Generative AI, the transformative technology of the moment, exploded onto the scene with the arrival in late 2022 of chatGPT, an AI-powered chatbot developed by the company OpenAI. 

After only five days, a million users had tried the app; after two months: 100 million, the fastest growth ever seen for a consumer application. TikTok, the previous record holder, took nine months to reach a 100 million users; Instagram had taken 2.5 years.

Optimists thrill to the potential AI offers humanity (“Why AI Will Save The World”), while doomers catastrophize (“The Only Way To Deal With The Threat From AI? Shut It Down”). Consultants and bankers offer frameworks and roadmaps and persuade anxious clients they are already behind. Just this week, McKinsey predicted that generative AI could add $4.4 trillion in value to global economy. Morgan Stanley envisions a $6 trillion opportunity in AI as a whole, while Goldman Sachs says 7% of jobs in the U.S. could be replaced by AI.

The one thing everyone — from Ezra Klein at the New York Times to podcasters at the Harvard Business Review — seems to agree on is that generative AI “changes everything.”

But if we’ve learned anything from previous transformative technologies, it’s that at the outset, nobody has any real idea how these technologies will evolve, much less change the world.  When Edison invented the phonograph player, he thought it might be used to record wills. The internet arose from a government effort to enable decentralized communication in case of enemy attack.  

As we start to contemplate – and are thrust into – an uncertain future, we might take a moment to see what we can learn about technology from the past.

***

Yogi Berra, of course, observed that “it is difficult to make predictions, especially about the future,” and forecasts about the evolution of technology bear him out. 

In 1977, Ken Olsen, President of Digital Equipment Corporation (DEC), told attendees of the World Futures Conference in Boston that “there is no reason for any individual to have a computer in their home.” In 1980, the management consultants at McKinsey projected that by 2000, there might be 900,000 cell phone users in the U.S.; they were off by over 100-fold; the actual number was above 119 million. 

On the other hand, much-hyped technologies like 3D TV, Google Glass, and the Segway never really took off. For others, like cryptocurrency and virtual reality, the verdict is still out. 

AI itself has been notoriously difficult to predict. For example, in 2016, AI expert Geoffrey Hinton declared:

“Let me start by just saying a few things that seem obvious. I think if you work as a radiologist, you’re like the coyote that’s already over the edge of the cliff but hasn’t yet looked down, so doesn’t realize there’s no ground underneath him. People should stop training radiologists now. It’s just completely obvious that within five years, deep learning is going to do better than radiologists because it’s going to get a lot more experience.  It might be 10 years, but we’ve got plenty of radiologists already.”

Writing five years after this prediction, in his book The New Goliaths (2022), Boston University economist Jim Bessen observes that “no radiology jobs have been lost” to AI, and in fact, “there’s a worldwide shortage of radiologists.”

As Bessen notes, we tend to drastically overstate job losses due to new technology, especially in the near term. He calls this the “automation paradox,” and explains that new technologies (including AI) are “not so much replacing humans with machines as they are enhancing human labor, allowing workers to do more, provide better quality, and do new things.” 

Following the introduction of the ATM machine, the number of bank tellers employed actually increased, Bessen reports. Same for cashiers after the introduction of the bar code scanner, and for paralegals after the introduction of litigation-focused software products.

The explanation, Bessen explains, is that as workers are more productive, the costs of what their making tends to go down, which often unleashes greater consumer demand – at least up to a point. 

For instance, automation in textiles enabled customers to afford not just a single outfit, but an entire wardrobe. Consequently, from the mid-nineteenth century to the mid-twentieth century, “employment in the cotton textile industry grew alongside automation, even as automation was dramatically reshaping the industry,” Bessen writes. 

Yet after around 1940, automation continued to improve the efficiency of textile manufacturing, but consumer demand was largely sated; consequently, he says, employment in the U.S. cotton textile industry has decreased dramatically, from around 400,000 production workers in the 1940s to less than 20,000 today.

The point is that if historical precedent is a guide, the introduction of a new technology like generative AI will be accompanied by grave predictions of mass unemployment, as well as far more limited, but real examples of job loss, as we’ve seen in recent reporting. In practice, generative AI is likely to alter far more jobs than it eliminates and will likely create entirely new categories of work.

For example, Children’s Hospital in Boston recently advertised for the role “AI prompt engineer,” seeking a person who skilled at effectively interacting with chatGPT.

