A central challenge of R&D, like many disciplines characterized by rare, outsized success, is how to think about risk, as well as the contributions of luck and skill. 

Two new books – How To Become Famous, by Cass Sunstein, and On the Edge, by Nate Silver, offer valuable perspectives. I’ll also highlight several articles that provide additional relevant insight, including two examples of conspicuous pharma failures.

How To Become Famous — Cass Sunstein

Cass Sunstein, a professor at Harvard Law School, is perhaps known for his interest in behavioral science, and is co-author, with Nobel Laureate Richard Thaler, of Nudge.

Sunstein’s more recent effort, How To Become Famous, examines success, and focuses on how remarkably contingent it can be. He acknowledges early on that the title is “a bit of a cheat” since “one of my main points is that there is no recipe for how to become famous… This is not a how-to manual.”

We learn, for example, about a fascinating study of authors of the Romanic period by University of Toronto Professor Heather Jackson, examining the trajectories of Wordsworth, Austen, Keats, and Blake on the one hand, and Crabbe, Southey, Cornwall, Hunt, and Brunton on the other. The names in the first group, but not the second, are now iconic, yet it wasn’t at all clear during their lives who would be remembered and who would be largely lost to history.

Jackson’s analysis, Sunstein says, “strongly suggests that accident, contingency, champions, and luck” played a “massive role” in guiding the outcome, rather than any difference in ability.  Roll the dice again and history might have come out differently.

Or consider music; we might believe that the most successful songs emerge from something intrinsically special about them.  Yet Sunstein presents data from an experiment called “Music Lab” revealing that preference for one (unfamiliar) tune over another can be driven by the putative relative popularity (which the experimenter can manipulate) of the two songs.  

“Everything turned on initial popularity,” Sunstein reports, and slight differences in initial preferences can play an outsized role in shaping the outcome. 

Success begets success, and popular songs tend to become even more popular.

Perhaps the most relevant lesson Sunstein has for biopharma executives is the fallacy of studies that select on the dependent variable. The typical examples here are approaches that examine outlier successes (prominent CEOs, startups, blockbuster drugs) and then extract common features, based on the idea these are critical factors presumably worthy of emulation. 

The problem, Sunstein says, is that this represents shoddy logic.  Such analyses, he argues, can offer “no idea whether the unifying characteristics were responsible for or even contribute to” the success. 

But he acknowledges that these narratives are hard to resist – hence the popularity (as Sunstein points out) of so many business books (and, I might add, consultant reports) premised on just this approach.   Everyone who has spent time in a pharma will recognize this pattern – generalizing on the “lessons learned” based on distinctive factors associated with an individual success. 

We might do well consider the insightful if uncomfortable reflections of former Pixar CEO Ed Catmull, who noted that there isn’t a template (or, in pharma parlance, a “playbook”) for a successful original film – you need to invent it anew each time.

Sunstein reminds us repeatedly about the ever-present role of randomness in success.  Consider, for example, the great (arguably the greatest) boxer Muhammad Ali.  When he (then Cassius Clay) was 12, Sunstein writes, growing up in Louisville, Kentucky, Clay’s bike was stolen, and he went to a police officer saying he wanted to “whup” the thief.  The officer he approached, Joe Martin, turned out to run a boxing gym, and told Clay that he might want to learn how to box.  He did.

Or consider the band Fleetwood Mac, best known for singer Stevie Nicks and guitarist Lindsey Buckingham. Founded in 1967 by drummer Mick Fleetwood, without either Nicks or Buckingham, it was a blues band and not very successful, Sunstein tells us. 

One day, when Fleetwood was evaluating a recording studio, an engineer happened to play a tape of Buckingham and Nicks (a pair of struggling musicians at the time) just to provide a sense of how the studio sounded.  A week later, one of Fleetwood’s guitarists left, and the band needed a replacement; they approached Buckingham, who agreed only if Nicks could join as well.   Before long, Fleetwood Mac’s dreams began to be realized.

Sunstein’s fundamental point is that “it’s a mistake to attribute spectacular success to the intrinsic qualities of those who succeed.  Of course, it is true that those who succeed may well be extraordinary…but their extraordinariness was hardly sufficient to get them where they ended up.  Countless extraordinary people never get very far.”

A member of the Obama administration from 2009-2012, Sunstein remembers a particularly salient observation made by the President about CEOs: “They’re lucky to be where they are.” 

Obama continued,

“They might be amazing, but still, they’re lucky to be where they are.  They got a lot of good breaks.  Some of them don’t seem to know that.  But it’s true.  Look at me.  I hope I’m doing a good job, but I had a lot of luck.”

On The Edge – Nate Silver

Best known for his political forecasts (he is founder of FiveThirtyEight, where he served as Editor-in-Chief until 2023), Nate Silver’s true passion, it turns out, is poker, which he played professionally before venturing into politics.

