20VC Why No Models Today Will Be Used in a Year, Why Open Will Always Beat Closed in AI, Why Proprietary Data is Less Important Than Ever And Why EU AI Regulation is a Disaster with Alex Lebrun, Founder & CEO @ Nabla

In a dynamic conversation with Harry Stebbings on "20vc," Alex Lebrac, co-founder and CEO of Nabla, delves into the AI landscape, sharing insights from his extensive experience, including leading engineering at Facebook AI Research and founding AI-driven companies like Wit AI and Virtuos. Lebrac critiques new European AI regulations, deeming them impractical and potentially stifling innovation by making current AI training methods illegal. He emphasizes the importance of timing and involvement of practitioners in regulation. Lebrac also envisions AI transforming healthcare by providing AI assistants to clinicians, improving efficiency, and patient care. However, he notes the challenges in healthcare startups, particularly around who pays for innovation. Lebrac predicts that in ten years, AI will play a significant role in high-level decision-making in healthcare and expresses his ambition to build a comprehensive, data-driven healthcare system.

• The new regulation in Europe is seen as a disaster by Speaker A.
• 100% of the Language Model Systems (LMS) trained in the last three years would be illegal under this regulation.
• Speaker A believes that the regulation's intent is good, but the practical limitations it imposes on training and operating models make current practices illegal.

"So the new regulation is a disaster. It means 100% of the LMS that were trained these last three years would be illegal in Europe."

The quote from Speaker A emphasizes the perceived negative impact of new European regulations on the legality of recently trained language model systems, highlighting a disconnect between regulatory intent and practical implications for AI development.

• Harry Stebings introduces the episode and guest Alice Labrac, cofounder and CEO at Nabla.
• Alice's background includes leading engineering at Facebook AI Research and founding Wit AI, which was acquired by Facebook.
• Previous companies founded by Alice include Virtuos, a pioneer in customer service chatbots, acquired by Nuance Communications.
• A thank you is given to Julianne Cordonur for making the episode possible.

"Welcome back to 20 vc with me, Harry Stebings, and today we continue the deep dive into the world of AI."

Harry Stebings opens the podcast, setting the stage for a conversation about AI advancements and introducing his guest, Alice Labrac, who has a significant background in AI and startups.

• The episode is sponsored by Tigas, Secureframe, and Deal.
• Tigas provides research resources for investors, including transcripts and financial data.
• Secureframe offers a platform for security and privacy compliance, aiming to help companies become audit-ready quickly.
• Deal is an HR tool designed to manage global teams, offering automation and integration for various HR functions.

"But before we dive into the show today, this episode is brought to you by Tigas, the Goto research destination for bold investing."

Speaker C introduces the sponsors of the episode, detailing their services and how they contribute to the fields of investment research, security and privacy compliance, and global HR management.

• Alice fell in love with a chatbot named Sibel 22 years ago, which inspired him to dedicate his life to building chatbots.
• He founded his first company focused on customer service bots early in his career.
• Alice reflects on the challenges and the evolution of AI over the years, including a humorous anecdote about a bot's misunderstanding.

"So, 22 years ago, I fell in love with a chatbot. Her name was Sibel."

Alice shares the initial inspiration behind his career in AI, recounting his fascination with chatbots and how it led to his first company.

• Alice initially thought chatbots would quickly become intelligent and replace humans in call centers.
• An early deployment with the French railway company revealed the limitations and challenges in AI.
• Alice recognizes the continuous development of AI, but notes that public perception tends to view progress as more sporadic and significant.

"When we started, I thought chatbots would become very, very intelligent after three, four years."

This quote reflects Alice's early expectations about the rapid advancement of AI and chatbots, and his subsequent realization of the complexities involved in achieving such intelligence.

• Alice founded Nabla after feeling a sense of complacency at Facebook and wanting to apply what he learned to the real world.
• The decision to leave Facebook was sparked by a moment of self-reflection, comparing his situation to a scene from the show Silicon Valley.