More generally, while it can be difficult to predict exactly how a new technology will evolve, we can learn from the trajectories previous technological revolutions have followed, as economist Carlota Perez classically described in her 2002 book, Technological Revolutions and Financial Capital.

Among Perez’s most important observations is how long it takes to realize “the full fruits of technological revolutions.” She notes that “two or three decades of turbulent adaptation and assimilation elapse from the moment when the set of new technologies, products, industries, and infrastructures make their first impact to the beginning of a ‘golden age’ or ‘era of good feeling’ based on them.” 

The Perez model describes two broad phases of technology revolutions: installation and deployment. 

The installation phase begins when a new technology “irrupts,” and the world tries to figure out what it means and what to do with it. She describes this as a time of “explosive growth and rapid innovation,” as well as what she calls “frenzy,” characterized by “flourishing of the new industries, technology systems, and infrastructures, with intensive investment and market growth.” There’s considerable high-risk investment into startups seeking to leverage the new technology; most of these companies fail, but some achieve outsized, durable success.

It isn’t until the deployment phase that the technology finally achieves wide adoption and use. This period is characterized by the continued growth of the technology, and “full expansion of innovation and market potential.” Ultimately, the technology enters the “maturity” stage, where the last bits of incremental improvement are extracted. 

As Perez explained to me, “A single technology, however powerful and versatile is not a technological revolution.” While she describes AI as “an important revolutionary technology … likely to spawn a whole system of uses and innovations around it,” she’s not yet sure whether it will evolve into the sort of full-blown technology revolution she has previously described. 

One possibility, she says, is that AI initiates a new “major system” – AI and robotics – within an ongoing information and communication technology revolution.  

At this point, it seems plausible to imagine we’re early in the installation stage of AI (particularly generative AI), where there’s all sorts of exuberance, and an extraordinary amount of investing and startup activity. At the same time, we’re frenetically struggling to get our heads around this technology and figure out how to most effectively (and responsibly) use it.

This is normal. 

Technology, as I wrote in 2019, “rarely arrives on the scene fully formed—more often it is rough-hewn and finicky, offering attractive but elusive potential.”

As Bessen has pointed out, “invention is not implementation,” and it can take decades to work out how best to use something novel. “Major new technologies typically go through long periods of sequential innovation,” Bessen observes, adding, “Often the person who originally conceived a general invention idea is forgotten.”

The complex process associated with figuring out how to best utilize a new technology may account, at least in part, for what’s been termed the “productivity paradox” – the frequent failure of a new technology to impart significant productivity improvement. We think of this frequently in the context of digital technology; economist Robert Solow wryly observed in a 1987 New York Times book review that “You can see the computer age everywhere but in the productivity statistics.”

However, as Paul A. David, an economic historian at Stanford noted in his classic 1990 paper, “The Dynamo and the Computer,” a remarkably similar gap was present a hundred years earlier, in the history of electrification. David writes that the dawn of the 20^th century, two decades after the invention of the incandescent light bulb (1879) and the installation of Edison central generating stations in New York and London (1881), there was very little economic productivity to show for it.

David goes on to demonstrate that the simple substitution of electric power for steam power in existing factories didn’t really improve productivity very much. Rather, it was the long subsequent process of iterative reimagination of factories, enabled by electricity, that allowed the potential of this emerging technology to be fully expressed.  

A similar point is made by Northwestern economic historian Robert Gordon in his 2016 treatise The Rise and Fall of American Growth. Describing the evolution of innovation in transportation, Gordon observes that “most of the benefits to individuals came not within a decade of the initial innovation, but over subsequent decades as subsidiary and complementary sub-inventions and incremental improvements became manifest.”

As Bessen documents in Learning by Doing (2015), using examples ranging from the power loom (where efficiency improved by a factor of twenty), to petroleum refinement, to the generation of energy from coal, remarkable improvements occurred during the often-lengthy process of implementation, as motivated users figured out how to do things better — “learning by doing.”

Many of these improvements (as I’ve noted) are driven by what Massachusetts Institute of Technology professor Eric von Hippel calls “field discovery,” involving frontline innovators motivated by a specific, practical problem they’re trying to solve.

Such innovative users—the sort of people who Judah Folkman had labeled “inquisitive physicians”—play a critical role in discovering and refining new products, including in medicine; a 2006 study led by von Hippel of new (off-label) applications for approved new molecular entities revealed that nearly 60% were originally discovered by practicing clinicians.