In his new book, On The Edge, Silver offers a glimpse into the mindset of what he sees as a distinct category of people who view life in terms of risk and probability, aspiring to identify situations where the math is in their favor (in their parlance: where the “expected value” or “EV” is positive). 

Silver refers to this “ecosystem” as the River, which includes not only gamblers, but also many investors (especially Silicon Valley VCs, with whom he seems especially enamored) and some philosophers. 

Silver contrasts Riverians with citizens of the Village, which consists generally of highly educated members of the media, government, and academia who he sees as comparatively risk adverse. 

Silver also believes many in the Village have an unhealthy tendency to “couple” – let their political beliefs intrude upon their analytics – rather than “decouple,” which he contends is “a type of intelligence that is valued in The River.” 

He describes decoupling operationally as the use of “yes, but” statements.  For example, he says, a Riverian might say: “Yes, I disagree with the Chick-fil-A CEO’s position on gay marriage, but they make a damn fine chicken sandwich.”  

Silver’s point is the speaker may or may not agree with the CEO’s politics but can separate in her head her view of the politics and her view of the food.   

While highlighting potential flaws within each camp, Silver proudly aligns himself with the River and describes a natural resonance with what he sees as their more probabilistic and contrarian view of the world.

Silver is clearly infatuated with poker and offers a level of detail reminiscent of Melville’s description of whaling.  Similarly, Silver’s extensive discussion of effective altruism and rationalism – not to mention Sam Bankman-Fried (Silver is not a fan, and considers him, in the words of Tim Wu, a “false prophet.) – will prove excessive for many readers (including this one).

Yet Silver’s ability to capture the mindset of Riverians feels relevant in R&D, a power law business driven by the exceedingly rare, outsized success – and where, I might add, key decisions are approached with a mindset that seems to combine extreme caution and faux mathematic precision.  It’s a domain where there’s an incredible premium on predicting success and “picking winners,” even though the actual ability of anyone to do this is arguably extremely low.

Of particular interest, Silver discusses how probabilistic thinking relates to two types approaches to knowledge. 

“The fox knows many things, but the hedgehog knows one big thing,” wrote the ancient Greek poet Archilochus, a formulation popularized in the mid-twentieth century by the philosopher Isaiah Berlin.  

A famous study by professor Philip Tetlock discovered (see this magnificent 2005 New Yorker review by Louis Menand), perhaps surprisingly, that the best predictors – “Superforecasters’ —  turned out to be foxes (particularly those with structured, probabilistic approaches), rather than hedgehogs.

Looking at Silicon Valley, Silver sees an ecosystem consisting of founders – hedgehogs – who tend to have a singular belief in the promise of their company, and foxes – VCs – whose job is to “herd hedgehogs,” assembling them into portfolios. 

Silver then invokes a useful poker analogy, envisioning a game where each turn, you need to either fold or go all-in.  He considers two strategies players might have – in one approach (“prudent”), the player decides based on her hand whether to fold or not.  In a second approach (“degenerate” or “degen”), the player goes all-in every single turn. 

What’s interesting here is that while the more cautious, arguably more skillful player does better on average, and winds up broke less often, the very highest scores are obtained by those who went all-in, even though on average, they do worse, and they go bankrupt most of the time. 

“Is this my way of saying that the richest founders in the world are just degenerate gamblers who got lucky?” Silver asks.  “No, I’m not saying that.  I think they’re highly skilled degenerate gamblers who got lucky.” 

Presumably, one might appropriately ask the same questions about the development of blockbuster drugs.  To give a particular candidate medicine the best chance of success, you need a highly skilled team – but ultimately, blockbusters likely require far more luck (and the resources to support the inevitable failures – see here) than the dialogue celebrating the rare success might suggest.

Four Articles Worth Reading

1. This thoughtful, concise read by Wharton Professor (and AI guru – see here) Ethan Mollick discusses how survivorship bias, together with compelling narratives, often lead to the confusion of skill and luck. His focus is entrepreneurship, but the argument applies readily to biopharma R&D as well.
2. A key paper cited by Mollick is this 2012 gem from Denrell and Liu, pointing out what savvy investors like Michael Mauboussin and Howard Marks have long recognized: in domains where luck plays a significant role in determining outcomes, the very highest performers are likely to have adopted an excessively risky strategy and happened to get lucky; in contrast, the most skilled performers are less likely to be at the very top most years, but will consistently do reasonably well.

Thus, seeking to emulate the characteristics of the top performers will not necessarily lead to the best results.  As the authors warn, “widespread use of this heuristic to identify whom to learn from can lead to diffusion of very risky behavior, and ‘nudges’ may be necessary to help people resist the temptation to praise, blame, or learn from extreme performers.”

A key point made by Nassim Taleb in Fooled by Randomness, and reinforced by Mollick, is that because of survivorship bias, it’s both essential and difficult to search out examples of failures and missed opportunities, to provide at least a measure of narrative balance. 

