation. | |
| Evaluating AI Founders and Team Dynamics | 3 | 5 | 2 | 0 | Matt asks how AIX evaluates technical AI founders versus domain experts. Richard shares lessons on founder dynamics and critiques past tech trends where speech recognition was misapplied to visual search problems. | |
| AI Industry Value Layers: Foundation, Middleware, and Applications | 5 | 6 | 3 | 1 | Matt presents a structured three-layer stack framework (foundation, middleware, applications) and asks about future growth trajectories. Richard pushes back against doom and AGI hype, explaining that AI progress will follow S-curve limits due to physical and human consumption bottlenecks. |
Statements from this episode (15)
Socher: MetaMind joined first FDA trial for radiology deep learning
“I think we were in the first FDA proof trial of a deep learning classifier in a production radiology system.”
Socher: Salesforce research team invented prompt engineering in 2018
“As part of the research team, we actually invented, ah, prompt engineering in 2018”
Socher: You.com launched the first internet-connected LLM search chatbot
“Then we were the first last year to bring chats into the search context also by having the first LN shipped to the world, as far as we know that connected and had a connection to the internet.”
Socher: You.com solved AI hallucinations before anyone else
“We figured that, I don't know, problem of hallucinations out before anyone else.”
Socher: Microsoft Bing copied You.com's AI image generation feature
“That was eventually copied, and Bing also has that feature now within Bing.com.”
Socher: Don't train LLMs from scratch without hundreds of millions of responses
“And yeah, I don't think you should attempt it unless you have, you know, hundreds of millions of responses that to really like train your own model versus just fine-tune it a little bit. Like certainly not from scratch.”
Socher: Larger Nvidia GPUs will enable running bigger LLMs on single GPUs
“There's just beyond a certain size that started running on a single GPU, but that will also change in the future as Nvidia you know, builds larger and larger memory GPUs so that you can then run log and log your language models on a single GPU, and so that wil…”
Socher: Consumers dislike paying for services even more than ads
“Truth is people don't like paying for things and they like it even less than advertisements.”
Socher: Studies show developers using You.com are more efficient than on Google
“Yeah, we actually did some studies and developers who use you.com are way more efficient than developers who use Google.”
Socher: A good chunk of developer queries will move into IDEs
“I think ultimately a good chunk of developer queries will move to be directly in the IDE, an integrated developer environments that the quotas are using but not all of them.”
Socher: Generating protein sequences with LLMs will transform medicine next decade
“They use large language models to generate new protein sequences. And I think that will change all of medicine in the next decade in a massive way.”
Socher: The best startups are usually founded by two co-founders
“A lot of companies die because the founders don't get along, but the best companies are usually you know, founded by two founders.”
Socher: Founders returned AIX check after co-founder friction
“I had once this dinner, some really incredible security researchers and each of them was a rock star, but at the dinner, I kind of sensed a little bit of you know, some disagreements and it kind of interrupted each other a little bit here and there. So I put j…”
Socher: Only a few dozen horizontal AI companies will see major success
“I do think there will be probably only a few dozen horizontal companies that are going to be really major breakout successes. But there will be thousands of vertical application, what you call it, sort of vertical companies that are successfully using AI for t…”
Socher: Plumbers will eventually earn more than marketers as AI automates knowledge work
“Probably at some point, your plumber might get more than your marketer because the marketer can use tools and be way, way more efficient, but the plumber AI is going to take a very long time because None of the plumbing companies are collecting data, having ro…”