***

What does this history of innovation mean for the emerging technology of the moment, generative AI? 

First, we should take a deep breath, and recognize that we are in the earliest days of technology evolution, and nobody knows how it’s going to play out. Not the experts developing it, not the critics bemoaning it, not the consultants trying to sell work off our collective anxiety around it.

Second, we should acknowledge that the full benefits of the technology will take some time to appear. Expectations of immediate gains in productivity by plugging in AI seem as naïve, the dynamo-for-steam substitution all over again. While there are clearly some immediate uses for generative AI, the more substantial benefits will likely require continued evolution of both technology and workflow processes.

Third, it’s unlikely that AI will replace most workers, but it will require many of us to change how we get our jobs done – an exciting opportunity for some, an unwelcome obligation for others.  AI will also create new categories of work, and introduce new challenges for governance, ethics, regulation, and privacy.

Fourth, and perhaps most importantly: As mind-blowing as generative AI is, the technology is not magic. It doesn’t descend from the heavens (or Silicon Valley), deus ex machina, with the ability to resolve sticky ethical challenges, untangle complex biological problems, and generally ease the woes of humanity. 

But while technology isn’t a magic answer, it’s proved a historically valuable tool, driving profound improvements in the human condition, and enabling tremendous advances in science. The discovery of the microscope, telescope, and calculus all allowed us to better understand nature, and to develop more impactful solutions.

Technology changes the world in utterly unexpected and unpredictable ways. 

How exciting to live in this moment, and to have the opportunity — and responsibility — to observe and shape the evolution of a remarkable technology like generative AI. 

Yes, there are skeptics. I have a number of friends and colleagues who have decided to sit this one out, reflexively dismissing the technology because they’ve heard it hallucinates (it does), or because of privacy concerns (a real worry), or because they’re turned off by the relentless hype (I agree!)

But I would suggest that we owe it to ourselves to engage with this technology, familiarize ourselves, through practice, with its capabilities and limitations. 

We can be the lead users generative AI — like all powerful but immature transformative technologies — requires to evolve from promise to practice.

Recent Astounding HealthTech columns on Generative AI

• Generative AI: No Humbug
• Welcome to the AI Irruption
• New Book Contemplates How Generative AI Will Serve Man
• Tech, Pharma, and the Uneven Distribution of the AI-Enabled Future
• Big, If True: Opportunities and Obstacles Facing AI
• Pharma R&D Execs Offer Extravagant Expectations for AI But Few Proof Points

As the excitement around generative AI sweeps across the globe, biopharma R&D groups (like most everyone else) are actively trying to figure out how to leverage this powerful but nascent technology effectively, and in a responsible fashion.

In separate conversations, two prominent pharma R&D executives recently sat down with savvy healthtech VCs to discuss how generative AI specifically, and emerging digital technologies more generally, are poised to transform the ways new medicines are discovered, developed, and delivered.

The word “poised” is doing quite a lot of work in the sentence above. Both conversations seamlessly and rather expertly blend what’s actually been accomplished (a little bit) with the vision of what might be achieved (everything and then some).

The first conversation, from the a16z “Bio Eats World” podcast, features Greg Meyers, EVP and Chief Digital and Technology Officer of Bristol Myers Squibb (BMS), and a16z Bio+Health General Partner Dr. Jorge Conde.  The second discussion, from the BIOS community, features Dr. Frank Nestle, Global Head of Research and CSO, Sanofi, and Flagship Pioneering General Partner and Founder and CEO of Valo Health, Dr. David Berry.  (Readers may recall our discussion of a previous BIOS-hosted interview with Dr. Nestle, here.)

Rather than review each conversation individually, I thought it would be more useful to discuss common themes emerging from the pair of discussions.

Theme 1: How Pharma R&D organizations are meaningfully using AI today

AI has started to contribute meaningfully to the design of small molecules in the early stages of drug development.  “A few years ago,” Meyers says, BMS started “to incorporate machine learning to try to predict whether or not a certain chemical profile would have the bioreactivity you’re hoping.”  He says this worked so well (producing a “huge spike” in hit rate) that they’ve been trying to scale this up.

Meyers also says BMS researchers “are currently using AI pretty heavily in our protein degrader program,” noting “it’s been very helpful” in enabling the team to sort through different types of designs.

Nestle also highlights the role of AI in developing novel small compounds.  “AI-empowered models” are contributing to the design of modules, he says, and are starting to “shift the cycle times” for the industry.