In this spirit, readers might consider:

3. This captivating, insider account by distinguished endocrinology researcher (and former Harvard Medical School Dean) Jeffrey Flier, describing an early effort to pursue several novel approaches to diabetes, including GLP-1, in conjunction with a biotech company (CalBio), and partnered with a large pharma (Pfizer). The collaboration was born in the 1990s.  

Concluding that injectable diabetes medicines (other than insulin) had no future – Pfizer abandoned the partnership. That led the biotech to abandon it as well. Notably, GLP-1s achieved blockbuster status for the treatment of diabetes even before the obesity indication was added.

4. This terrific STAT article by Jason Mast entitled “How Pfizer’s Grand Gene Therapy Ambitions Crumbled.” According to one Pfizer source with whom Mast spoke, “I think we promised probably too much without really understanding the limitations…We projected hope rather than reality.  We didn’t know.” 

One founder, Jude Samulski, explained that “We all expected it was going to be turnkey” – a standard delivery vehicle with different genes for different diseases.  He added “we were naïve in thinking that it was going to be a universal delivery system, in universally everybody, in universally every disease.”

Also recommended

Success and serendipity:

• This discussion of Seth Stephens-Davidowitz’s Don’t Trust Your Gut, focused in particular on how to increase exposure to positive serendipity — a “how to take action” component palpably absent from Sunstein’s analysis.
• This WSJ review of Malcolm Gladwell’s Outliers
• This discussion of Michael Mauboussin’s insightful perspective on luck and skill.
• This discussion of the role of luck and skill in film and pharma
• This 2008 Financial Times commentary about serendipity and pharma, co-authored with Nassim Taleb.

Narrative Bias:

• This discussion of narrative bias and a caution about tidy success stories.
• A discussion of VC Ali Tamaseb’s Superfounders, emphasizing the need to liberate founders from narrative bias.
• This discussion of the how heroic founder narratives, amplified by VCs and journalists, can adversely impact entrepreneurs.

Therapeutics Stories:

• This discussion of discovery of Keytruda.
• This discussion of the arduous development of GLP-1 medicines.
• This discussion of the early work on ibrutinib.
• This WSJ book review about the discovery and development of Botox.

Probability, forecasting, and longshots:

• This discussion of the fragility of pharma forecasts.
• This WSJ review of a new book discussing Bayes Theorem.
• This critique of “sterile information” associated with pharma forecasts.

The last word: Dr. Judah Folkman:

• This WSJ review of Safi Bahcall’s Loonshots. To quote from the last paragraph of the piece, “… it is often impossible for an organization to determine, in advance, whether an innovative team is working on a hit or a flop. As the late Harvard researcher Judah Folkman used to tell his students: ‘If your idea succeeds, everybody says you’re persistent. If it doesn’t, you’re obstinate.’” 

The phrase “assent without belief” has been used to describe the concept of going along with the outward manifestations of an ideology without true conviction. Most commonly, this is used to describe a familiar contemporary approach to religious observance.  

It also seems to describe how the vast majority of biopharma colleagues view AI and other emerging digital and data technologies. 

While few pharma employees are likely to stand up and object loudly when their CEOs boast in Davos about embracing AI, most also don’t consider AI as particularly relevant for their own work. Rather, if they think about AI at it, it’s typically viewed as merely the latest shiny object, breathlessly venerated by CEOs, and eagerly operationalized (at least in theory) by management consultants. 

The upshot is that in practice, for most in pharma, AI specifically, and digital technologies more generally, tend to be seen as vaguely conceived top-down initiatives you need to navigate around in order to do you actual job.

Leading tech companies like OpenAI, Alphabet, Meta, and Microsoft, on the other hand, have profound conviction around AI. This is true not only for their executives but also for most of the engineers who are flocking to join. There is nothing performative about how these organizations are approaching AI. 

Tech’s seriousness about AI is backed by investment. As a recent New York Times story notes, top tech executives are continuing to spend gobs of money on AI; their capital expenditures in the last quarter have increased 63% compared to a year ago, mostly AI-related. Even so, as the article’s headline notes, “a payoff still looks a long way away.”

The lack of an immediate return has troubled analysts at Goldman Sachs (as I recently discussed), and others, prompting concerns of a bubble. 

Biopharma readers might take note of a striking recent report: a pharma CIO, after test driving an AI-enhanced version of Microsoft Office in his organization for six months, cancelled the upgrade, saying the additional cost (an extra $30/month per user) wasn’t worth it. Rather damningly, he described AI-produced PowerPoint slide decks as similar to “middle school presentations.”

Of particular interest – and something that I’m aware of occurring at other pharmas as well – the CIO noted that the ability to automatically take notes during Teams video meetings was regarded as such a liability by the Legal department (presumably because the document would be both legally “discoverable” and potentially incorrect or misleading) that this feature wasn’t activated by the company.

Ethan Mollick: AI Whisperer

These contrasting perspectives on AI would likely not surprise one of the most thoughtful voices on AI, Ethan Mollick. He is an Associate Professor of Management at Wharton, where he co-directs (with his wife, Lilach) the Generative AI Lab.

