“We’re not doing 30 bets a year”: Vijay Pande on betting small after running $4 billion at a16z

It utilized to beryllium that Vijay Pande was amended known successful world circles than capitalist circles. That changed beauteous abruptly a twelve years ago, erstwhile Marc Andreessen and Ben Horowitz — who’d spent their firm’s archetypal 5 years explicitly avoiding healthcare and beingness sciences — decided the class was worthy betting connected aft each and handed the keys to Pande. At the time, helium was a Stanford chemistry prof who was champion known for gathering Folding@home, the distributed-computing task that turned millions of location PCs into a supercomputer for illness research. Over the adjacent decade-plus, helium grew a16z’s stake into a signifier managing adjacent to $4 billion.

So it was somewhat unexpected erstwhile successful June of past year, Pande walked distant from it each to commencement thing overmuch smaller. In fact, his caller firm, VZVC, co-founded with longtime capitalist Zach Werner, is built astir a fistful of concentrated bets a twelvemonth alternatively than dozens, it has nary associates, and it relies heavy connected AI for its day-to-day operations.

To larn much astir Pande’s hard pivot, we talked with him this week astir wherefore he’s making conscionable a fistful of concentrated bets alternatively than spreading himself bladed successful the existent marketplace — and astir 1 of the much absorbing conundrums successful AI-driven biotech: dissimilar text, biologic information can’t beryllium scraped disconnected the internet, truthful astir each institution ends up gathering its ain walled-off dataset. What does that mean for each the advances AI successful medicine has promised, and who really gets entree to them?

This speech has been edited for magnitude and clarity. You tin besides perceive to the fuller speech (below).

You’ve said biology is moving from a “science of discovery” to thing you tin engineer. What does that mean?

For a batch of the mode drugs person been developed, determination was precise overmuch a fortuitous facet to it. I deliberation what’s shifted is that AI and instrumentality learning let computers to wrapper their benignant of knowing astir thing very, precise complicated… to effort to fig retired what targets you privation your drugs to hit, for circumstantial diseases, to beryllium capable to marque those drugs, and present adjacent to assistance successful the objective trials — which are the astir costly portion of the process.

I thought objective trials were getting cheaper due to the fact that cause developers are utilizing much synthetic data, truthful not arsenic galore radical are needed for these trials.

That’s, I think, precise overmuch an aspiration.

The outgo and clip to get to objective trials has been shrinking, particularly with AI, but it could inactive outgo hundreds of millions of dollars to tally a trial, which is wherefore drugs are precise expensive. The probability of a cause going successfully from the archetypal proceedings to the extremity of the 3rd proceedings is conscionable 20%. If 8 retired of 10 fail, and these things outgo hundreds of millions of dollars, the amortized outgo gets truly high. The crushed they neglect typically is not that the biologist did thing wrong; it’s that each the experiments these drugs were designed connected were connected carnal models similar mice, and successful the end, carnal models are conscionable not precise predictive of humans. The AI exemplary is not going to beryllium perfect, but it’s going to beryllium mode amended than immoderate carnal exemplary would be, and erstwhile it crosses that bar, that’s wherever it gets truly exciting.

[The signifier aft that is]: Is the cause the close cause for me?

You mean personalized medicine. . .

The jargon present is alleged precision medicine. If you spell to a doc with thing not trivial, they person to conjecture what’s going on, due to the fact that there’s lone truthful overmuch they tin tell. Then they springiness you a cause — and if that doesn’t work, they springiness you different drug, past different drug. This happens successful cancer, it happens successful tons of antithetic areas. We would each beryllium overmuch amended disconnected if the archetypal cause was the close one. Typically, your humor trial values are compared to colonisation averages. But really, they should beryllium compared to: is this [result] weird for you? What we’re starting to bash besides connected the medicine broadside is [the ability] to conscionable recognize what would beryllium close for the individual.

Would you accidental the way to this infinitesimal has been dilatory and steady, oregon did it spike much recently?

I think it’s lots of antithetic things [coming together]. So for instance, precision medicine for the longest clip was based connected genomics. But the world is your genome is kind of like the blueprint for your location connected time one, but your location is fairly different now compared with the infinitesimal it was built. So determination are galore different things that radical tin present measurement successful proteomics and truthful connected that are overmuch much applicable for knowing illness and wherever your assemblage is now. There has also been [a batch of] automation successful robotic measurements that is people tied into AI, and those 2 spell manus successful hand really well. 

Over the past decade, there’s been this steady clip fforin some AI for biology and AI for chemistry. The biology portion is like, however tin we dainty this disease? And past the chemistry portion is, however tin we come up with a cause to spell aft that circumstantial protein? There person really been precise significant advances implicit those 10 years. 

You mentioned that biology is 1 of the fewer places AI can’t conscionable scrape information disconnected the internet. What does that mean for however the tract develops?