"I was sitting at Terrace at Facebook headquarters in Menlo park."

Alice describes the moment that led to the founding of Nabla, influenced by his experiences at Facebook and a desire to return to the entrepreneurial arena.

• Alice was impressed by Facebook's efficiency and speed, even with a large number of employees.
• He learned about the power of having resources and agility at a company like Facebook.
• Working with Mark Zuckerberg, Alice observed his preparedness for meetings and quick decision-making.

"So first thing that amazed me when I arrived at Facebook is I thought all big companies were slow."

Alice shares his initial surprise at Facebook's efficiency and the lessons he learned about operating effectively within a large, fast-moving organization.

• Alice admired Mark Zuckerberg's efficient use of time and preparation for meetings.
• Zuckerberg's approach to meetings involved challenging ideas to understand them better, which Alice found very insightful.

"One thing I like, he spent not a lot of time outside the company."

Alice discusses Zuckerberg's focused and data-driven approach to meetings, which he found to be a valuable lesson in leadership and decision-making.

• Alice notes that starting a new company does not get easier with experience.
• He discusses the "Kim Jong-un entourage trap," where being unchallenged can lead to mistakes.
• Alice emphasizes the importance of surrounding oneself with people who provide honest feedback and challenge decisions.

"It's not getting easier. You learn some lessons, you don't do the same mistakes again, hopefully, but get new dangers, new issues."

Alice reflects on the continuous challenges of entrepreneurship, highlighting that while some lessons are learned, new obstacles always arise.

• Alice combats the entourage trap by seeking external advice from successful individuals with no incentive to please him.
• Internally, he encourages his team to challenge decisions, which has led to a more balanced level of scrutiny within the company.

"The solution first is, I think, is with external people, even my investors, who are too close to me."

Alice outlines his strategy for avoiding the entourage trap, which involves seeking candid feedback from external sources and fostering a culture of challenge within his team.

• Chat GPT has generated significant consumer excitement around AI.
• Alice believes that advancements in AI are continuous, but public perception often sees them as sudden leaps.
• He discusses the evolution of AI models like GPT-3 and their gradual improvements leading to the current state of technology.

"For the general public, it looks like a very big step function with huge advancements every ten years."

Alice comments on the discrepancy between the AI community's view of gradual progress and the public's perception of sudden breakthroughs in AI technology.

• Alice finds the VC hype cycle around AI to be somewhat ridiculous, given his long-term perspective on AI development.
• He notes the cyclical nature of AI hype and the strategic timing of when to highlight AI in company presentations.

"From our standpoint, it's really ridiculous."

The quote captures Alice's view on the cyclical and often exaggerated excitement that venture capitalists exhibit towards AI, based on his extensive experience in the field.

• The criticism of generative AI as a thin layer over foundational models is disputed by Alice.
• He compares the emergence of generative AI to the introduction of the C programming language and databases, emphasizing the potential for substantial applications.
• Alice discusses the complexity and challenges of working with large language models (LLMs) and the competitive advantage of understanding them deeply.

"Today, I don't think it's fair."

Alice argues against the notion that generative AI applications are merely superficial layers on top of foundational models, suggesting that there is significant value and complexity in developing these applications.## LLM Output and Hallucination

• LLMs are designed to generate outputs even when there's nothing substantive to say, leading to potential "hallucinations."
• Emad's reference to design likely pertains to this unavoidable output generation.

"I mean, by construction, the LLM has to output something, and so if there is nothing to say, it'll make something up that looks natural. So this is probably what Emad meant by design."

This quote explains that the nature of LLMs is to produce content regardless of the input's substance, which can result in fabricated or nonsensical responses. This is a fundamental aspect of how LLMs are constructed and relates to their design as per Emad's comments.

• Agreement that current models will not be in use in a year due to rapid technological progress.
• The analogy with car usage over ten years illustrates the expected obsolescence of today's models.

