Welcome back to The Hidden Layer. I’m Ian Krietzberg, and yes, I did see Nvidia’s announcement that it will invest $100 billion in OpenAI so that OpenAI can… buy more Nvidia chips. We’ll dig into that one on Thursday.

Today, though, I’ll take a close look at an overlooked A.I.-pharmaceutical partnership. The multihundred-million-dollar race to leverage artificial intelligence for drug discovery has overshadowed a potentially more consequential project: using A.I. to accelerate enormously costly, time-consuming clinical trials.

Also mentioned in today’s issue: Tuhin Srivastava, Mark Zuckerberg, Sam Altman, Microsoft, Satya Nadella, Jensen Huang, QuantHealth, Orr Inbar, and many more…

And to anyone celebrating, Shana Tova—may the new year be a good one for you and your loved ones.

Let’s get into it…

Three Things to Know…

Engineers, developers, and vibe coders: As generative A.I. systems proliferate, “vibe coding”—a Merriam-Webster–approved slang term for largely A.I.-driven software development—has become enormously popular. It’s also a major point of concern for many software engineers worried about their career prospects. But vibe coders and professionals, at least so far, aren’t quite the same thing. Vercel, one of the major vibe-coding platforms, recently analyzed the internal data for its 3.3 million users: 63 percent of them aren’t even developers, and 62 percent of them only vibe code outside of working hours.

Meanwhile, Tuhin Srivastava, the C.E.O. and co-founder of Baseten, an A.I. deployment platform, told me it’s wrong to think that software engineers will be entirely outsourced to A.I. “We’re hiring infrastructure engineers who are running production services. There’s no way we’re vibe coding that service,” he said. “People building the models aren’t vibe coding new architectures. People doing kernel optimizations to make a model faster—that’s not being vibe coded.”

At the same time, Srivastava added that front-end development—the coding of less complicated applications—is probably going away. That’s still concerning because it threatens the pipeline for extremely valuable, highly skilled engineers. For now, though, the space remains enormously competitive: Look no further than Mark Zuckerberg’s attempt to poach engineering talent with nine-figure deals.
The American A.I. awareness matrix: Over the past few years, artificial intelligence has seeped into the cultural zeitgeist, thanks largely to ChatGPT. Just three years into the A.I. race, Pew Research Center has found that 95 percent of Americans have heard at least a little about the technology, and nearly half have heard a lot. The survey included other interesting data points: 57 percent of Americans believe they have “not too much or no control” over how A.I. is integrated into their lives, and 61 percent would like more control—a six-point increase since Pew’s 2024 survey. Seventy-three percent are willing to let A.I. assist them at least a little with their daily lives, while only 13 percent seek a lot of assistance.

Perhaps most fascinating, half of respondents are more concerned than excited about the proliferation of A.I.—a marked increase from 2021, when that figure stood at 37 percent. More than half are worried that increased use of A.I. will hamper creative thinking, echoing concerns from many researchers in the field. And nearly three-quarters of those surveyed said it’s extremely or very important that people understand what A.I. is. (Reader, you’re in the right place.) If you have thoughts on these prompts and findings, just respond to this email. Let’s do our own survey.

A MESSAGE FROM OUR SPONSOR

EFFORTLESSLY ADAPTED
Responsive handling. Connected convenience. Enter the next phase of sporting luxury with the Range Rover Sport.
EXPLORE

The paradox of hand-wringing: On Monday, a group of several hundred Nobel laureates, computer scientists, and world leaders signed a rather grave statement noting that A.I. poses potentially “unacceptable” dangers, and “could soon far surpass human capabilities and escalate risks such as engineered pandemics, widespread disinformation, large-scale manipulation of individuals including children, national and international security concerns, mass unemployment, and systematic human rights violations.” The group advised that governments around the world collaborate to establish clear “red lines,” and that enforcement mechanisms be put in place by the end of next year, before our window to act closes.

Of course, this is not the first letter of its kind. In 2023, more than 33,000 people signed a now-infamous letter from the Future of Life Institute, calling for a six-month pause on training A.I. systems more powerful than GPT-4 to establish safety protocols. Obviously, that didn’t happen. (Notably, Elon Musk was a signatory.) In 2024, a similar letter was published by a small group of anonymous former employees of major A.I. firms, which claimed that the technology’s risks “range from the further entrenchment of existing inequalities, to manipulation and misinformation, to the loss of control of autonomous A.I. systems potentially resulting in human extinction.” Light stuff.