AI is also now contributing to the development of both digital and molecular biomarkers.  For example, Meyers described the use of AI to analyze a routine 12-lead ECG to identify patients who might have undiagnosed hypertrophic cardiomyopathy.  (Readers may recall a very similar approach used by Janssen to diagnose pulmonary artery hypertension, see here.)

Nestle offered an example from digital pathology.  He described a collaboration with the healthtech company Owkin, whose AI technology, he says, can help analyze the microscope slides with classically stained tissue samples.

Depending on your perspective, these use cases are either pretty slim pickings or an incredibly promising start. 

I’ve not included what seemed to me as still exploratory efforts involving two long-standing industry aspirations:

• Integrating multiple data sources to improve target selection for drug development;
• Integrating multiple data sources to improve patient selection for clinical trials.

We’ll return to these important but elusive ambitions later, in our discussion of “the magic vat.”

I’ve also not included examples of generative AI, because I didn’t hear much in the way of specifics here, probably because it’s still such early days.  There was clearly excitement around the concept that, as Meyers put it, “proteins are a lot like the human language,” and hence, large language models might be gainfully applied to this domain.

Theme 2: Grand Vision

The aspiration for AI in in biopharma R&D were as expansive as the established proof points were sparse.  The lofty idea seems to be that with enough data points and computation, it will eventually be possible to create viable new medicines entirely in silico.  VC David Berry described an “aspiration to truly make drug discovery and development programmable from end to end.”  Nestle wondered about developing an effective antibody drug “virtually,” suggesting it may be possible in the future.  Also possible, he suggests: “the ability to approve a safe and effective drug in a certain indication, without running a single clinical trial.”

Both Nestle and Meyers cited the same estimate – 10^60 – as the size of “chemical space,” the number of different drug-like molecular structures that are theoretically possible.  It’s a staggering number, more than the stars in the universe, and likely far beyond our ability to meaningfully comprehend.  The point both executives were making is that if we want to explore this space productively, we’re going to get a lot further using sophisticated computation than relying on the traditional approaches of intuition, trial and error.

The underlying aspiration here strikes a familiar chord for those of us who remember some of the more extravagant expectations driving the Human Genome Project. For instance, South African biologist Sydney Brenner reportedly claimed that if he had “a complete sequence of DNA of an organism and a large enough computer” then he “could compute the organism.”   While the sequencing of the genome contributed enormously to biomedical science, our understanding of the human organism remains woefully incomplete, and largely uncomputed.  It’s easy to imagine that our hubris – and our overconfidence in our ability to domesticate scientific research, as Taleb and I argued in 2008 – may be again deceiving us.

Theme 3: Learning Drug Development Organization

For years, healthcare organizations have strived towards the goal of establishing a “learning health system (LHS),” where knowledge from each patient is routinely captured and systematically leveraged to improve the care of future patients.  As I have discussed in detail (see here), the LHS is an entity that appears to exists only as ideal with the pages of academic journals, rather than embodied in the physical world.

Many pharma organizations (as I’ve discussed previously) aspire towards a similar vision, and seek to make better use of all the data they generate.  As Meyers puts it, you “want to make sure that you never run the same experiment twice,” and you want to capture and make effective use of the digital “exhaust” from experiments, in part by ensuring it’s able to be interpreted by computers.

Berry emphasized that a goal of the Flagship company Valo (where he now also serves as CEO) is to “use data and computation to unify how… data is used across all of the steps [of drug development], how data is shared across the steps.”  Such integration, Berry argues, “will increase probably of success, will help us reduce time, will help reduce cost.”

The problem – as I’ve discussed, and as Berry points out, is that “drug discovery and development has historically been a highly siloed industry.  And the challenge is it’s created data silos and operational silos.” 

The question, more generally is how to unlock the purported value associated with, as Nestle puts it, the “incredible treasure chest of data” that “large pharmaceutical companies…sit on.”

Historically, pharma data has been collected with a single, high-value use in mind.  The data are generally not organized, identified, and architected for re-use.  Moreover, as Nestle emphasizes, the incentives within pharma companies (the so-called key performance indicators or “KPIs”) are “not necessarily in the foundational space, and that not where typically the resourcing goes.”  In other words, what companies value and track are performance measures like speed of trial recruitment; no one is really evaluating data fluidity, and unless you can directly tie data fluidity to a traditional performance measure, it will struggle to be prioritized.

