Alex Lebrun on AI Assistants for Doctors, Model Limits and Building Nabla

Harry Stebbings welcomes Alex Lebrun, co-founder and CEO of Nabla, to discuss AI, startups and healthcare. Lebrun's short answer to whether AI will replace doctors is that doctors who use AI may replace those who do not: a prediction, not an established outcome.

Lebrun traces his fascination with chatbots to a summer night in Paris more than two decades earlier. He founded VirtuOz to build customer-service bots, believing at first that they would quickly become intelligent enough to replace call-center workers. The longer he worked on them, the more difficult the task appeared.

A 2004 deployment for the French railway illustrates the gap. When a customer said her name was wrong on a reservation, the bot treated “wrong on the reservation” as her name. Lebrun says both technology and public familiarity have since advanced, and he thinks the market may finally be ready.

After Facebook acquired his second company, Wit.ai, Lebrun spent four years there. Sitting on a terrace at its headquarters, he recognized the comfortable situation of employees in the television show Silicon Valley and decided to return to building. He wanted to apply what he had learned in AI research to the real world.

Facebook surprised him by combining scale with speed. In 2015 he proposed hiring 200 human concierges to help train a bot, despite having no budget for it. He recalls Mark Zuckerberg challenging the idea in a short meeting and approving it within a week.

Lebrun says Zuckerberg prepared for successive short meetings by requiring a note explaining the decision requested. If it did not arrive 24 hours beforehand, the meeting was canceled. Zuckerberg would review the material and gather information before the meeting, making a fast decision possible.

For the concierge proposal, Lebrun prepared answers to objections. Instead, Zuckerberg said that if the idea worked they should hire 100,000 people, prompting Lebrun to explain why the team should be cautious. Whether intentional or not, reversing the premise helped him examine his own project more honestly.

Starting a third company did not become easier, Lebrun says: old lessons prevent some mistakes but new risks emerge. He describes an entourage trap in which, after two exits, investors and employees were too inclined to agree with him, leaving early decisions insufficiently challenged.

At Nabla, an engineering team initially launched a virtual primary-care clinic to learn healthcare. Lebrun considers that decision useful but expensive; in his account, the company then took too long to recognize it needed to return to a business-to-business model. He sought external critics with fewer incentives to please him and encouraged internal disagreement.

Stebbings asks whether recent AI excitement reflects a technical leap or years of cumulative work. Lebrun says progress looks discontinuous from outside but more continuous to researchers: each wave builds on previous models and techniques. His recollection of particular model and paper dates is approximate.

Having lived through several funding cycles, Lebrun jokes that entrepreneurs must know when to add or remove “AI” from their decks. Investor enthusiasm has fluctuated repeatedly since his first fundraising in the early 2000s; he welcomes the attention while recognizing the cycle.

He rejects the idea that an application built on a foundation model is necessarily a valueless thin layer. Like earlier programming languages and databases, an LLM is infrastructure, but developers still have to build useful products and manage hallucinations, nondeterminism and rapidly changing models.

Lebrun says expertise helps select a model, understand where it fails and build supporting machine learning. Companies will need to evaluate new models and change them without breaking existing users' products. His expectation that models would be switched frequently was a 2023 prediction.

Asked about Emad Mostaque's description of hallucination as a feature, Lebrun suggests he may mean that a language model produces plausible continuations even when it lacks a factual answer. Lebrun agrees with the prediction that models then in use might be replaced within a year; the forecast is his judgment at the time.

Investors often emphasize proprietary data, but Lebrun thinks its importance has changed as pretrained models improved. He says Nabla assembled 30,000 medical consultations with patient consent and paid doctors to build a dataset for its product; the figure and collection account are his company's description.

He explains pretraining as learning from a large corpus by predicting the next token, followed by fine-tuning to shape behavior for a task. He points to the then-recent LIMA research as an example of what high-quality, relatively small fine-tuning datasets might achieve. He cautions that researchers were still learning how these systems worked.

Asked how small a useful dataset could become, Lebrun says it depends on the domain and that he does not know the limit. He expects some information will always be needed to specify the desired behavior; his suggested lower bound is speculation rather than an established threshold.