Mollick, a former entrepreneur (and a scholar of entrepreneurship) who describes himself as drawn to technology but not a technologist himself, has become a celebrity of sorts for his ability to move effortless between the worlds of education, technology, and industry.  He published the wildly popular Co-Intelligence in April, writes the “One Useful Thing” Substack, and advises policy makers including the Biden White House and companies such as J.P. Morgan, Google, and Meta.

Mollick is especially well-known for his compelling presentations and interviews; a recent podcast discussion with 20VC host Harry Stebbings provides a useful entry point into Mollick’s perspective and offers relevant guidance for those in biopharma trying to reconcile the excessive hype and ardent enthusiasm.

Mollick’s view is that while “everybody” has tried ChatGPT or similar models (like Google’s Gemini), only 5-10% of people in almost any typical company have even tried to use it more seriously, and only 2-3% of people have used it for at least the 10 hours Mollick believes is the minimum required to start to get a sense of what it can and can’t do.

In this sense, Mollick observes, “almost nobody uses these systems.”

One issue, Mollick argues, is that most large companies are fairly skittish about the use of AI.  “The [local] regulatory environment is unclear,” he says, and argues that between absolute prohibitions, severe limits, and ambiguous restrictions (perhaps the most common challenge), many employees are understandably reluctant to use the technology. Because most don’t play around with the technology, Mollick argues, they’re unlikely to get familiar enough with it to start appreciating what it can and cannot do, and to figure out how to incorporate it into their daily work and life — a practice Mollick strongly encourages. 

This is absolutely what I’ve both directly observed and also have heard about in biopharma specifically – lots of celebrating the idea of AI, but exceptional restrictions and anxiety around the actual use. 

(There are exceptions: in my last role, I was able to partner with a spectacularly forward-thinking senior executive, Colleen Beauregard, and an imaginative data science colleague, Iksha Herr, to explore immersively the promise of genAI with Beauregard’s enthusiastic team.)

Consequently, Mollick explains, the use of generative AI in most organizations tends to occur in secret, as those who are adept at leveraging tend not to advertise this, fearing potential repercussions. This obviously tends to slow the adoption of the technology.

Mollick also emphasizes that generative AI is “jagged” – it’s “really good at some stuff, really bad at other things,” he says, adding “As a result, it can’t sub in for all of human work…The question is, can that jaggedness get overcome?”

While AI enthusiasts tend to focus on the technology’s “top line capabilities,” he argues, they often overlook the context — “the human systems” and “the organizational systems” — that these technologies “have to interact with.” 

He adds, “We can’t be naïve about the work that needs to be done here to make this stuff operate. You can’t just drop these [technologies] in.”

Mollick contends that “we’re not even at the stage of integrating” generative AI into human systems yet. He adds that the way we interact with the underlying large language models is almost entirely via chatbot. He calls that an “insane process.”  Mollick is excited about the downstream potential but anticipates “we have 10 years or so of integrating [generative AI] slowly into human systems” to get through first.

Wanted: “Skilled Artisans” or “Lead Users”

To effectively leverage AI, Mollick suggests, we should learn the lesson from the steam engine.  The advances in productivity didn’t arise directly from Watt’s invention, he explains, but rather from “having skilled artisans in your factory, who said, I’ve got to the thing that can make power go back and forth. How do I create the gearing to connect that to [the tasks I want to do]?”

He continues, “It was the skilled artisans that made all of this work and made the manufacturers capture all the money. So you want to be a skilled artisan right now, you want to figure out how to take the back-and-forth power of an LLM and convert that into usable work inside your organizations.”

This perspective – I have typically favored Eric von Hippel’s term, “Lead Users,” but “Skilled Artisans” also works – aligns perfectly with perhaps the central message this column has repeatedly conveyed for years (see here in particular, also here, here, and here), informed by the work of James Bessen, Carlota Perez, and others. 

In short:

• Transformative emerging digital and data technologies hold exceptional promise for the discovery, development, and delivery of impactful new medicines for patients.
• This potential isn’t likely to be captured by merely substituting the new technology for an earlier technology.
• Rather, leveraging new technology requires re-imaging and re-inventing the work, in ways that were not readily achievable with previous technologies.
• Beyond the reimagination, successive incremental improvements (as James Bessen and Robert Gordon have described, as discussed here and here) are also required to unlock meaningful productivity gains from new technologies.
• Productively integrating transformative technologies takes time, as Paul David, James Bessen and Carlota Perez have all argued. Productively integrating AI (contra rosy consultant predictions) will also take time, as Mollick eloquently explains and as this column has consistently emphasized.
• The key driving force for leveraging new technologies are the lead users/skilled artisans who are impassioned about solving critical problems and open to adopting new technology if it can help get the job done better.
• The fascinating question for biopharma where the effective adoption of new technologies like AI will come from. The contenders are:
  • Large pharmas: well-resourced but exceedingly restrained, fairly impatient with their capital (given the demands of investors), and often defined by a famously challenging bureaucratic culture that can suppress innovation, as Safi Bahcall has chronicled (see here, here).
  • Smaller biotechs and innovative techbio startups: more willing to take risks, but generally less well-resourced, and often have neither the time nor the financing to absorb the inevitable costly failures so common in drug development – see here.
  • Ardent tech AI champions: have the appetite for risk and perhaps the financial resources and necessary patience but may lack the domain expertise (as emphasized by Mollick in the context of education) to know what they don’t know.
  • My bet: adoption will be driven by agile biotechs and techbios who successfully re-imagine a critical aspect of the drug discovery and development process and – this is key — are lucky enough to be able to persuasively demonstrate this value before their funding runs out.
• Bonus recommendation: For more on the central role of contingency in success, see both this recent piece on parallels between success in film and pharma, and Cass Sunstein’s thoughtful, if somewhat dispiriting new book, How To Become Famous.