"Absolutely agree with that. Will you drive the car you drive today in ten years? I don't think so."

This quote emphasizes the speaker's agreement with the idea that the rapid pace of progress in LLMs will render current models obsolete within a year, similar to how cars are replaced over a decade.

• Proprietary data was crucial in the past but is becoming less important due to advancements in model training.
• Nabla's creation of a medical consultation dataset with 30,000 entries highlights the initial need for proprietary data to bootstrap models.
• Pretrained models and fine-tuning now allow for efficient use of smaller datasets.

"I think this is always one train late. So having a lot of proprietary data was very, very important for the last cycle, five years ago, maybe it's less and less true."

This quote suggests that while proprietary data was once essential for AI development, its importance is diminishing as new methods allow for effective training with less data.

• Large language models (LLMs) undergo unsupervised pretraining followed by fine-tuning.
• Pretraining involves predicting the next word in a text without human intervention.
• Fine-tuning uses machine learning techniques to train models to follow specific instructions.

"So a large language model is trained in two phases. First phase is unsupervised pretraining... And then the second phase is fine tuning..."

This quote describes the two-stage training process for LLMs, with the initial phase focusing on general language comprehension and the second phase refining the model's responses to specific tasks or instructions.

• New research (Lima models) demonstrates efficient fine-tuning with only 1,000 examples, performing nearly as well as larger models like GPT-3 and GPT-4.
• This suggests that a small amount of high-quality data can significantly impact model performance.

"This lima paper shows that with only 1000 question and answer examples, they get something that performs better than GPT-3 and almost at the level of GPT-4 with only 1000 QA for fine tuning."

The quote highlights a recent study indicating that high-quality fine-tuning can be achieved with considerably smaller datasets, challenging the notion that vast amounts of data are always necessary for effective model training.

• Most foundational model companies likely already exist, but new ones could still emerge with the right resources.
• Success in creating a foundational model company requires significant scientific expertise and financial investment.

"We probably have most of the existing companies, maybe a few will be started."

This quote acknowledges the possibility of new companies entering the foundational model space, though it suggests that many current players are already established.

• While incumbents have distribution advantages, they also face challenges such as slower innovation and a tendency to enhance existing products with AI features rather than creating disruptive technologies.
• Startups may have the agility to introduce novel AI-driven paradigms.

"So incumbents have a huge advantage through distribution... But I think the best incumbents can benefit from AI to be competitive in their existing markets, but I don't think they will invent totally disruptive things."

The quote points out that established companies can leverage AI to improve their market position, but they are unlikely to pioneer completely disruptive AI applications, which could be an opportunity for startups.

• The speaker predicts open foundational models will prevail, but cautions that even open models can be unpredictable and not fully understandable.
• The term "open" does not necessarily equate to transparency or explainability in the context of complex AI models.

"So obviously I think the foundational model that will win will be open."

This quote reflects the speaker's belief in the dominance of open AI models, although they note that "open" does not guarantee complete clarity into the model's functioning.

• The speaker is skeptical about the creation of national data sets, suggesting industry-specific data sets are more relevant for improving model quality.
• Even with curated data, LLM outputs cannot be assumed to be reliable due to the probabilistic nature of the models.

"I'm not sure about national data sets... But even if you feed an LLM with curated data, it's not guaranteed that the output will be good, will be perfect, that you can trust the output."

This quote expresses doubt regarding the utility of national data sets for AI and emphasizes that curated inputs do not ensure accurate or trustworthy outputs from LLMs.

• The speaker agrees with Jan that there is no inherent correlation between AI intelligence and a desire for dominance.
• Skepticism is expressed towards the idea that more intelligent AI would pose a threat to humanity.

"I fully agree with Janin. 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."

The speaker concurs with Jan's dismissal of the notion that AI's increasing intelligence would lead to a desire to dominate, attributing such claims to a lack of understanding or a desire for attention.