On new foundation-model companies, he cites the recently formed Mistral and says exceptionally strong researchers, focus and very large funding could allow another startup to compete. He describes billions of dollars as the scale he has in mind, an investment estimate rather than a guarantee.

Incumbents have distribution, Lebrun says, while startups may more readily pursue new product paradigms. Stebbings counters that companies such as Adobe and Notion have moved quickly. Lebrun accepts that some incumbents are fast but argues that adding AI features to existing workflows may leave room for more fundamental redesign.

He offers an explanation for why large firms did not release a ChatGPT-like product sooner: the benefit of spending heavily on a large model was uncertain, and individual incentives and legal concerns made the project difficult to prioritize. This is his interpretation of those companies' decisions.

Lebrun predicts open foundation models will ultimately win as infrastructure, while warning that open weights do not make a model transparent or predictable. Even when training data and parameters are available, a large model can remain difficult to explain. In healthcare, he says, applications need extra work to manage its limitations.

Stebbings asks about national datasets. Lebrun is unsure they are the right organizing principle and instead sees a case for domain-specific data, such as medical text. But curated input does not guarantee a correct output: the model can combine pieces into fluent yet factually wrong statements.

He describes an LLM as a probabilistic autocomplete system. Its confident form can hide errors in facts or reasoning, even if it was trained on trusted material. Removing a noisy source alone, in his view, would not eliminate wrong answers.

On existential risk, Lebrun agrees with Yann LeCun that intelligence does not automatically imply a desire to dominate. He says he does not know why Geoffrey Hinton sees the issue differently. He speculates about the incentives of some people calling for a pause; those are his opinions, not established motives.

After decades working in AI, Lebrun feels both vindicated by public interest and frustrated when persuasive narratives outrun technical understanding. He acknowledges that some people raising concerns are credible, although he disagrees with many claims about catastrophic risk.

He believes adoption has reached a meaningful moment: people are already using ChatGPT to change how they work. That does not mean a chatbot is ready to replace every worker. For years, he says, he first tried to convince customers the technology was ready and later had to caution them against overestimating it.

Turning to healthcare, Lebrun predicts that AI will change clinical work and that doctors who embrace it will have an advantage. He argues that computerization has often added administrative obligations to clinicians rather than freeing them to care for patients.

He cites estimates that doctors spend 49% of their time on administrative work and that three in four have experienced burnout symptoms. He connects documentation burdens to reimbursement and legal protection, but the interview does not establish the source, population or causal basis for those figures.

Lebrun describes electronic health record systems as cumbersome, citing a clinical action that he says required 227 clicks, with some clinicians reporting more. His proposed assistant would follow the consultation and help with documentation so the doctor could attend to the patient.

He says Nabla captures consultation audio to obtain context but does not store the audio, and envisions connecting the assistant to patient records. He describes varying degrees of interoperability among record systems and says browser extensions can provide an entry point where an API is unavailable. These are company claims, not an independent privacy or security audit.

Stebbings asks whether improving productivity could threaten jobs and create political opposition. Lebrun responds that clinician shortages create demand for help, citing a World Health Organization projection. The exchange does not establish how every occupation or healthcare system would be affected.

Lebrun thinks many healthcare startups, including Nabla initially, aim too quickly at a patient-facing automated doctor. In his view, technical limitations, missing data, regulation and the complexity of healthcare make that leap premature. He prefers first helping clinicians within existing workflows and learning alongside them.

He recalls meeting French emergency physician Patrick Pelloux, who invited him to observe an emergency call center overnight. There he watched call takers gather addresses and descriptions from distressed callers before a physician assessed the appropriate response. His figures for call volumes are presented as recollections, not independently verified statistics.

The call takers' expertise lies in calming people and understanding complex situations, Lebrun says. He envisions an assistant listening alongside them and drafting notes, allowing them to focus on the caller. His claim that an early trial worked well is an anecdotal company report, not evidence of clinical effectiveness.

Asked how often calls result in emergency services being sent, he estimates roughly one in five based on a few nights of observation and explicitly says it is not an official figure. He distinguishes different response teams and the most critical cases.

