Jun 19, 2023 · 1h 16m · news
Alex Lebrun: Why the EU's AI Regulation is a Disaster; How Zuck Prepares for Meetings | E1027 · 20VC with Harry Stebbings
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 extensive interview with Harry Stebbings, Nabla Co-founder and CEO Alex Lebrun reflects on his 22-year career in AI, analyzing the mechanics of LLMs and the strategic opportunities for startups in highly regulated sectors. He shares critical insights into Nabla's bottom-up strategy to solve global doctor burnout, while warning against bureaucratic EU regulations and challenging common investor myths regarding AI defensibility and data moats.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Harry holds 21.1% of the talking time here. How this is scored →
Harry as informed peer 3.4 Guest teaching 4.5 Guest disagreement 2.3 Harry pushing back 2.8
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 1:01:03 EU AI Regulation called a disaster
Alex strongly attacks European policy decisions, warning that mandatory consent requirements on past training data make existing LLMs illegal and force European companies to migrate.
Hardest push from Harry ▶ 23:52 Harry challenges startup disruption vs incumbent speedHarry refuses the premise that incumbents are inherently slow at implementing AI, citing concrete examples like Adobe, Notion, and Navan rapidly deploying generative AI.
Biggest teaching moment ▶ 30:14 Curated training data does not eliminate LLM hallucinationsAlex corrects the common misperception that clean input data yields reliable output, detailing how probabilistic autocomplete mechanisms stitch together factually wrong sentences regardless of source quality.
Harry holds his own ▶ 52:45 Harry bringing personal angel investment experience in healthcare techHarry draws on his direct experience with healthcare messaging startup Pando to challenge Alex on the harsh reality of NHS procurement and budget resistance.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| Market Timing and Early Chatbot Hurdles | 2 | 3 | 1 | 1 | Harry introduces Alex and asks about his early chatbot history, Facebook experience, and working with Mark Zuckerberg. Alex shares stories about early bot errors and how Mark Zuckerberg used reverse psychology to test his conviction on hiring concierges. | |
| The "Kim Jong-un" Paradox and Healthcare Mistakes | 2 | 3 | 1 | 2 | Alex discusses the 'Kim Jong-un' entourage trap where successful founders are not challenged enough by investors or teams. Harry prompts Alex to detail early mistakes made at Nabla during its pivot from B2C primary care to B2B. | |
| The VC Hype Cycle and Continuous AI Progress | 3 | 4 | 2 | 2 | Harry shares his perspective as a VC feeling contrarian for making non-AI deals. Alex reframes public perceptions of step-function AI leaps as a continuous 22-year technical progression. | |
| Application AI as More Than a "Thin Layer" | 4 | 5 | 3 | 4 | Harry presses Alex on the common VC critique that application-layer AI is merely a thin, valueless wrapper on LLMs. Alex directly counters this framing, drawing parallels to how C programming language and databases transformed software infrastructure. | |
| Model Churn and Demystifying Hallucinations | 3 | 6 | 2 | 3 | Harry asks about model longevity and cites Emad Mostaque on hallucinations being a feature. Alex educates Harry on unsupervised pre-training versus fine-tuning, referencing the Lima paper to demonstrate high-quality data efficiency. | |
| Incumbent Slowdown, Startup Disruption, and the Google Dilemma | 5 | 4 | 3 | 5 | Harry challenges Alex's stance on slow incumbents by citing Notion, Adobe, and Navan moving fast on generative AI. Alex acknowledges their feature releases but argues incumbents rarely create self-disruptive paradigm shifts due to cost and cannibalization concerns. | |
| Open vs. Closed Models and the Black Box Dilemma | 3 | 5 | 3 | 3 | Alex dismantles common misconceptions about open models and curated datasets, explaining that open parameters do not solve black-box opacity. He clarifies that training on clean data still does not prevent LLMs from hallucinating incorrect sentence joins. | |
| Existential AI Risk and Elon Musk's "Pause" Petition | 4 | 4 | 4 | 4 | Harry brings up Geoff Hinton's fear of existential AI risk and Elon Musk's petition to pause AI development. Alex dismisses apocalyptic scenarios and argues pause petitions are self-serving moves by market leaders attempting to freeze competition. | |