Peter Attia is a prominent physician-turned-California longevity guru, and (to paraphrase Woody Allen) as California longevity gurus go, he’s one of the best, striving to remain grounded in science and data.

Known for his popular book Outlive, and his affection for “rucking” (look it up), Attia is also the host of a long-form podcast called The Drive, and an engaging guest on other worthwhile podcasts including Patrick O’Shaughnessy’s “Invest Like the Best” (here) and Bari Weiss’s “Honestly” (here).  He’s also the founder of Early, which describes itself as a science-based health program to “build your own, personalized, end-to-end longevity playbook.”

Attia’s latest podcast features a lengthy (two hour), exceptionally illuminating interview about AI and medicine with Zak Kohane, a physician-scientist at Harvard Medical School and head of the Department of Biomedical Informatics (with which I’m affiliated). 

The discussion starts with a quick, knowledgeable review of the history of AI (interested readers might also enjoy several of the books summarized here), and then explores how AI is likely to impact health and medicine.

The entire episode – with thoughtful questions from Attia and nuanced replies from Kohane – is unquestionably worth the time (and extra credit if, like me, you listen while jogging, even if not quite rucking). 

However, I wanted to highlight three related topics raised by these two senior medical experts that were especially striking:

• How AI may fit into the evolution of medicine (tl;dr AI augmentation of non-MD health providers, initially in domains where there aren’t enough doctors);
• What’s most likely to keep the promise of AI in health from being fully realized (tl;dr incumbent healthcare systems maintaining their stranglehold on patient data)
• What’s the most promising driver of AI-enabled disruption in medicine (tl;dr empowered, exceptionally motivated patients and their families)

AI and the future of medicine

AI is already poised to impact medicine. Kohane highlights the promise of AI in image-based specialties such as dermatology and radiology, where high-resolution images, fed into machine learning models, can identify image abnormalities as well as (and in many cases even better than) the human eye. 

He also points out that in contrast to the infamous prediction of legendary AI expert Geoff Hinton (who suggested in 2016 that radiologists would be obsolete in 5-10 years), AI’s “are not replacing the doctors because image recognition process is only part of their job.” 

Focusing on the use of AI for physician augmentation (rather than replacement), Attia suggests the adoption of AI would likely occur in two stages.  The first specialties to be impacted, he says, will probably be those he describes as visually focused: pathology, radiology, dermatology, and cardiology (especially those most involved in interpreting cardiac imaging studies).  Next would be specialties he sees as focused on integrating “language data” and visual data, like primary care doctors and pediatricians.

While Kohane says he agreed with Attia’s proposed order of adoption, he emphasizes that a more immediate consideration is the shortage of doctors in key areas like primary care. 

“You have to ask yourself, how can we replace these absent primary care practitioners with nurse practitioners, with physician assistants augmented by these AIs, because there’s literally no doctor to replace.”

More broadly, medicine’s challenge and AI’s opportunity, as Kohane describes it, is that fewer doctors are choosing to go into specialties not associated with highly financially remunerative procedures.  While there doesn’t appear to be a shortage of future interventional radiologists or dermatologists (which he describes as high-paying and high prestige), there aren’t enough traditional radiologists or primary care doctors. 

“I go around medical schools and ask who’s becoming a primary care doctor?  Almost nobody,” Kohane said.  “So primary care is disappearing in the United States. In fact, Mass General and Brigham announced officially they’re not seeing [new] primary care patients [see here].”

Kohane continues,

“There’s a huge gap emerging in the available expertise. So, it’s not what we thought it was going to be, that we had a surplus of doctors that had to be replaced [by AI]. It’s just we have a surplus in a few focused areas which are very popular. And then for all the work of primary care and primary prevention kind of stuff that you [i.e. Peter Attia] are interested in, we have almost no doctors available.”

The gap in “expert clinicians,” Kohane suggests, might be filled by non-MDs like nurse practitioners and physician-assistants augmented by AI.