But none of these letters have translated into action. If their target audience was global regulators, that group has instead been hard at work pushing infrastructure build-outs and driving nationwide adoption efforts across schools, hospitals, and governments—almost the exact opposite of the letters’ intended effect. Many researchers, meanwhile, view certain elements of this hand-wringing as more industry hype than substance, and note that some of the perceived threats—particularly around that “human extinction” category—lack clear evidence. That said, given that only a few actors are making decisions that are already impacting the rest of the world, a global cooperative effort—however unlikely—is certainly not a bad idea.

Quote of the Week

“It is dangerous to distract ourselves with a fantasized A.I.-enabled utopia or apocalypse which promises either a ‘flourishing’ or ‘potentially catastrophic’ future. Such language that inflates the capabilities of automated systems and anthropomorphizes them … deceives people into thinking that there is a sentient being behind the synthetic media. This not only lures people into uncritically trusting the outputs of systems like ChatGPT, but also misattributes agency.”
—Researchers Timnit Gebru, Emily M. Bender,and Angelina McMillan-Major in 2023, writing in response to the publication of the Future of Life Institute’s “A.I. Pause” letter

And now for the main event…

Your A.I. on Drugs

Investors are shoveling hundreds of millions of dollars into the quixotic promise of A.I.-assisted drug discovery, but a handful of tech companies are pursuing a parallel, potentially more impactful endeavor: leveraging artificial intelligence to speed up enormously costly, time-consuming clinical trials.

Ian Krietzberg

It’s still unclear how rapidly A.I. technology will develop, but at least some of the breathless hype of the earliest years has calmed as the tech’s near-term limitations have become more clear. The notion that artificial intelligence would immediately help us cure cancer, for instance, has been refined to the (still exciting) expectation that the technology can help drive major advancements in drug discovery. In 2024, Nvidia chief Jensen Huang declared that we are on the precipice of “computer-aided drug design.” Microsoft C.E.O. Satya Nadella has similarly talked about how the company’s latest A.I. models will advance the effort.

Surrounding all of these pronouncements, of course, is an environment of investments, partnerships, and deals between A.I.-focused biopharmaceutical companies and many of the major A.I. players. Formation Bio, which is backed by Sam Altman, raised some $372 million in 2024 from the usual Silicon Valley V.C. suspects; Retro Biosciences, also backed by Altman, was in talks to complete a $1 billion funding round earlier this year; and Google spinoff Isomorphic Labs completed a $600 million round in March. The list is vast, growing, and overflowing with zeros.

And yet, while A.I.’s role in drug discovery is indeed promising, it’s by no means a silver bullet. The reality is that the drug development pipeline takes, on average, between 10 and 15 years. The molecule selection process is time-consuming, but much of that time is spent gathering data to convince regulators a given drug is safe. That’s why a handful of tech companies, like the Israeli firm QuantHealth, have built their businesses around using A.I. to speed up those costly, time-consuming clinical trials—an overlooked race that’s quietly running in parallel with efforts to advance A.I.-assisted drug discovery.

Orr Inbar, the C.E.O. and co-founder of QuantHealth, told me that clinical trials “account for a larger budget than drug discovery by a wide margin.” Yet he said a “herd mentality” has consumed investors, who have largely neglected the promise and challenges of clinical trial optimization. “This problem that QuantHealth solves is really, really hard to solve, and it just so happens that this space is actually not a niche by any stretch of the imagination,” he told me. “It’s a $100 billion-plus market that’s received very little solutions.”

A MESSAGE FROM OUR SPONSOR

EFFORTLESSLY ADAPTED
Responsive handling. Connected convenience. Enter the next phase of sporting luxury with the Range Rover Sport.
EXPLORE

Indeed, a 2021 Congressional Budget Office report found that an outsize portion of larger drug companies’ R&D budgets is “devoted to conducting clinical trials.” A 2020 study, meanwhile, found that the median cost of developing a single drug, without adjusting for the costs of failed trials, is around $319 million; when you include the price of failed trials, the median price tag is upward of $1.1 billion. “Pharma has always wanted to do this. This is the holy grail for them,” Inbar claimed. “It’s just really hard. It’s taken us $30 million, five years, and 50 people to build this. You have to be a little obsessive to tackle a problem like this. Luckily for me, I have my obsessions.” Eli Lilly and Pfizer—both of which are listed as partners on QuantHealth’s website—didn’t return requests for comment regarding how they use QuantHealth’s technology and what effect it has had.

