Sarah Guo and Elad Gil: The Future of AI Investing

Detailed Summary

Introduction (0:00 - 3:45)

• Michael Mignano, a partner at Lightspeed, introduces the episode, noting a departure from his usual interviews with founders to speak with AI investors Sarah Guo and Elad Gil.
• Sarah Guo, founder and partner at Conviction, describes her firm's focus on AI-native software 3.0 companies, highlighting the revolutionary nature of AI in human-computer interaction.
• Elad Gil, an investor and advisor to companies like Character.ai, Perplexity, and Pika, shares his background in early ML systems at Google and his extensive investment in AI and non-AI companies.
• They mention their shared investment in Pika and their podcast, "No Priors," which covers a broad survey of AI across researchers, founders, and large companies.

The AI Investment Boom (3:45 - 11:12)

• Sarah Guo admits she didn't predict the consumer reaction to ChatGPT, expecting a slower adoption, but recognized the scaling walls and effectiveness of Transformers.
• Elad Gil, having worked in ML for nearly two decades, noted GPT-3 as a significant step function, followed by GPT-3.5, which prompted him to fund many AI ventures.
• He describes five waves of human capital entering the AI space: AI-native builders, hardcore dev/infra people, B2B app developers, consumer app developers, and finally, enterprise adoption.
• The discussion emphasizes that the current AI boom feels like an early stage with significant growth potential.

The Go-Forward Opportunity for Foundational Models (11:12 - 16:31)

• Sarah Guo discusses the unpredictability of venture funding for foundational models, noting that widespread investment only occurred after ChatGPT's revenue growth became apparent.
• She highlights the success of companies like Mistral, which demonstrated the capability of smaller, open-source models, challenging previous dismissals.
• While scaling language model training is becoming increasingly expensive, efficiency in training is improving, and open-source contributions are significant.
• Elad Gil believes there will be new model companies, specializing in various modalities (voice, image, video, audio, code) and scientific domains (biology, physics, materials, math).
• He anticipates that most funding will go to late-stage model companies due to capital intensity, but the number of new companies will be higher in tooling and applications.

What Does the Next App Wave Look Like? (16:31 - 23:35)

• Elad Gil categorizes the next app wave into B2B and consumer, noting that B2B apps will likely emerge first, followed by consumer applications.
• He quantifies the opportunity by comparing the half-trillion-dollar software market to the 3.5 trillion dollars in services headcount payroll addressable by generative AI.
• Sarah Guo emphasizes that AI enables software to perform tasks previously paid for, opening new markets for prosumers and small businesses.
• The success of consumer-focused AI companies like OpenAI (with ChatGPT's consumer revenue) and Midjourney suggests a significant opportunity in the consumer space.

SMBs and AI: New BigCo’s or New Tools? (23:35 - 31:30)

• Elad Gil cautions against overstating AI's immediate impact on making every company a one-person operation, emphasizing the continued need for human involvement.
• He uses Harvey, a legal AI tool, as an example: it's a fantastic tool for legal professionals but won't immediately obviate the profession.
• Sarah Guo notes a shift in entrepreneurial mentality towards building capital-efficient companies, aiming for high revenue with fewer employees.
• Elad Gil highlights that historically, many successful tech companies (Microsoft, Dell, Yahoo, eBay, Google, Instagram, YouTube) were capital-efficient.
• They discuss the challenges of the SMB market, where businesses are busy and only invest in essential tools, making it a difficult market for many products.

Vertically Integrated Models (31:30 - 39:54)

• The discussion turns to the defensibility of vertically integrated AI companies that build both products and their own foundational models, like Midjourney, OpenAI, and Pika.
• Sarah Guo explains that the line between fine-tuning and pre-training is blurring, especially with open-source base models.
• She argues that defensibility in AI applications still relies on understanding customer workflows, delivering ROI, and effective distribution, similar to successful SaaS companies.
• Elad Gil notes that the first wave of AI founders often over-invested in model building rather than customer focus.
• He predicts a continued subset of companies building their own models, especially in image, video, and audio, where training is cheaper than large language models.
• The market will see a mix of generalizable large language models and specialized, performant models, with vertically integrated companies more common in science and robotics.
• Sarah Guo points out that benchmarks are often weak proxies for real-world performance, and true value comes from how models work in specific applications.

Chat as an Interface: The Future or the Past? (39:54 - 49:19)

• Sarah Guo acknowledges the existence proof of chat as a compelling interface (e.g., AI girlfriend apps) but doubts it will be chat-only, expecting multimodal interfaces.
• The bull case for chat as a primary interface is its naturalness and the ability of AI to understand intent, potentially leading to a new platform and hardware.
• Elad Gil believes there will be a series of interfaces, with chat working well for many tasks, alongside multimodal interactions like GPT-V for image understanding.
• He draws a parallel to the early PDA era, where handwriting input was initially favored over typing, highlighting how people often overthink interface design.
• Sarah Guo clarifies that skepticism about chat interfaces often conflates user experience with technological possibility, arguing that AI-driven intent understanding can make interactions much easier.

Hardware x AI (49:19 - 55:45)

• Elad Gil discusses his long history of investing in hardware-plus-AI companies like Samsara (fleet management) and Anduril (defense), where hardware enables interaction with the physical world.
• He suggests that most consumer AI applications will likely integrate into existing platforms (smartphones), with standalone devices needing to offer truly new capabilities.
• Sarah Guo emphasizes that new hardware can enable new applications, especially when it offers capabilities like better battery management, different sensors, or contextual intelligence.
• Elad Gil adds robotics and self-driving cars as significant areas where specialized hardware combined with foundational models will lead to breakthroughs.

Web3 and Authenticity (55:45 - 1:00:11)

• The conversation shifts to Web3's potential role in AI, particularly for content authenticity and provenance.
• Sarah Guo sees identity and provenance as classic application areas for Web3, though usability for creators and IP owners remains a challenge.
• Elad Gil has long believed in blockchain-resident identity for content origin and agent credentials, allowing for secure, partial data revelation.
• He notes the complexity of content provenance, especially with derivative art, where even removing original works from training data might not prevent style replication due to existing derivative works.

Where Is Regulation Headed? (1:00:11 - 1:02:13)

• The discussion touches on the political nature of AI regulation, particularly regarding training data and fair use.
• Sarah Guo believes a strong case for fair use exists, provided there are safeguards around model outputs, and expects battles to be fought at the application layer.
• She reiterates Ilia Sutskever's point that collecting new data will be crucial for continued model progress, and blockchain-based microtransactions and identity systems could facilitate this by enabling abuse-resistant data contributions.

Conclusion (1:02:13 - 1:05:05)

• Michael Mignano thanks Sarah Guo and Elad Gil for the insightful conversation.
• They plug their podcast, "No Priors," available at noyenpri.com or by searching on podcast platforms.
• The hosts express appreciation for the discussion and the opportunity to share their perspectives on the future of AI investing.
• The episode concludes with a reminder to rate and review the podcast and to follow Lightspeed for more content.