Impediment to AI-Innovation in Healthcare

The tremendous potential of AI today, Kohane explains, can be attributed to the compounding effects of large data sets, neural networks, and suitable computer chips (GPUs). 

Kohane describes the tremendous importance of rich clinical datasets, the sort of information one might associate with electronic health records (EHRs). 

However, he notes that the EHR “turns out not to be the answer in the United States. Why? Because in the United States, we move around. We don’t stay in any given health care system that long. So very rarely will I have all the measurements made on you.”

However, he adds, this isn’t the case in other countries, and cites the example of the Clalit in Israel, which he notes “published all the big COVID studies looking at the efficacy of the vaccine. And why could they do that? Because they had the whole population available, and they have about 20, 25 years’ worth of data on all of their patients in detail and family relationships.”

In the U.S., he asks, “Where is that data going to come from?” 

Some organizations in the U.S. believe “they can get enough data” on their own, he says, and cites the Mayo Clinic as one example. He also notes that “there are some data companies that are trying to get relationships with health care systems where they can get de-identified data,” although he doesn’t sound very optimistic about this approach.

Attia, for his part, strongly agrees that health data is exceedingly “unfriendly,” and cites a pain point from his own practice.

When lab tests are obtained for patients, he says, and “we want to generate our own internal reports based on those…it’s almost impossible to scrape those data out of the labs because they’re sending you PDF reports. Their APIs are garbage. Nothing about this is user friendly.”

“Just because our patients own the data doesn’t make it easy to get,” he continues. “There is no aspect of my practice that is more miserable and more inefficient than data acquisition from hospitals. It’s actually comical, absolutely comical.”

Attia adds, “Is there a more user-hostile industry from a data perspective than the health industry right now?”

Kohane agrees, and explains, “There’s a good reason why. Because they’re keeping you captive.”

Going forward, he says, he’s most concerned about the tendency of “the medical establishment” to “pour concrete over practices.” 

He explains that hospitals are huge entities that bring in huge revenue but have very small margins. In this situation, he argues, “you’re going to be very risk averse and you’re not going to want to change.” 

Consequently, while Kohane said he remains excited about the potential of new businesses that come up with ways of delivering data-driven insights to patients about their individual health needs, he’s concerned about what these startups are up against. Specifically, he warns of “hospitals doing a bunch of information blocking … that will make it harder for these new businesses to get created.”

Can Patients Drive Disruption?

A key theme of both Kohane’s recent book about GPT-4 (including in particular the sections written by co-author Carey Goldberg) as well as Kohane’s conversation with Attia is the embrace of generative AI tools by patients.  “They’re being used by patients all the time in ways that we had not anticipated,” Kohane says.

Kohane cites the example of a mom whose child had confusing symptoms that doctors struggled to synthesize into a diagnosis.  Frustrated, she typed the information into ChatGPT, which suggested something called “tethered cord syndrome,” subsequently confirmed by imaging.

At this point, Attia asks a critical question: was it surprising that none of the doctors had tried to use ChatGPT to assist with the diagnosis?

Kohane’s deeply discouraging response was that “most clinicians I know do not have what I used to call the Google Reflex” – the instinct to look up and pursue curious and confusing observations.

The reason, he suggests, is that “doctors are in a very unhappy position these days. They’re really being driven very, very hard,” and morphing “into data entry clerks.” 

Essentially, Kohane explains, physicians are too busy with the operational aspects of medicine – in particular, documentation and billing.  This means they have less time to think, and perhaps less incentive to think. 

(I anticipated and discussed this challenge – preserving creativity in medicine – 20 years ago, see here.)

Consequently, Kohane sees patients, and those invested the most in them (especially their families), as the most powerful disruptive force in medicine. 

(I’ve also emphasized the critical role of patients and families as key drivers of care – including in the NYT here, here, as well as in The Atlantic here).

Kohane points to patients without primary care doctors who might be worried about a rash, for example. They recognize that they have a power tool, he says, and given the choice between waiting months to potentially be seen or using the tool, they’ll use the tool.

“It’s better than no doctor for sure. And maybe better.”

Kohane also emphasizes that because patients have the legal right to their own data, even though accessing these data can often be difficult, the legal right creates opportunities. 

He points to Apple Health, which he says has partnered with 800 hospitals and makes it easy to get some of your key health data on your phone or computer. 

“Now, there’s not a lot of companies that are taking advantage of that,” he says, “but right now, that data is available on tens of millions of Americans.”

Kohane predicts that in “the next 10 years, there’ll be a company, at least one company that figures out how to use that patient’s right to access through dirty APIs, using AI to clean it up, provide decision support with human doctors or health professionals to create alternative businesses.”

Yet, he acknowledges, “I don’t want to underestimate the medical establishment’s ability to squish threats, so we’ll see.”

Bottom Line

Rapidly improving AI is likely to impact medicine by augmenting the capabilities of specialists, particularly those focused on visual diagnoses.  AI might also provide the expert insight to enable non-MDs to provide care in situations where MDs are increasingly unavailable.  The development of AI for clinically focused applications (in healthcare as well as in drug development) depends critically on the availability of relevant clinical data, which (despite the existence of at least some legislation) remains excruciatingly, maddingly, disturbingly difficult to pry from the grip of incumbent health organizations. 