Lebrun says doctors have become more receptive as healthcare systems face growing pressures. He attributes access problems to clinician shortages and rising demand, including chronic illness and an aging population. These are his broad observations, not a country-by-country assessment.

Stebbings describes investing in a communications product for healthcare staff and asks how to build a business when those who benefit do not control the budget. Lebrun says the payer, provider and patient often have different incentives, making adoption difficult even when a product creates value.

A modest improvement may fit existing workflows, while an ambitious redesign can be too disruptive to sell. Lebrun advises healthcare founders to find a balance: the product must fit well enough for adoption without abandoning the larger ambition.

His first question for a healthcare startup is who pays. Nabla's vision of an emergency-call assistant faced a difficult public-sector buying process, he says, whereas documentation presented a significant problem for US providers. He warns that endless pilots can consume a startup without leading to deployment.

Lebrun says thousands of doctors were using Nabla daily in the US at the time. The company offered a web application and browser extension that individual physicians could try and then recommend to their clinics. Both the usage figure and go-to-market account come from the founder.

Asked whether AI founders must be in Silicon Valley, he recalls meeting Siri co-creator Adam Cheyer there and gaining confidence to reject early acquisition offers for Wit.ai. He thinks talent is now more distributed and advises founders to stay close to customers; the earlier networking experience still mattered to him.

Lebrun praises French mathematical education and AI engineering talent but believes the country has historically struggled to grow companies. Turning down a life-changing acquisition offer is difficult, especially for a first-time founder. He hopes more successful scale-ups will provide examples for the next generation.

In this 2023 interview, he sharply criticizes a proposed European AI regulation, arguing that requirements concerning training data could make model development impractical. He explicitly notes the measure was not final. His reading and claims about its potential legal effects are his opinion, not a statement of current law.

Rather than regulate immediately, Lebrun would have waited a year or two to learn from practical uses and involve model builders, users and the public. He says regulation and attention to real risks are needed, but argues that timing and contact with the field matter.

On China, he speculates that access to large datasets could be an advantage while a more closed environment and research incentives could create disadvantages. His statements about access to Chinese health data are conjecture. He sees the US as well positioned but considers immigration restrictions a potential obstacle to attracting distributed talent.

In quick-fire questions, Lebrun predicts that AI will enable new entrants to challenge existing businesses and gives a five-year horizon. He also wants greater diversity in AI research. These forecasts are his opinions at the time, not evidence that displacement will occur on that schedule.

He recalls a Facebook image model that appeared to recognize tennis balls but relied on surrounding context; changing features of people in the image could alter its classification. He credits a Senegalese researcher with identifying the problem and argues that different perspectives can help uncover model bias. The example is his recollection.

Lebrun expects AI to disrupt the generation of media content more than investigative journalism. On a personal level, he says people can grow and correct weaknesses, but he has learned not to assume their fundamental characteristics will change. Both are judgments, not universal rules.

He considers perceived consciousness a common misconception: a fluent chatbot answer can lead people to infer understanding or sentience. He invokes ELIZA, the early pattern-matching chatbot, as an example of how presentation can create that impression. This is his interpretation of the history and the technology.

Asked what makes a good venture investor, Lebrun says it depends on the founder and stage. At Wit.ai, he valued Andreessen Horowitz providing fast help on request while leaving him in control. Earlier in his career he would have welcomed more proactive coaching on strategic and go-to-market mistakes.

He says European venture capital has changed and contrasts the financial backgrounds of investors he encountered there with the entrepreneurial backgrounds of many US investors. That comparison describes his experience, not every investor in either region.

Looking toward 2033, Lebrun hopes every physician will have an assistant and imagines AI helping allocate emergency resources at a city scale. His suggestion that doctors could become ten times more efficient is an aspiration and forecast, not a measured result.

His own dream is to build a healthcare system from scratch, encompassing the patient experience, providers, hospitals and data-driven prevention. He says he still needs to learn much more about healthcare and would need substantial funding. Stebbings thanks him and the conversation closes.