• The speaker finds Elon Musk's call for a pause in AI development contradictory to his actions in attempting to build a competing team.
• It is suggested that those calling for regulation may be trying to maintain their competitive edge.

"It's weird. At the same time, he said that I know he was trying to build a team to compete with OpenAI."

This quote highlights the perceived inconsistency between public statements advocating for a halt in AI development and private actions that suggest a desire to advance in the field.

• The speaker believes we are now at a significant moment for AI adoption, with tools like Chat GPT changing how jobs are performed.
• The potential impact of AI is considerable, although it may not be ready to fully replace human roles.

"Finally, it's here. When I see people using Chat GPT and learning to do their job differently with the help of Chat GPT, I really feel we are on the verge of finally having a huge impact with chatbots."

The quote expresses the speaker's view that we are on the cusp of a transformative period in AI adoption, particularly with the integration of tools like Chat GPT into various professions.

• AI will not replace doctors but will enhance the capabilities of those who adopt it.
• Doctors who use AI effectively will have an advantage over those who do not.

"AI will not replace doctors, but doctors who use AI would replace doctors who don't."

This quote suggests that AI's role in healthcare will be augmentative rather than substitutive, providing a competitive edge to doctors who leverage AI technologies.## Burnout Symptoms and Administrative Work

• Physicians are experiencing burnout symptoms due to the pressure of administrative tasks.
• Clinical documentation is the primary administrative burden for doctors.
• Electronic Health Records (EHRs) are outdated and inefficient, requiring excessive input from clinicians.
• Insurance reimbursement and legal protection are the main reasons for extensive documentation.
• EHRs are described as "dinosaur systems" with a specific clinical action requiring hundreds of clicks.

The biggest work they have to do is clinical documentation.

This quote highlights the primary source of administrative workload for healthcare providers, emphasizing the time-consuming nature of clinical documentation.

You need to have a very solid file with all that. Otherwise the insurance will take the first opportunity, the first pretext not to pay your claim.

This quote explains the financial necessity of thorough documentation for insurance reimbursement purposes in the U.S. healthcare system.

Your best protection is to document this in the right way.

This quote underscores the importance of proper documentation as a legal safeguard against medical malpractice lawsuits.

• AI is expected to provide an assistant for every clinician, improving efficiency in documentation and patient care.
• AI assistants would capture audio during patient visits and integrate with EHRs without storing sensitive data.
• The new law in the U.S. mandates EHR interoperability, despite opposition from major companies.

It will bring an AI assistant to every doctor, to every clinician.

This quote describes the envisioned role of AI in healthcare, where it serves as a support tool for clinicians, streamlining their workflow.

We don't store it, we just capture it and then drop it.

The quote clarifies that while AI assistants would capture audio for context, they would not retain this data, addressing privacy concerns.

They have an API, it's easy to integrate.

This quote indicates that despite their outdated nature, modern EHR systems are designed to allow integration with other applications, like AI tools, via APIs.

• There is a global shortage of clinicians, with an estimated 18 million needed by 2030.
• AI can alleviate the workload on healthcare providers without causing unemployment.
• Efficiency improvements in healthcare may face political resistance due to potential job losses.

We are missing 18 million clinicians by 2030.

This quote from the World Health Organization emphasizes the dire shortage of healthcare workers, suggesting that AI could help bridge this gap.

The health systems are collapsing everywhere.

The quote reflects the global crisis in healthcare, with systems under strain and the need for innovative solutions like AI becoming more urgent.

• Startups often face challenges by moving too quickly towards patient-facing solutions.
• Healthcare systems are complex, and regulatory hurdles can impede rapid innovation.
• Focusing on clinicians first is a more viable strategy for integrating AI into healthcare.

We are trying to go too fast to a patient facing perfect system.

This quote identifies a common mistake made by startups, which is attempting to revolutionize patient care without first addressing the needs and processes of clinicians.

