Jan 2, 2019 · 41m · a16z
a16z Podcast | Breaking Into Bio
Daphne Koller · 19m spoken Atul Butte · 12m spoken Vijay Pande · 4m spoken Sonal Chokshi · 38s spoken
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
The a16z Bio podcast episode features a panel discussion with Dr. Atul Butte and Dr. Daphne Koller, moderated by Vijay Pande, exploring how researchers and technologists can successfully transition from academia to build high-impact biological and healthcare startups. The panelists share strategic insights on navigating complex regulatory ecosystems, selecting complementary co-founders, applying pragmatic machine learning, and bridging the cultural gap between computer science and medicine.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The host as informed peer 3.8 Guest teaching 4.9 Guest disagreement 1.3 The host pushing back 1.8
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 15:30 Daphne rejects high-risk framing of bio startups
Daphne explicitly dissents from Vijay's premise regarding career risk, arguing that computer scientists face minimal actual risk because Machine Learning skills remain highly durable.
Hardest push from the host ▶ 22:17 Vijay challenges soft-selling adviceVijay directly refuses Daphne's recommendation to soft-sell pitch goals, pointing out that founders lacking established academic brands cannot afford to under-promise in competitive markets.
Biggest teaching moment ▶ 17:30 Daphne details ML data regimes in biologyDaphne educates the host and audience on the historical plateau of neural networks and explains why biology datasets currently require explicit structural modeling rather than out-of-the-box ML architectures.
The host holds their own ▶ 38:46 Vijay explains failure modes of separate domain teamsVijay demonstrates sharp operational expertise by detailing why pairing separate top-tier computer scientists and biologists fails due to language barriers, arguing strongly for hiring hybrid talent instead.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The host as informed peer | Guest teaching | Guest disagreement | The host pushing back | Why |
|---|---|---|---|---|---|---|
| Transitioning Academic Research into Bio Startups | 3 | 5 | 1 | 1 | Vijay frames the challenge of bridging academic tech to real products and agrees on healthcare's antiquated state. Atul and Daphne educate on deep domain pain points and non-technical adoption barriers. | |
| Go-to-Market Challenges and Healthcare Ecosystem Adoption | 3 | 4 | 1 | 1 | Vijay agrees with guests that claims about AI replacing doctors are naive and stupid. Atul and Daphne detail regulatory realities like HIPAA/IRB and go-to-market strategies. | |
| Finding and Evaluating the Right Co-Founder | 4 | 4 | 2 | 2 | Vijay humorously probes Atul's formulaic co-founder criteria by asking what happens if the co-founder applies the same strict criteria back. Daphne lightheartedly notes the game-theoretic equilibrium. | |
| Competition, Fundraiser Strategy, and Founder Lessons | 3 | 5 | 1 | 1 | Vijay prompts guests for unknown knowns in founder strategy. Atul discusses global competition and compares first startups to ruined first pancakes, which Vijay enthusiastically reinforces. | |
| Joining a Bio Startup and Managing Career Risk | 5 | 5 | 3 | 3 | Vijay offers a provocative hypothesis that people attend grad school because they are inherently risk-averse. Daphne explicitly dissents, reframing career risk and emphasizing that ML skills do not go obsolete. | |
| Machine Learning and Data Regimes in Modern Biology | 4 | 6 | 1 | 2 | Vijay playfully asks about ML pre-dating deep learning. Daphne breaks down the historical arc of neural networks, data saturation plateaus, and why current bio data regimes require domain-specific model structures. | |
| Avoiding AI Hype and Building Brand Trust | 6 | 5 | 2 | 7 | Vijay directly pushes back on Daphne's advice to under-promise, arguing that unknown founders without established personal brands cannot afford soft-selling. Atul and Daphne respond by outlining how unknown founders can build trust through rigorous peer-reviewed publications. | |
| Applying Practical Machine Learning in Bio Startups | 5 | 5 | 1 | 1 | Vijay uses a Raiders of the Lost Ark analogy to contrast academic elegance with practical startup execution. Guests agree, advising against Kaggle-style metric chasing in favor of solving real clinical problems with simple models. | |
| Strategies for Computer Scientists Entering Healthcare | 3 | 5 | 1 | 1 | Vijay asks how CS candidates can transition into bio. Guests advocate attending medical Grand Rounds to absorb lingo and adopting Silicon Valley humility when asking questions across fields. | |
| Audience Q&A: Evaluating Execution Quality in Competing Startups | 3 | 6 | 1 | 1 | During audience Q&A, guests explain that execution quality trumps idea quality and outline strategies for securing medical data through proof-of-concept consulting. Vijay reinforces the importance of working with A-plus talent. | |
| Audience Q&A: Identifying Talent Shortages in Bio and ML | 5 | 4 | 1 | 1 | Vijay demonstrates domain expertise by explaining why pairing separate pure CS and pure bio experts fails due to communication friction, favoring single hybrid candidates instead. Daphne strongly agrees with his assessment. | |
| Audience Q&A: US Healthcare Complexity versus International Markets | 2 | 5 | 1 | 1 | Guests address an audience query on international medical markets, highlighting the massive scale of US health data (e.g., 15M UC records) despite regulatory complexity, before Vijay wraps up the panel. |
Statements from this episode (25)
Butte: Skin mole AI apps tackle easy problems, ignoring complex medicine
“Deep learning of moles, cancer. Because that seems like a very intuitive kind of problem, but that's an easy one. We have many, many harder problems in biology medicine, but they take time to learn, so you have to be patient.”