“It’s a Lift”

Inbar, who co-founded QuantHealth in 2020, said he started developing the technology as a graduate student. Several years ago, he was conducting research on medical records and found it was possible to build a model capable of predicting how a patient would respond to a preexisting drug. But the real challenge, in his view, was predicting how a patient would respond to a novel drug. He started looking beyond clinical and pharmacological data, and found that “there’s a huge amount of data that does, in fact, characterize the biology of almost all the drugs that pharma would like to develop. It’s just scattered and highly disorganized.”

When he founded QuantHealth, the company set about assembling—and licensing—those data sources, while building a series of algorithms designed to stitch them together. When applied to simulated clinical trials, Inbar said, the resulting A.I. system can predict “how any patient responds to any therapy,” which lets R&D teams at pharma companies adjust the design of their clinical trials to avoid costly failures. In one example the company shared, the suggestions of a QuantHealth simulation apparently led a major pharmaceutical company to refine its patient population, resulting in $31 million in savings.

QuantHealth operates in a space without any off-the-shelf model to iterate on, so the company has to build everything from scratch. That means proprietary “biological foundation models” trained with data from more than 100 million patients on average. These models, which Inbar said are much more complex than L.L.M.s, are also transformer-based deep learning systems, but they are designed with more classical statistical approaches in mind. Inbar said that QuantHealth develops one of these foundation models for each therapeutic area it’s targeting—such as oncology or immunology—and trains each one roughly twice a year. “It’s a lift,” he said.

Last year, QuantHealth closed a $17 million Series A round led by Accenture, among others, at an undisclosed valuation. As with all A.I. companies, the cost of talent—necessary for data collection, curation, and model-building—is a major source of QuantHealth’s expenses, in addition to the cost of compute. Of course, further investment will likely depend on whether QuantHealth can address some of the early skepticism surrounding the platform. This has involved back-testing on previously completed clinical trials, and testing on older trials on the verge of completion. The company, Inbar said, has completed over 100 tests with the first method and more than 35 with the second, achieving an 85 percent accuracy rate on both. “Pharma has to make these multidimensional decisions that carry enormous risk and R.O.I. for them, but they’re also really, really hard to quantify,” Inbar told me. “So we kind of step in there right at the moment of trial design when those key questions are being asked, and help de-risk the overall process.”

What I’m Reading…

Last week, Italy became the first country in the E.U. to pass its own law regulating artificial intelligence. The legislation established rules for copyright, child welfare (kids under 14 will need parental consent to access an A.I. system), and general misuse—with prison terms for people who spread A.I.-generated content that causes harm. [The Guardian]

Last year, Mira Murati, OpenAI’s C.T.O. at the time, was asked by The Wall Street Journal what content the company used to train Sora, its video generation model. Her response—a pause, a grimace, and a “not sure”—became an instant meme. The Washington Post recently conducted a series of tests in which Sora generated clips that closely resemble copyrighted I.P. from sources including Disney and Universal—which, per the paper, “suggests a version of the originals appeared in the tool’s training data.” [Washington Post]

I broke down my recent reporting on A.I. therapy on the Beyond Well podcast. Give it a listen! [A.I. Chatbots and Mental Health]

That’s all for today. I’ll see you Thursday.

Ian

Puck sports correspondent John Ourand and a rotating cast of industry insiders take you inside the executive suites and owners boxes where the decisions that shape the entire sports business are made. You’ll hear interviews with players, network execs, and everyone in between. The Varsity is an extension of John’s private email for Puck by the same name. New episodes publish every Wednesday and Sunday.

Unique and privileged insight into the private conversations taking place inside boardrooms and corner offices up and down Wall Street, relayed by best-selling author, journalist, and former M&A senior banker William D. Cohan.

Iger’s Boogeyman

KIM MASTERS

Fashion’s October Anxieties

LAUREN SHERMAN

Jeffries’ Pressure Cooker

ABBY LIVINGSTON

Need help? Review our FAQ page or contact us for assistance. For brand partnerships, email ads@puck.news.

You received this email because you signed up to receive emails from Puck, or as part of your Puck account associated with {{customer.email}}. To stop receiving this newsletter and/or manage all your email preferences, click here.

Puck is published by Heat Media LLC. 107 Greenwich St, New York, NY 10006