• Emergency call centers function like a triage system, sorting through calls to determine the appropriate response.
• AI could assist in documenting calls, but human empathy and decision-making are crucial.
• The potential for AI in emergency services is significant, but government funding is a complex and challenging go-to-market strategy.

It would be so easy to do with what we do at Nabla and what we know how to do with AI today.

This quote suggests that current AI technology is capable of significantly improving the documentation process in emergency services.

• Healthcare startups must navigate a complex system of payers, providers, and patients.
• The party that benefits from a product is often not the one who pays for it.
• Startups must find the right balance between fitting into the existing system and being disruptive.

Who is paying for the product?

This quote emphasizes the importance of understanding the payment structure in healthcare, which is critical for a startup's success and scalability.

It's impossible to go to market with something too ambitious.

The quote advises against overly ambitious projects that do not align with the current healthcare system's structure, as they are likely to fail in the market.

• The talent for AI development is becoming more distributed globally.
• Proximity to customers is more important than being located in traditional tech hubs like Silicon Valley.
• France has a strong education system that produces skilled AI engineers, but French startups often sell too early.

It used to be true, I think that you needed to be in the valet with having these kind of people around you. Now I think it's less true because everything is distributed.

This quote reflects the changing landscape of AI talent distribution, suggesting that location is less critical than it once was for success in AI.

We produce lots of good engineers, but we

This incomplete quote hints at France's ability to produce skilled engineers, but suggests that there may be challenges in leveraging this talent to its full potential within the startup ecosystem.## Company Growth Challenges

• Companies often struggle with scaling and may sell too early due to attractive offers.
• Founders find it difficult to refuse large sums of money, especially if it is significantly more than they have previously dealt with.
• There is a lack of discipline and focus among some founders, as well as a shortage of role models who have grown companies to a large scale.
• The environment in Europe is changing with the emergence of successful scale-ups.

"We are not bad at that. I don't know why. Maybe we lack discipline. Maybe we." "So if you are normal, it's hard to refuse a $1 billion offer if you haven't done 10 million before." "Hopefully it's changing. We have very good scale ups in France now, but it takes some time too."

These quotes highlight the speaker's self-reflection on why they may not excel at growing companies, the temptation to accept large buyout offers, and the evolving startup ecosystem in Europe.

• Europe is lagging behind the US and China in AI development by approximately ten years.
• New regulations in Europe could further delay progress in AI by making current AI models illegal due to data accountability requirements.
• European startups may need to relocate or find alternative ways to train their models to navigate the regulatory environment.

"So Europe is probably ten years relates compared to us and China." "So the new regulation is a disaster." "If the regulation, if it's really like this and nobody can challenge it, startups like us may have to move."

These quotes express the speaker's concern about Europe's position in AI development and the potentially detrimental impact of new regulations on AI startups.

• Good regulation requires appropriate timing and the involvement of individuals actively working in the field.
• Premature regulation can stifle innovation, especially when the actual risks are not yet fully understood.
• A balanced approach to AI regulation would involve waiting to observe real-world applications and risks before implementing comprehensive laws.

"Good regulation is about good timing and involving the right people who are actually doing stuff in this domain." "First, I wait a little bit, because it's too early."

The speaker suggests that regulators should wait and involve practitioners in the field before enacting AI regulations, emphasizing the importance of timing and real-world experience.

• China's lack of stringent data regulation, like GDPR, gives it an advantage in AI development.
• The abundance and quality of data available in China, particularly in healthcare, provides a significant edge for AI research and application.
• Government support in China may facilitate access to vast amounts of data for AI development.

"They have a key advantage that there is no GDPR or very few regulation internally." "They probably have everything about every individual in China."

These quotes underline China's competitive edge in AI due to less restrictive data regulations and extensive data availability.

• China has advantages in AI but also faces challenges due to increasing isolation and potentially misaligned research incentives.
• Europe's strict regulations could hinder its progress in AI, exacerbating its current lag behind other regions.
• The US maintains a strong position in AI, but future immigration policies could impact its ability to attract global talent.