In contrast, companies like Valo; other Flagship companies like Moderna; and some but not all emerging biopharma companies are constructed (or reconstructed — eg Valo includes components of both Numerate and Forma Therapeutics, as well as TARA biosystems) with the explicit intention of avoiding data silos.  This concept, foundational to Amazon in the context of the often-cited 2002 “Bezos Memo,” was discussed here. 

In contrast, pharmas have entrenched silos; historically, data were collected to meet the specific needs of a particular functional group, responsible for a specific step in the drug development process.  Access to these data (as I recently discussed) tends to be tightly controlled.

Data-focused biotech startups tend to look at big pharma’s traditional approach to data and see profound opportunities for disruption.  Meanwhile, pharmas tend to look at these data-oriented startups and say, “Sure, that sounds great.  Now what have you got to show for all your investment in this?” 

The result is a standoff of sorts, where pharmas try to retrofit their approach to data yet are typically hampered by the organizational and cultural silos that have very little interest in facilitating data access.  Meanwhile, data biotech startups are working towards a far more fluid approach to data, yet have produced little tangible and compelling evidence to date that they are more effective, or are likely to be more effective, at delivering high impact medicines to patients. 

Theme 4: Partnerships and External Innovation

Both BMS and Sanofi are exploring emerging technologies through investments and partnerships with a number of healthtech startups, even as both emphasize that they are also building internal capabilities. 

“We have over 200 partnerships,” Meyers notes, “including several equity positions with other companies that really come from the in silico, pure-play sort of business.  And we’ve learned a ton from them.”

Similarly, Nestle (again – see here) emphasized key partnerships, including the Owkin relationship and digital biomarker work with MIT Professor Dina Katabi.

Meanwhile, Pfizer recently announced an open innovation competition to source generative AI solutions to a particular company need: creating clinical study reports.

In addition to these examples, I’ve become increasingly aware of a number of other AI-related projects attributed to pharma companies that upon closer inspection, turn out to represent discrete engagements with external partners or vendors who reportedly are leveraging AI.

Theme 5: Advice for Innovators

One of the most important lessons from both discussions was the challenge for aspiring innovators and startups.

Berry, for example, explained why it’s so difficult for AI approaches to gain traction.  “If I want to prove, statistically, that AI or an AI component is doing a better job, how many Phase Two clinical readouts does one actually need to believe it on a statistical basis?  If you’re a small company and you want to do it one by one, it’s going to take a few generations.  That’s not going to work.”

On the other hand, he suggested “there are portions of the drug discovery and development cascade where we’re starting to see insights that are actionable, that are tangible, and the timelines of them and the cost points of them are so quickly becoming transformative that it opens up the potential for AI to have a real impact.”

Meyers, for his part, offered exceptionally relevant advice for AI startups pitching to pharma (in fact, the final section of the episode should be required listening for all biotech AI founders). 

Among the problems Meyers highlights – familiar to readers of this column – are the need “for companies that are focused on solving a real-world problem,” rather than solutions in search of a problem.  He also emphasized that “this is an industry that will not adopt something unless it is, really 10x better than the way things are historically done.” 

This presents a real barrier to the sort of incremental change that may hard be appreciate in the near term but can deliver appreciable value over time.  Even “slight improvements” in translational predictive models, as we recently learned from Jack Scannell, can deliver outsized impact, significantly elevating the probability of success while reducing the many burdens of failure.

Meyers also reminded listeners of the challenges of finding product-market fit because healthcare “is the only industry where the consumer, the customer, and the payor are all different people and they don’t always have incentives that are aligned.”  (See here.)

On a more optimistic note, Berry noted that one of the most important competitive advantages a founder has is recognizing that “a problem is solvable, because that turns out to be one of the most powerful pieces of information.”  For Berry, the emergence of AI means that “we can start seeing at much larger scales problems that are solvable that we didn’t previously know to be solvable.”  Moreover, he argues, once we realize a problem is solvable, we’re more likely to apply ourselves to this challenge.

Paths Forward

In thinking about how to most effectively leverage AI, and digital and data more generally, in R&D, I’m left with two thoughts which are somewhat in tension.

“Pockets of Reducibility”

The first borrows (or bastardizes) a phrase from the brilliant Stephen Wolfram: look for pockets of reducibility.  In other words – focus your technology not on fixing all of drug development, but on addressing a specific, important problem that you can meaningfully impact. 