It’s a spot wherever you don’t person immoderate of this information that radical tin conscionable each bid the aforesaid thing, and your information can’t beryllium distilled from 1 exemplary to another. It’s a truly absorbing play from conscionable the axenic AI sense.

Doesn’t that echo a acquainted occupation successful medicine, though — doctors operating successful [territorial, often competitive] silos?

You’re onto thing truly large here. Let’s accidental [someone] has immoderate benignant of cancer, and it’s some an contented successful oncology and endocrinology — those 2 doctors truly don’t sync unneurotic precise well. What is truly intriguing astir AI is that it can, successful principle, beryllium a specializer successful everything, and it tin commencement to spot things that truly immoderate azygous quality being couldn’t. It would beryllium equivalent to having a squad of the precise champion doctors each clamoring unneurotic successful that moment.

But is determination capable information sharing for that imaginativeness to really beryllium realized? I recognize wherefore founders and investors privation to support their [respective findings], but . . .

I deliberation 1 of the bigger trends is that we’re starting to spot a displacement toward gathering these atlases of biologic accusation — which, from a exertion standpoint, are typically instauration models. And arsenic they go much common, I deliberation we’ll spot the aforesaid happening that’s happened with open-source LLMs, which bash precise good against the firm ones: open-source instauration models successful biology having a precise wide impact.

You’re progressive with Genesis Therapeutics, which came retired of your laboratory astatine Stanford, and Insitro, the drug-discovery institution launched by Daphne Koller, a erstwhile workfellow astatine Stanford. You accidental you’re besides incubating a institution with a laminitis you’ve known for 20 years. What are you looking for successful founders, and successful what areas?

There are 2 areas that I’ve been spending astir of my clip on. One is AI for healthcare delivery, which I did a ton astatine a16z arsenic well, and past AI for objective trials.

One of the things that’s astir important to maine [about founders] is that we tin truly spot each different — founders that person precocious integrity, that bash what they accidental they’re gonna do… I’m expecting this narration to beryllium 5, 10 years positive into, ideally, their adjacent company. I privation to enactment with radical who are reasoning agelong word similar that. Ideally, these are radical who are not conscionable trying to triumph and bushed different people, but truly reasoning astir the question: however bash we triumph together?

What person you gotten close and incorrect successful your investing vocation truthful far?

When I started talking astir AI and instrumentality learning and exertion and medicine and bio 10 positive years ago, determination was a batch of absorption and a batch of radical saying, ‘Oh, that’s ne'er going to happen. That’s never going to beryllium useful,’ and truthful on. That absorption is mostly gone and seeing this arc is precise fulfilling.

I deliberation it took maine immoderate clip to truly admit that arsenic seductive arsenic the coolest technologies are, it truly ever comes backmost to go-to-market. I archer my founders, particularly the ones who are coming from the subject oregon the merchandise side, for them to instrumentality each their brilliance and creativity and truly use it to the go-to-market side, that the go-to-market portion is astatine slightest arsenic hard oregon harder than the exertion side.

Help america recognize however you’re designing this caller steadfast differently, compared with what you were moving astatine a16z.

Right now, we’re doing thing truly rather different… VZ is named aft me, Vijay, and my co-founder, Zach Werner — he’s the Z. We’re intentionally truly rather small… connected the concern side, it’s truly conscionable the 2 of us. We were really intending connected hiring associates, but it turned out, with the agents that we’ve built up, not to beryllium thing that we request to do.

How concentrated is “concentrated”?

We’re [not] driving 30 bets per year… we’re talking astir about apt five, not a batch of investments — precise concentrated. Adding a institution astatine a emblematic money is similar adding a Facebook person — that’s thing you bash beauteous quickly. For Zach and I, it’s much similar . . . wanting to person different child. This is simply a large woody for us.

With that structure, who are you competing against for deals?

The comic happening astir this exemplary is that typically we’re not trying to vie for a blistery circular — radical marque country for us. It’s a precise antithetic happening than trying to get the blistery Series A oregon Series B. Largely, radical privation america arsenic investors due to the fact that of what Zach and I tin do, and however hands-on we tin be. When I look astatine radical who are inspirations, I look astatine idiosyncratic similar Antonio Gracias astatine Valor — he’s well-known present due to the fact that of the SpaceX deal, but he’s been doing what he’s been doing for 20 years. What Thrive has done, with a much concentrated portfolio, is besides a existent inspiration. Obviously, a16z is benignant of successful my DNA arsenic well, but I deliberation those different ones are caller additions to however we deliberation astir things.

What’s overhyped close present successful AI and biotech?

The world is that AI tin find insights that we can’t get from conscionable humans alone. The happening that ever gets tricky is erstwhile there’s this telephone that AI is going to cure each everything. The crushed for hesitance determination is not due to the fact that of immoderate uncertainty astir AI — it’s astir uncertainty of the data. LLMs enactment due to the fact that there’s truthful overmuch information to larn from. When the information is conscionable simply not there, past AI can’t magically lick that problem.