But if these organizations are an immovable object, then patients and their families might represent an irresistible force, determined to leverage all available tools to accelerate diagnoses and identify potential treatments (David Fajgenbaum’s own miraculous story, as well as the journeys chronicled by Amy Dockser Markus, come to mind – see here).  Ideally, despite institutional and occupational hurdles, inquisitive physicians such as those celebrated by Judah Folkman will also drive the use and acceleration of AI technologies in service of improved patient care.

“It was all mixed into one, enormous, overflowing stew of very real technological advances, unfounded hype, wild predictions, and concerns for the future.  ‘Artificial intelligence’ was the term that described it all.” – Cade Metz, Genius Makers

 

The buzzy excitement around artificial intelligence (AI), and most recently generative artificial intelligence (genAI), has inspired some biopharma leaders, exasperated many others, and touched almost everyone.

Leading management consulting firms have sold an enormous amount of business by persuading biopharma companies that:

• They are already lagging dangerously behind their competitors on AI adoption;
• There are tremendous productivity gains to be found, and value to be created, in the expeditious adoption of AI.

Within biopharma R&D departments, most researchers remain predictably skeptical of the incessant hype, even as many are authentically curious about promising advances (like AlphaFold, whose inventors received the 2023 Lasker Award for Basic Medical Research).  They are also politically astute enough to genuflect to senior management’s imperative to demonstrate the organization’s embrace of AI. 

One result has been an AI version of innovation theater, where there’s all sorts of demonstration projects, working groups, PowerPoint decks, partnerships, and celebratory speechifying.  A huge amount of heat is generated, but so far, relatively little light.

At one level, none of this is surprising.  As I discussed in 2019 in the context of precision medicine, and more expansively in 2023 in the context of AI, it historically takes a very long time for us to figure out how to productively use new technologies.  As economic historians like Paul David, Carlota Perez, James Bessen, Robert Gordon, and others have consistently reminded us (as I discussed here), we don’t tend to wring productivity of out new technologies overnight; more typically, it takes decades, and many rounds of successive incremental innovations.

Yet, it’s easy to imagine, in the context of breathless pitches and extravagant promises, that perhaps this time it’s different – perhaps AI has found a way to beat the historical odds, and is leading to the sort of immediate, measurable productivity gains that enthusiasts promise and biopharma executives desperately seek.

For biopharma in particular, the excitement is understandable.  As visionaries like Mustafa Suleyman argues in The Coming Wave (my 2023 WSJ review here) and Jamie Metzl argues in Superconvergence (my just-published WSJ review here), the thesis that accelerating revolutions in biotech and AI are compounding each other, and leading us towards a promising, tech+bio future is not just compelling but directionally correct.  The question, of course, is when will we realize this AI-infused bio-rapture?

According to a recent, arresting Goldman Sachs (GS) report, entitled “GenAI: Too Much Spend, Too Little Benefit?” perhaps we shouldn’t hold our breath.

The GS report is worth reading in its entirety, but I’ll focus on several salient sections: an interview with the distinguished scholar and MIT economist Daron Acemoglu; an interview with GS’s Global Head of Equity Research Jim Covello, and an interview with two GS Senior Equity Research Analysts, Kash Rangan and Eric Sheridan.

Before we get to these details, it’s worth noting how refreshing it is to read a corporate document that conveys multiple, at times conflicting viewpoints around complex issues like the future state and anticipated economic impact of AI.  While top management consulting firms typically offer a singular, consensus view on topics like the path forward for AI, this report from GS acknowledges and systematically explores differences in perspectives and assumptions. The result is an unusually substantive and credible report that conveys nuance and embraces uncertainty.

Daron Acemoglu: Hopeful Skeptic

Daron Acemoglu is a distinguished economist at MIT and the co-author, most recently of Power and Progress.  A WSJ review by Deirdre McCloskey described Acemoglu as “a shoo-in for Nobel Prize” in economics, and said the book expressed the authors’ view that “The invisible hand of human creativity and innovation…requires the wise guidance of the state.”

In his conversation with GS, Acemoglu expressed excitement about the promise of genAI, noting it “has the potential to fundamentally change the process of scientific discovery, research and development, innovation, new product and material testing, etc. as well as create new products and platforms.”

However, he cautioned, “these truly transformative changes won’t happen quickly and few—if any—will likely occur within the next 10 years.”

Instead, he suggests, “AI technology will instead primarily increase the efficiency of existing production processes by automating certain tasks or by making workers who perform these tasks more productive.”

He also thinks AI, even with more data and fancier chips, will still struggle with open-ended tasks, like improving “a customer service representative’s ability to help a customer troubleshoot problems with their video service.” 

