Interview, Fireside Chat
Alex Lebrun: Why the EU's AI Regulation is a Disaster; How Zuck Prepares for Meetings | E1027
20VC with Harry StebbingsAlex Lebrun, Zuck, Mark Zuckerberg, Emad Mostaque, Yann LeCun, Jeff Hinton, Elon Musk, Adam Scheer, Patrick Peloux, Andres, Arvids
- Core Market Insight: AI will not replace doctors, but "doctors who use AI will replace doctors who don't."
- The primary disruption will be administrative efficiency rather than clinical replacement.
- Speaker Background: Alex (Francesc) Campoay founded his first company, Virtuose, 22 years ago focused on chatbots.
- He joined Facebook via the acquisition of his second startup, Wit.ai.
- He is currently building Nabla, his third venture, in the healthcare AI space.
- Market Timing Analysis:
- Early predictions (2004–2016) of AI replacing humans in call centers were premature due to technical limitations.
- Market timing is now favorable because technology has matured and consumer/professional adaptation has occurred over the last 10–12 years.
- Key Takeaways from Facebook:
- Large organizations can operate with the speed and efficiency of a "well-oiled machine" when properly structured.
- Decision-making requires pre-meeting documentation; meetings are fast, data-driven, and decision-oriented.
- Mark Zuckerberg's strategy of "reverse thinking" forces founders to rigorously validate the truth of their assumptions by challenging them from the opposite angle.
- Startup Governance & Leadership:
- Campoay warns against the "Kim Jong-un entourage trap" where investors and teams agree with the founder to avoid conflict.
- Solution: Engage external advisors who have no incentive to please the founder to provide unbiased feedback.
- Internally, he mandates a culture where the team is explicitly empowered to challenge decisions.
- Nabla's Evolution & Mistakes:
- Initial strategy involved starting as a B2C virtual primary care clinic to gain domain expertise.
- Lesson Learned: The company pivoted back to a B2B model after realizing it moved too far from the core product and target market before validating the path.
- Generative AI Landscape:
- Public perception views AI progress as discontinuous (huge steps), while the reality is continuous evolution (e.g., GPT-4 based on GPT-3, Transformers since 2016).
- Hype Cycle: Campoay notes the cycle repeats every 3–4 years, advising entrepreneurs to know when to include "AI" in their pitch decks based on investor sentiment.
- Technical Reality vs. Perception:
- LLMs as Infrastructure: Large Language Models are a new, non-deterministic resource similar to C language or databases, requiring novel engineering to control.
- Model Switching: Products must expect to switch foundational models every few weeks due to rapid advancement; static models will become obsolete quickly.
- Hallucinations: Described as a feature of design (the model must output something), not just a bug, due to the probabilistic nature of the technology.
- Data & Training Strategies:
- Fine-tuning Efficiency: New research (e.g., the "Lima" paper) suggests high-quality models can be fine-tuned with as few as 1,000 Q&A examples, rivaling larger models like GPT-4.
- Proprietary Data: While proprietary data is less critical for bootstrapping new models due to better pre-trained architectures, it remains essential for specific domain fine-tuning (e.g., Nabla's dataset of 30,000 medical consultations).
- Trust Fallacy: Feeding an LLM with trusted, curated data does not guarantee trustworthy output; the model can logically combine true facts into false conclusions.
- Geopolitical & Regulatory Risks:
- Europe: Faces a significant disadvantage due to the EU AI Act, which Campoay calls "prohibitive" and "illegalizing" current training practices (e.g., requiring explicit consent for all training data).
- Impact: Startups may need to relocate model training operations outside the EU or move to the UK.
- China: Holds an advantage in data quality and quantity due to a lack of strict privacy regulations like GDPR.
- US: Remains a leader but faces future risks from immigration laws that could limit talent influx.
- Europe: Faces a significant disadvantage due to the EU AI Act, which Campoay calls "prohibitive" and "illegalizing" current training practices (e.g., requiring explicit consent for all training data).
- Healthcare Sector Specifics:
- Current State: Doctors spend 49% of their time on administrative documentation, leading to burnout in 3 out of 4 physicians.
- Systemic Failure: Electronic Health Records (EHR) are described as "dinosaur systems" requiring up to 300 clicks for single actions.
- AI Application: The immediate value is an "ambient AI assistant" that listens to calls/visits and auto-generates clinical notes, reducing cognitive load.
- Go-to-Market Strategy:
- Must target the payer/provider dynamic; in the US, selling to private clinics (bottom-up adoption by physicians) is more viable than waiting for government (hospital) procurement.
- Starting with patient-facing (B2C) products in healthcare is often too risky due to regulatory complexity and lack of immediate payment clarity.
- Global Supply/Demand: There is a shortage of 18 million clinicians by 2030, making efficiency gains critical rather than a threat to employment.
- Fundamental Misconceptions:
- Consciousness: AI is not conscious; the perception of sentience is a historical trap dating back to the ELIZA bot in 1966.
- Bias: All models contain bias; diverse teams (e.g., including researchers from different cultural backgrounds) are essential to identify blind spots in training data.
- Timeline for Disruption: Existing incumbents will likely fail to disrupt themselves; new startups built on new AI paradigms will replace legacy models within 5 years.
- VC & Ecosystem Observations:
- Best Investors: Effective VCs act as on-demand strategic partners (e.g., helping with real estate, marketing, hiring) rather than micromanagers, providing help within 20 minutes of a request.
- European VC Gap: Historically dominated by former bankers rather than ex-entrepreneurs, leading to a lack of operational guidance; this is slowly improving.
- French Startup Culture: French engineers are highly skilled, but the ecosystem historically suffers from selling too early due to a lack of discipline or examples of large-scale growth.
- Future Outlook (2033 Vision):
- Physician Role: Every physician will have an AI assistant, becoming 10x more efficient and spending more time on patient care.
- Systemic Decision Making: AI will eventually handle high-level resource allocation and emergency services regulation at the city/country scale.
- Ultimate Goal: To build a full-stack, data-driven healthcare system from scratch, removing current regulatory and technological constraints.