| AI Adoption in Healthcare and Doctor Burnout | 3 | 5 | 2 | 2 | Alex presents stark healthcare metrics, detailing that 3 out of 4 doctors face burnout and spend 49 percent of their time on administrative tasks. He cites extreme examples where executing a single clinical action requires over 200 mouse clicks in legacy EHRs. | |
| Ambient AI Assistants and EHR Interoperability | 4 | 4 | 2 | 3 | Harry asks about data privacy and interoperability in medical records. Alex explains how ambient audio processing functions without storing raw recordings and how Nabla uses Chrome extensions to bypass legacy EHR integration hurdles. | |
| Clinical Shortages and the Failure of B2C Healthcare AI | 5 | 5 | 3 | 5 | Harry pushes back that automating healthcare tasks conflicts with political incentives to protect nursing jobs. Alex counters with World Health Organization data predicting an 18 million clinician shortage by 2030, showing automation is essential for survival. | |
| First-Hand Research in Paris Emergency Call Centers | 2 | 4 | 1 | 1 | Alex recounts conducting field research during overnight shifts at emergency call centers in Paris. He describes dispatchers struggling to type structural notes while managing high-stress emergency calls. | |
| Global Doctor Shortages and Aging Populations | 5 | 6 | 2 | 4 | Harry draws on his own angel investment experience in NHS messaging software to press Alex on healthcare monetisation difficulties. Alex outlines the complex payer-provider-patient dynamic and explains Nabla's bottom-up strategy targeting individual clinicians. | |
| Silicon Valley vs. Europe and Why Startups Sell Too Soon | 3 | 4 | 2 | 3 | Harry asks whether European founders lack Silicon Valley ambition and sell out too early. Alex agrees that French founders produce world-class mathematics talent but historically lacked local role models who held out for multi-billion dollar outcomes. | |
| The EU AI Regulation Disaster and Brexit's Opportunity | 4 | 5 | 4 | 3 | Alex forcefully condemns the EU AI Regulation, describing it as an impractical disaster that would render existing LLM training pipelines illegal due to unfeasible consent mandates. He suggests European startups may be forced to relocate to the UK or US. | |
| Geopolitics of AI: US, China, and Europe's Strengths and Weaknesses | 3 | 4 | 2 | 2 | Alex compares regional AI dynamics, highlighting China's massive state-backed healthcare data advantage, Europe's regulatory self-sabotage, and the US reliance on open immigration policy to retain top global talent. | |
| Quick-Fire Round: Simulations, Consulting Failures, and Model Bias | 3 | 5 | 3 | 2 | In the quick-fire section, Alex asserts that we live in a simulation and rejects the idea that consulting services will dominate AI value accrual. He shares a striking anecdote about computer vision bias at Meta discovered by research scientist Moustapha Cissé. | |
| Journalism, Human Nature, and the Sentient AI Illusion | 3 | 5 | 2 | 2 | Alex discusses journalism and human nature, identifying the illusion of sentient AI as society's biggest misconception. He notes that people fell for MIT's simple Eliza bot in 1966 for the same reasons they project consciousness onto modern LLMs. | |
| Elite Venture Capital, Yann LeCun, and Nabla's Ten-Year Vision | 3 | 4 | 2 | 2 | Alex praise angel investor Yann LeCun and describes how Andreessen Horowitz provided frictionless, low-overhead support during his previous startup. He concludes with his ten-year vision to build a full-stack, AI-native healthcare provider system. |
Statements from this episode (40)
Mark Zuckerberg Cancels Meetings Missing 24-Hour Advance Prep Memos
“He prepared really well every meeting, so you have to send in advance, like a short note about why, what decision you are expecting from him. If you fail to send his document, 24 hours, you know, by the minute before your meeting gets cancelled.”
Serial Founders Face Deference Traps That Cause Uncorrected Early Mistakes
“Starting my third company after two exits, I felt like my investors were always agreed, my team always agreed people around us, and so I wasn't challenged enough, and we made some early mistakes that probably were not challenged enough.”