"That will keep Europe behind and we are already behind." "Immigration laws may be something that will be a problem eventually in the US."

The speaker analyzes the factors that could influence the success of different regions in AI, highlighting Europe's regulatory challenges and potential issues with US immigration policy.

• AI will not only enhance existing businesses but also enable new players to disrupt industries.
• Traditional service companies will adapt to include AI in their offerings, potentially displacing incumbents.
• Some AI service companies have failed, suggesting that the future of AI business may not lie solely in service models.

"AI will enable a new generation of players in every industry that will kill the incumbents eventually." "Existing services companies like consulting companies are embracing AI."

These quotes predict the transformative impact of AI on various industries and the potential for AI to create new market leaders.

• The AI community needs more diversity to uncover biases and improve model accuracy.
• Diverse perspectives can lead to the discovery of significant flaws in AI models that may otherwise go unnoticed.
• An anecdote from Facebook highlights how a model's bias was discovered by a researcher with a different cultural background.

"Make it more diverse." "Somebody discovered that the model...didn't learn to find the tennis ball. It learned the context of a tennis ball, like a racket."

The speaker emphasizes the importance of diversity in identifying and correcting biases within AI models, using a real-world example to illustrate the point.

• AI is unlikely to replace the core investigative functions of journalism.
• However, AI may disrupt the content generation and distribution aspects of media.

"I don't think AI will kill investigative journalism." "Now, the generation part at the output probably will be disrupted a lot by AI."

The speaker believes that while AI will change certain aspects of media, it will not eliminate the need for human-driven investigative journalism.

• People have inherent traits that are unlikely to change fundamentally.
• Recognizing and accepting this can prevent unrealistic expectations and potential disappointments.

"That fundamentally people don't really change." "You can help people to grow... But I think you have to accept that some of the fundamental characteristics... won't change."

The speaker reflects on the nature of personal growth and the limits of change, suggesting that fundamental characteristics are stable over time.

• A common misconception is that AI systems are conscious.
• People are often misled by the appearance of intelligence in AI responses, mistaking form for consciousness.

"It's that they think it's conscious." "In 1966, when the mother of all chat bots was released at the MIT Eliza... people around the world thought that AI was solved."

The speaker addresses the misconception of AI consciousness and cites historical examples of people being deceived by the superficial appearance of intelligence in AI systems.

• Beliefs can change over time, and people often forget their previous convictions.
• The speaker shares a personal realization about the disconnect between demonstrating value and securing payment in business.

"It's very hard because when we change belief, we tend to forget that we have the contrary belief before." "If patients love your healthcare product enough and you prove the health benefits of your product, then payers will pay for it."

These quotes discuss the evolution of beliefs and the speaker's personal learning experience regarding value and payment in the healthcare industry.

• The value of a VC is not in their financial expertise but in their entrepreneurial experience.
• European VCs have historically come from finance, while US VCs are often former entrepreneurs.
• The best VCs provide support without being overbearing and offer practical assistance when requested.

"There is nothing about financial industry finance in being a VC a little bit, of course, but it's mostly about entrepreneurship." "But for my first startups, I did some strategic mistakes, go to market mistakes, and then I wish my vcs back then would have been more involved with me and coach me."

The speaker contrasts the backgrounds of European and US venture capitalists and shares personal experiences with VC support during different stages of company growth.

• AI is expected to significantly enhance the efficiency of physicians and healthcare decision-making.
• The speaker aspires to build a comprehensive healthcare system that leverages AI for improved patient care and data-driven prevention.

"Every physician has their AI assistant doing all a lot of stuff for them and helping them to be ten times more efficient." "So my dream is to build a healthcare system from scratch without any of all these limitations and constraints we mentioned."

The speaker outlines a vision for the future of healthcare, emphasizing the role of AI in augmenting medical professionals and optimizing healthcare systems.