In addition, like many others, Acemoglu worries “where more high-quality data [to power future AI models] will come from and whether it will be easily and cheaply available to AI models.”

He recognizes the future possibilities of genAI, he says, and hopes AI creates new tasks, products, business occupations, competences – but adds this is “not guaranteed.” 

While emphasizing that “Every human invention should be celebrated, and generative AI is a true human invention,” he is concerned that “too much optimism and hype may lead to the premature use of technologies that are not yet ready for prime time.”

Jim Covello: Pessimistic Skeptic

As Jim Covello, the Head of Global Equity Research at GS sees it, one critical concern around AI relates to the “substantial cost to develop and run” the technology; the investment only makes sense if AI can “solve extremely complex and important problems for enterprises.”  But solving complex problems, he says, is something the technology “isn’t designed to do.”

Covello challenges several familiar assertions used to justify current costs. 

“Many people attempt to compare AI today to the early days of the internet,” he says.  “But even in its infancy, the internet was a low-cost technology solution that enabled e-commerce to replace costly incumbent solutions.” 

In contrast, he argues, AI technology is starting out “exceptionally expensive.”

He adds that the idea that technology “typically starts out expensive before becoming cheaper is revisionist history.”  E-commerce, he asserts, “was cheaper from day 1.”

He also says we can’t count on AI prices declining significantly; the dramatic historical decrease in the price of semiconductor chips, he says, was due to fierce competition; but at least today, “Nvidia is the only company cable of producing the [computer chips] that power AI.”  Consequently, Nvidia is likely to maintain pricing power in the near-term.

Covello is also skeptical about the transformative potential of AI, arguing “people generally substantially overestimate what the technology is capable of today.”  He adds, “I struggle to believe that the technology will ever achieve the cognitive reasoning required to substantially augment or replace human interactions.”

He also contends, “Humans add the most value to complex tasks by identifying and understanding outliers and nuance in a way that it is difficult to imagine a model trained on historical data would ever be able to do.”

Covello points out that he was a semiconductor analyst when smartphones arrived and followed the evolution of smartphone functionality closely.  As he remembers it, the ensuing roadmap was clear, “with much of it playing out just as the industry had expected.”

In contrast, he argues, “No comparable roadmap exists today. AI bulls seem to just trust that use cases will proliferate as the technology evolves. But 18 months after the introduction of generative AI to the world, not one truly transformative—let alone cost-effective—application has been found.”

Of particular relevance for biopharma, he notes that “companies outside of the tech sector … face intense investor pressure to pursue AI strategies even though these strategies have yet to yield results. Some investors have accepted that it may take time for these strategies to pay off, but others aren’t buying that argument.”

He warns that “The more time that passes without significant AI applications, the more challenging the AI story will become. And my guess is that if important use cases don’t start to become more apparent in the next 12-18 months, investor enthusiasm may begin to fade.”

Rangan and Sheridan: Long-term Optimists

A somewhat more positive perspective on genAI, at least in the long-term, was expressed by Kash Rangan and Eric Sheridan, both Senior Equity Research Analysts at GS.

Noting that “hardly a week goes by without reports of a new, and better, AI model,” Rangan said he remained enthusiastic about genAI long-term potential, but acknowledged, “we have yet to identify AI’s ’killer application.’ ”  

Similarly, while acknowledging that the technology “is still very much a work in progress,” Sheridan said “it’s impossible to sit through demonstrations of generative AI’s capabilities at company events or developer conferences and not come away excited by the long-term potential.”

Rangan acknowledges that “AI technology is undoubtedly expensive today” but argues (in contrast to Covello) that “the cost-equation will change.” 

Pointing out that “people tend to overestimate a technology’s short-term effects and underestimate its long-term effect,” Rangan adds “Nobody today can say what killer applications will emerge from AI technology. But we should be open to the very real possibility that AI’s cost equation will change, leading to the development of applications that we can’t yet imagine.”

Sheridan agrees the economics of AI are challenging now: “I readily acknowledge that the return on invested capital (ROIC) visibility is currently low, and the transformative potential of AI will remain hotly debated until that becomes clearer.” 

He concludes, “people didn’t think they needed smartphones, Uber, or Airbnb before they existed. But today it seems unthinkable that people ever resisted such technological progress. And that will almost certainly prove true for generative AI technology as well.”

Concluding thoughts

I remain optimistic and energized by the promise to be found at the intersection of AI and biotechnology, and view digital technologies like AI as increasingly essential tools for understanding biology as well as for effectively managing biopharma R&D.  But even if at times genAI seems magical, we can’t treat it as magic.  We can’t operationalize it as magic.  We can’t invoke it as a force that will descend from the rafters, deus ex machina, and somehow fix what ails our organizations.  Nor should we set it aside, dismissing it simply the newest new thing.  By steering a path between credulous mysticism on the one hand, and reflexive cynicism on the other, we can inquisitively explore and thoughtfully interrogate this powerful emerging technology, and identify meaningful opportunities for productive application in biopharma R&D.