Alex Lebrun: AI progress is continuous, not discontinuous step functions
“So from the outside, from, for the general public, it looks like a very big step function with huge advancement every 10 years. I think from the inside, it's much more continuous.”
Founders Must Add or Remove 'AI' From Pitch Decks Every 3-4 Years
“As an entrepreneur, the only thing you need to know is when to add or remove AI from your deck, and it would change about every three or four years.”
Lebrun: AI application startups are not merely thin layers over LLMs
“So I think it's really not fair to think that any AI application on top of LLM is just a thin layer.”
Lebrun: Companies will continuously switch between LLMs to remain competitive
“It's how it will work, of course. You have, you know, every two weeks, incredible things coming out of research, and if you just assume it's a static, in a year from now, your product will be very dumb compared to your competitor's products. Or more expensive …”
No AI Models in Use Today Will Be Used in One Year
“Absolutely agree with that. You know, will you drive the car you drive today in 10 years? I don't think so. And 10 years in car industry is like one week in machine learning gravity zone. And so there is, of course, there is so much progress so fast That I don…”
VCs Are Late to Realize Proprietary Data Moats Are Diminishing
“I think VCs are always one train late. So proprietor, having a lot of proprietary data was very, very important for the last cycle five years ago. Maybe it's less and less true.”
Nabla Paid Doctors to Collect 30,000 Medical Consultations for AI Training
“We built a data set of 30,000 consult, medical consultations with patient consent. We have to pay doctors. It's very hard to build this because we need this to bootstrap our product.”
Lebrun: LIMA fine-tuned on 1,000 examples beats GPT-3, rivals GPT-4
“Three weeks ago, there was a paper about Lima. So, so this Lima paper shows that with only 1000 question and answer examples, so very, very small data sets they get something for, use for fine tuning, so the second stage, they get something that performs bette…”
Lebrun: Startups with billions in funding can match OpenAI's capabilities
“I think a new startup with enough talent and enough money, and when I say enough money, I'm talking about billions, just to be clear, then I'm sure you can do as, at least as well as OpenAI.”
A New AI Paradigm Will Displace Notion and Google Docs
“It probably something will come and destroy Notion and Google Docs and all of them with a totally new paradigm that is made possible by AI, and certainly these incumbents won't do it.”
Incentives and Legal Fears Kept Google and Meta From Leading LLMs
“Nobody could predict that LLMs would be so useful and powerful before you train one at this scale. And who in the Google org chart had the incentive to invest five hundred million dollars and just to see this without any business benefit for the company? If yo…”
Lebrun: Open-source foundational AI models will win over closed models
“Open, obviously. You know, LLM are AI in general is an infrastructure in the future, and like every infrastructure, I'm just paraphrasing Jan, but is, is open. Open wins always with infrastructure. And so obviously I think the financial, the foundational model…”
Open-Source LLMs Are Not Inherently Explainable or Transparent
“Even an open model trained with open data, to me, is not that open, because when you have, like, three hundred billion parameters, and it's a huge black box, you don't understand why the output is what it is, is it really open? And so, I'm just putting a littl…”
Lebrun: Training LLMs on Curated Data Does Not Guarantee Trustworthy Outputs
“Even if you feed an LLM with curated data, you don't, it does, it's not guaranteed that the output will be will be good, will be perfect, that you can trust the output. So, feeding an LLM with trusted data doesn't make the output trustable because of how LLM w…”
AI Existential Risk Warnings Are Driven by Publicity Seekers
“I fully agree with Jan and not just because he was my boss at Meta, but I think if you really understand how machine learning works and you're not looking for free publicity, I don't see why you would say something like that, that, that getting more intelligen…”
AI Pause Petitions Are Cynical Attempts to Block Competitors
“It looks to me that the people who are proponent of this pause or to more regulation Are the ones who feel that they are in advance. And so it's like a way to say, guys, let us, we are in front. I don't want more people to start the race.”
Lebrun: Doctors spend an average of 49% of their time on administrative tasks
“Doctors spend on average 49% of their time doing this kind of administrative task, as opposed to caring for patients.”
Lebrun: Three out of four doctors suffered burnout symptoms in the last year
“Three out of four doctors have suffered from burnout symptoms in the last year.”
