Why Google failed to make GPT-3 + why Multimodal Agents are the path to AGI — with David Luan of Adept

Episode Summary

Executive Summary: David Luan traces the evolution of modern AI from early RL and Transformers to today’s agent era, arguing that compute, data, and product-user co-design now drive progress. He explains Adept’s enterprise-first strategy: build highly reliable agents that can use computers like humans, learn from real workflows, and become AI teammates rather than chatbot demos.

Main Topics: David Luan’s background and AI lineage (Priority: 5/5): Luan recounts his path from Dextro and Axon to early OpenAI, Google Brain/LLM leadership, and founding Adept, framing his perspective as shaped by multiple AI eras. How AI progress has changed over time (Priority: 5/5): He argues AI moved from small-team research breakthroughs to a compute-centric scaling era, and is now entering a phase where product, users, and technology must co-evolve. Agents as the long-term direction (Priority: 5/5): Luan defines AGI around doing anything a human can do on a computer, making agents the natural endpoint of LLMs plus RL lessons, behavioral cloning, and multimodal perception. What Adept is building (Priority: 5/5): Adept is positioned as an enterprise agent company that helps humans delegate complex computer workflows reliably, rather than a consumer chatbot, API, or open-source model company. Reliability, augmentation, and enterprise deployment (Priority: 4/5): He emphasizes that enterprise use requires nines-level reliability, and that Adept’s augmentation-first approach creates a data flywheel and better human oversight. Model and product stack: multimodal, UI interaction, and sensors/actuators (Priority: 4/5): Adept’s research focuses on fast multimodal models, screen understanding, dense OCR, and computer control as the practical path to agents that can operate across software tools. Industry positioning and talent/research strategy (Priority: 4/5): Luan argues foundation models will commoditize, so advantage comes from vertical integration around agents, customer feedback, and grounded evals tied to real workflows.

Key Arguments: AI progress is increasingly driven by deep co-design between product, users, and technology, not just bottom-up research. Agents are the correct long-term direction because AGI should mean a model that can do anything a human can do on a computer. LLMs are a form of behavioral cloning of human knowledge; multimodal models will extend that to the visual world. RL taught the field goal-directed optimization, but pure RL is too slow; the best path combines RL lessons with human demonstrations and data. Adept’s enterprise focus forces reliability, which is necessary for real workflows and creates a stronger data flywheel than consumer demos. Using computers like humans is more practical than relying only on APIs, because most real workflows do not have clean API coverage. Foundation models alone will be commoditized by large labs and open source; durable advantage comes from an agent stack built on top of them. Augmentation is strategically better than full automation because it keeps humans in the loop, improves adoption, and generates training data from hard tasks.

Data Points: OpenAI hire rank: 30th–35th hire - Luan describes joining OpenAI very early in the company’s history. Time at OpenAI: About 2.5 to 3 years - He led engineering and helped organize early teams and research direction. Time leading Google LLM efforts: About 1 year - He says he led Google’s LLM effort and co-led Google Brain during that period. Adept founding year: 2022 - He says Adept was started in 2022. Adept founding month: January 2022 - He notes the company started in January 2022 when agents were not yet widely understood. Adept funding raised: $420 million - Mentioned in the discussion of fundraising and company scale. Latest round: $350 million Series B - Referenced when discussing Adept’s funding and partner meetings. Q1 capacity: Sold out - Luan says Adept is sold out for Q1 in terms of onboarding bandwidth. Model reliability target: In the nines - He says enterprises cannot use systems that are not highly reliable. Act One demo: Public demo from early Adept - Used as a reference point for agent demos and workflow automation.

Pivotal Quotes: "the number one driver of AI progress over the next couple of years is going to be the deep co-design and co-evolution of product and users for feedback and actual technology" — David Luan: He explains why the AI field is shifting from pure research to product-driven iteration. "we want to build an AI agent that can do ... anything a human does on a computer" — David Luan: His core definition of Adept’s mission and the company’s AGI-oriented vision. "AI has really been the story of compute and compute plus data and ways in which you could change one for the other" — David Luan: His closing summary of the industry’s main historical driver.

Implications: The conversation suggests AI’s next phase will reward companies that combine strong models with real workflows, reliability, and user feedback. For listeners, the key takeaway is that agents—not chatbots—may become the dominant interface for knowledge work and enterprise automation.