For instance, I was speaking earlier this week with one of the world’s experts on data standards.  I asked him how generative AI as “universal translator” (to use Peter Lee’s term) might obviate the need for standards.  While the expert agreed conceptually, his immediate focus was on figuring out how to pragmatically apply generative AI tools like GPT-4 to standard generation so that it could be done more efficiently, potentially with people validating the output rather than generating it.

On the one hand, you might argue this is disappointingly incremental.  On the other hand, it’s implementable immediately, and seems likely to have a tangible impact. 

(In my own work, I am spending much of my time focused on identifying and enabling such tangible opportunities within R&D.)

The Magic Vat

There’s another part of me, of course, that both admires and deeply resonates with the integrated approach that companies like Valo are taking: the idea and aspiration that if, from the outset, you deliberately collect and organize your data in a thoughtful way, you can generate novel insights that cross functional silos (just as Berry says).  These insights, in principle, have the potential to accelerate discovery, translation (a critical need that this column has frequently discussed, and that Conde appropriately emphasized), and clinical development.

Integrating diverse data to drive insights has captivated me for decades; it’s a topic I’ve discussed in a 2009 Nature Reviews Drug Discovery paper I wrote with Eric Schadt and Stephen Friend.  The value of integrating phenotypic data with genetic data was also a key tenet I brought to my DNAnexus Chief Medical Officer role, and a lens through which I evaluated companies when I subsequently served as corporate VC.   

Consequently, I am passionately rooting for Berry at Valo – and for Daphne Koller’s insitro and Chris Gibson’s Recursion.  I’m rooting for Pathos, a company founded by Tempus that’s focused on “integrating data into every step of the process and thereby creating a self-learning and self-correcting therapeutics engine,” and that has recruited Schadt to be the Chief Science Officer.   I’m also rooting for Aviv Regev at Genentech, and I am excited by her integrative approach to early R&D.

But throughout my career, I’ve also seen just how challenging it can be to move from attractive integrative ambition to meaningful drugs.  I’ve seen so many variations of the “magic vat,” where all available scientific data are poured in, a dusting of charmed analysis powder is added (network theory, the latest AI, etc), the mixture is stirred, and then – presto! – insights appear. 

Or, more typically, not.  But (we’re invariably told) these insights would arrive (are poised to arrive) if only there was more funding/more samples/just one more category of ‘omics data, etc. — all real examples by the way. 

It’s possible that this time will be the charm – we’ve been told, after all, that generative AI “changes everything” — but you can also understand the skepticism.

My sense is that legacy pharmas are likely to remain resistant to changing their siloed approach to data until they see compelling evidence that data integration approaches are, if not 10x better, then at least offer meaningful and measurable improvement.   In my own work, I’m intensively seeking to identify and catalyze transformative opportunities for cross-silo integration of scientific data across at least some domains, since effective translation absolutely requires it.

For now, big pharmas are likely to remain largely empires of silos – and will continue to do the step-by-step siloed work comprising drug development at a global scale better than anyone.  Technology, including AI, may help to improve the efficiency of specific steps (eg protocol drafting, an example Meyers cites). Technology may also improve the efficiency of sequential data handoffs, critical for drug development, and help track operational performance, providing invaluable information to managers, as discussed here.

But foundationally integrating scientific knowledge across organizational silos?  Unless a data management organization already deeply embedded within many pharmas – perhaps a company like Veeva or Medidata – enables it, routine integration of scientific knowledge across long-established silos, in the near to medium term, seems unlikely.  It may take a visionary, persistent and determined startup (Valo? Pathos?) to capture persuasively the value that must be here.

Bottom Line:

Biopharma companies are keenly interested in leveraging generative AI, and digital and data technologies more generally, in R&D.  To date, meaningful implementations of AI in large pharmas seem relatively limited, and largely focused on small molecule design, and biomarker analysis (such as identifying potential patients through routine ECGs).  Nevertheless, the ambitions for AI in R&D seem enormous, perhaps even fanciful, envisioning virtual drug development and perhaps even in silico regulatory approvals.  More immediately, pharmas aspire to make more complete use of the data they collect but are likely to continue to struggle with long-established functional silos.  External partnerships provide access to emerging technologies, but it can be difficult for healthtech startups to find a permanent foothold with large pharmas.  Technology focused on alleviating important, specific problems – “pockets of reducibility” – seems most likely to find traction in the near term.  Ambitious founders continue to pursue the vision of more complete data integration.