A Single Clinical Action in an EHR System Required 227 Mouse Clicks
“I checked the other day, there was one specific clinical action that took 227 clicks, mouse clicks, to be performed in the system.”
Epic Systems Spent Millions Lobbying to Block US EHR Interoperability Laws
“Epic spent billions to millions to try to block, but it finally was after, I think, 20 years of battle was passed”
Nabla Uses Chrome Extensions to Extract Data From API-Less EHR Systems
“At Nabla, you know, we use things like Chrome extensions to get the data from the browser. There is no API. We don't care.”
Lebrun: WHO projects global shortage of 18M clinicians by 2030
“The World Health Organization says that we are missing eighteen million clinicians by 2030.”
Lebrun: Healthcare AI startups fail by focusing on patients over doctors
“I think the mistakes that many AI startups are doing in healthcare, and what we did initially at Nabla too, is we focused directly on the patients.”
Lebrun: AI assistants will cut emergency call response times to 20 seconds
“And it will be faster too. So it's a win-win, you know, to be like maybe 20 seconds instead of one minute and with more empathy.”
NHS Messaging App Reached 100,000 Users but Generated Zero Revenue
“My first ever investment, fun fact, was like a WhatsApp for doctors and nurses in the UK. We went to a 100,000 doctors and nurses in the NHS, and then it came to getting paid. And it was... And they just don't pay.”
Lebrun: Healthcare startups must identify who pays before defining the problem
“In healthcare, even starting with a problem is not enough. Start with who is paying, and then How you do frame the problem for this person to pay, and then what is the solution for this problem eventually?”
Lebrun: AI founders no longer need to be in Silicon Valley
“The talent is very, very distributed. It will get more distributed, and I think you should get close to your customers.”
France Produces Great AI Engineers but Sucks at Scaling Companies
“In France, we are really good in we have very good AI engineers, machine learning engineers, because the education system is free, is very focused on mathematics. So we produce lots of good engineers, but we suck at growing companies.”
Europe Is 10 Years Behind the US and China in AI
“Europe is probably 10 years late compared to U.S. And China, and with a new regulation, we are probably going to take 50 years more.”
Proposed EU AI Regulation Would Make All Existing LLMs Illegal
“One of the percent of the LMs that were trained these last three years would be illegal in Europe.”
Lebrun: Lack of Privacy Regulations Gives China Major AI Data Advantage
“So, so China, they have, like, a key advantage that, like, there is no GDPR or very few regulation internally and so the data quality they have and the amounts, both the quality and quantity of data is incredible in China.”
Lebrun: Strict Immigration Restrictions Will Eventually Harm US AI Dominance
“Immigration may be laws, may be something that will be a problem eventually in the US because the talent is so distributed with the new, you know, you can learn deep learning very easily from everywhere in the world on a tablet today, and it will be more and m…”
Lebrun: AI-native startups will displace incumbents across all industries within five years
“I think AI will enable a new generation of players in every industry that will kill the incumbents eventually.”
Element AI Failed and Was Acquired for Roughly Its Total Capital Raised
“There was a company called Element AI a few years ago in Canada, very ambitious company would try to do services like that. And they eventually failed. They were acquired by ServiceNow for, sort of, amounts are raised.”
Lebrun: Believing AI chatbots are conscious is a persistent illusion
“It's that they think it's conscious, and they are influenced by the form, like, oh, perfect answer, the perfect form, and it's a chatbot, it's, it answered my question, and it's, and influenced by the form, they make conclusion on the deep inside, you know, of…”
Andreessen Horowitz Held Zero Board Meetings and Responded in 20 Minutes
“Personally, you know, I worked with Andreessen Arvitz with my previous startup. It was incredible because they never call. They never ask for anything. We didn't even have board meetings. And when I, every time I need something, you know, 20 minutes I get huge…”
Lebrun: Founders Need VCs Who Are Ex-Entrepreneurs, Not Former Bankers
“10 years ago, VCs in Europe were all from the financial industry, you know, there is nothing about financial industry, finance in being a VC, a little bit, of course, but it's mostly about entrepreneurship, and so 10 years ago, all the VCs I was working with i…”