Koller: Healthcare tech adoption depends on workflow integration, not ML complexity
“It's not really about the machine learning inside the box. It's about how do you get it so that the physician doesn't even have to think about how to use your system. It just happens naturally.”
Pande: Healthcare has the hardest go-to-market strategy of any industry
“The go-to-market for healthcare in general is probably the hardest go-to-market.”
Butte: Pitching AI as a doctor replacement immediately alienates healthcare customers
“I think there are other kind of silly things that companies do sometimes that are unnecessary, like say things like physicians are going away, right, with AI and deep learning. Yeah, the greatest way to Make us not want to accept your product, right? That kind…”
Butte: Healthcare startups should target community hospitals before Stanford or UCSF
“There are other hospitals around here, medical systems and practices that are not Stanford and UCSF, and that's the first account you should try to get, not Stanford and UCSF.”
Koller: Tech founders in healthcare need domain co-founders or industry experience
“You really need to either spend serious time in either a hospital or a company, an existing company that actually has that as a market, or you get a co-founder who's had that.”
Butte: The surge in AI talent will cause inevitable startup collisions
“Learn what others are doing, not to kind of steer around, but just be aware, like, what is the pace you're going to have to keep up with, what are the milestones you're going to have to get to, because it is getting, especially in the computer and AI and machi…”
Pande: Many students attend graduate school out of low risk tolerance
“A lot of people go to grad school because they're actually Not very risk tolerant. You know, that grad school is a comfortable thing to do.”
Koller: In 2011, a big biology dataset was a couple hundred samples
“When I started Coursera back in 2011, 2012, a big data set was a couple hundred samples. That was really big, ok?”
Koller: Biological ML requires exploiting domain structure due to dataset limits
“We're still not in the large, large data regime where, you know, blind architectures that don't exploit structure of the problem can just work out of the box. So you really have to understand your problem domain and figure out how to exploit the structure that…”
Koller: General artificial intelligence is not right around the corner
“I think one of the big risks that we run as a machine learning community is the incredible amount of hyperbole that's going on right now, where it's like, we're gonna have general intelligence right around the corner. We're not. Ok, we really aren't.”
Butte: Healthcare branding equals trust and is often ignored by startups
“A lot of companies in this space really ignore the importance of branding. Branding equals trust, ok?”
Koller: Theranos never published peer-reviewed papers or disclosed raw data
“We all know Theranos, you know, that's an extreme example, but, ah, the fact that they never had a peer-reviewed publication, they never presented their data in any way, they kept even potential customers from looking at the raw data, I mean, those are all rea…”
Butte: I forbid my academic lab members from entering ML competitions
“I forbid anyone in my lab to enter those competitions, because that's playing someone else's game.”
Butte: Finding the right clinical question matters more than ML model accuracy
“The hardest part in everything we do is figuring out what is the question to ask? What is the pain point? That you realize this is askable and answerable. This is modelable now, right? Five years ago we didn't have the data, now we do. That's way more importan…”
Koller: AUC-ROC curves rarely measure real-world performance
“The area under the ROC curve is rarely the thing that you actually care about. That was devised for radars back in the fifties, ok?”
Butte: Hospital Grand Rounds reveal unsolved clinical problems to computer scientists
“They have this amazing concept at academic medical centers called Grand Rounds, and they bring in an expert, right, and they talk about a disease, and it's maybe 30 minutes of everything you know about the disease, and the last 10 minutes, this is everything w…”
Koller: Interdisciplinary innovators must risk looking foolish to ask basic questions
“If you're working at the boundary between two disciplines, you need to go in with the confidence that you are an expert in your domain, and it's okay for you to appear like a complete idiot in the other one, because if you're not going to ask those questions, …”
Butte: Healthcare inefficiency is driven by systemic resistance, not friction
“Friction is the wrong model for what needs to get fixed in the healthcare system. It's resistance. Somebody makes every dollar in the 3.2 trillion dollar economy here. They like making that money.”
Butte: Early-stage healthcare startups cannot serve both payers and providers
“Either you're going to go after providers or payers. You're rarely going to do both as a startup.”
Koller: Healthcare startups cannot achieve meaningful success in two to three years
“Going into the space is a long haul game. This is not a game you're going to win in two to three years.”
Koller: Demand for biologists who can code is exceptionally high
“And actually, there's more of them than you might think, but the demand for them is unbelievably high. Every pharma, every academic lab knows that they need people like that. So while there might be more of those than the unicorns that Atul was talking about, …”
Pande: A dual-skilled generalist outperforms separate bio and CS experts
“And maybe I would just add, you know, what happens if you don't have this, and you try to substitute this with an A-plus computer scientist joined with another person who's an A-plus biologist. The problem there is they just do not know what the other one does…”
Butte: University of California holds raw EHR data on 15 million patients
“We have fifteen million at the University of California. We have a total of fifteen million patients that we have raw EHR data on. So that's every drug, every dose, every vital sign. Every lab test result for fifteen million people, so that's about five percen…”
Butte: Small nation healthcare datasets deliver less impact than the US market
“Yeah, you could go to Estonia and Denmark and all these great places that have great data sets, but they're also small, and for impact, they may not have the same impact As it would have in our crazy healthcare non-system here.”