David Shaywitz

The worlds of technology and entrepreneurship are captivated by recent advances in generative AI and large language models (LLMs). 

The arrival of ChatGPT, developed by OpenAI (a startup partnered with Microsoft), caused Google to declare a “Code Red,” akin to “pulling the fire alarm,” the New York Times explained. The latest class of startups at Y Combinator are reportedly flocking to AI. Students in the Harvard MS-MBA biotechnology program I advise tell me they are using ChatGPT constantly in section, and say the page is always open. I’ve also heard many medical trainees are routinely using it.

But if the future has arrived, as science fiction writer William Gibson famously observed, it’s not evenly distributed.

“People in tech are freaking out about LLMs, but not in science yet,” writes University of Rochester chemist and AI expert Andrew White.

Andrew White

“AI hype may be too high,” White observes, “but we’re missing hype on smart people + AI…. this will be the biggest transformation in science since the internet. Universities must be completely rethinking education and research.”

White’s onto something. There is an absolute frenzy of activity in tech, and a rash of startups seeking to apply AI to some aspect of biopharma, often drug discovery. 

But the view from within established biopharma companies tends to be more reserved. 

Far, far more reserved. 

Curiosity, wariness, weariness, and concern

As best I can determine, the view of incumbent biopharmas reflects a combination of curiosity, wariness, weariness, and concern. 

There’s authentic curiosity around ChatGPT and AI more generally, wariness about the newest new thing, weariness about successive cycles of technology hype and disappointment, and concern about the many unknowns.

No one in the pharma C-suite wants to miss the AI train. Then again, no CEO wants to explain to the media how confidential company information was inadvertently leaked – as recently happened at Samsung. Biopharma also must be serious about patient confidentiality and strict adherence to regulatory process. Not surprisingly, the industry is loath to engage in anything that could put these vital priorities at risk.

Consequently, many biopharmas are taking a cautious approach, considering generative AI the way they routinely evaluate other emerging technologies. Rigorous guidelines for use are established, highly specified pilot projects are proposed, reviewed and subject to tiers of approvals. Once clear guardrails are established, the new technology is gently prodded and poked. The potential is gingerly explored.

It’s not just senior executives who are risk averse. Most of my colleagues in biopharma are skeptical. 

After successive cycles of hope and disappointment, many seasoned drug developers just aren’t smitten by what seems like the latest shiny tech object. (I was probably in this camp as well before I started speaking with grounded AI experts like Zak Kohane and Peter Lee, who had early access to GPT-4, and began to appreciate over time that the technology represents something profoundly different.)

Many are still trying to understand what this technology can do. For instance, a colleague this week told me he tried GPT-4 for search and concluded “Meh. I’m still going to use Google.”

Playing around (or not)

What’s abundantly clear to me from Kohane, Lee, and others who have invested time in understanding GPT-4 is that:

1. To really understand what GPT-4 can do, you need to spend a lot of time just playing with the technology, trying out all sorts of things, particularly in your area of expertise (like medicine).
2. The more time you spend with GPT-4, the more it starts to feel like you’re developing a relationship with an extraordinary alien intelligence – powerful, imperfect, mystifying, intriguing.

Ideally – as one expert I spoke with this week suggested – the best way, in theory, for pharma colleagues to fully leverage GPT-4 would be to play around with it. This exploration would help determine the extent to which the technology could enable everyday tasks, and eventually contribute more profound insights. 

Given the concerns about risk, I don’t see this happening in biopharma. While data scientists like Cloudera co-founder Jeff Hammerbacher famously aspire to “party on the data,” this unfettered, playful approach tends not to be the dominant mindset of large biopharmas, given their abiding concerns around potential downside consequences.

One way GPT-4 is likely to arrive in pharma is through consultant service companies like Tata Consulting Services, Accenture, and Cognizant, to use the three examples highlighted by Chamath Palihapitiya on the “All In” podcast. As Palihapitiya notes, these companies do coding-for-hire work at scale, and are thus likely to be the first to operationalize tools like GPT-4 (which can assist and enable coders) so these firms can accomplish more with fewer people. 

GPT-4 will also come to pharma through applications, ideally from established, trusted vendors.  Familiar Microsoft Office products supercharged by GPT-4 will be an early example. For many emerging tools, generative AI will operate under the hood, largely invisible. Users, understandably, are likely to focus on the capabilities an application is providing — the job to be done — rather than on the underlying technology making this possible.

Culture Contrast

The contrasting attitudes of established companies and startups towards emerging technology — including, but hardly limited to, GPT-4 — reflects radically different goals and relationships to perceived risk.

Large established companies are defined by their ability to execute reliably at scale. In biopharma R&D, this means being able to manage a portfolio of complex programs globally. Included: everything from the consistent manufacturing of a range of products to the safe, responsible, and efficient conduct of clinical trials to navigating different regulatory procedures across the world. 

It’s a huge challenge to orchestrate these activities in parallel. The work requires a deliberate focus on establishing and optimizing repeatable processes, and making careful, deliberate choices – often involving multiple stakeholders – to ensure any risks to the established business are identified and hopefully mitigated.

Startups typically can’t, and don’t, work this way. Wall Street Journal technology columnist Chris Mims recently highlighted the prismatic example of Noam Bardin, a startup CEO who was absorbed into Google after his company, Waze, was acquired.

Noam Bardin

“What seems natural at a corporation,” Bardin said, “multiple approvers and meetings for each decision—is completely alien in the startup environment: make quick decisions, change them quickly if you are wrong.”

Bardin also described to Mims the differences in incentives he noticed before and after the acquisition.  “Before the sale,” Mims reports, “everyone’s financial interest was aligned with the performance of the company’s products. Once Waze was a subsidiary, getting ahead was all about getting promoted.”

As I’ve discussed (see here, here) a nearly identical description is found in Safi Bahcall’s Loonshots. The author, a former biotech executive, explains how disruptive innovation can be stifled by corporate processes dominated by risk-aversion and enlightened careerism. 

Suffocating processes designed to reduce risks are reportedly what caused Google to fall behind OpenAI and Microsoft in generative AI, according to the WSJ. “Now Google, the company that helped pioneer the modern era of artificial intelligence, finds its cautious approach to that very technology being tested by one of its oldest rivals [Microsoft],” the Journal reports (also covered nicely in this WSJ-associated podcast).

The risk aversion of large, established companies creates an opportunity for risk arbitrage. Startups can naturally shoulder more risk because they have less to lose and potentially more to gain. As I’ve discussed, this is reportedly how PayPal was able to defeat eBay’s attempt to develop a rival online payment product with Wells Fargo (called Billpoint). It is almost certainly why the most important generative AI applications for biopharma are likely to be developed by someone else. 

Bottom Line

While many technology companies view generative AI as essential technology that must be adopted as an existential, urgent imperative, established biopharma companies are taking a more cautious approach, curious about the opportunities, but wary of the many uncertainties and risks. 

Additional Astounding columns on generative AI:

• Generative AI: No Humbug
• Welcome to the AI Irruption
• New Book Contemplates How Generative AI Will Serve Man

David Shaywitz

In 1935, Astounding Science — the premier science fiction magazine of the day, and perhaps ever — published “Proxima Centauri.” It was a short story envisioning an encounter between a voyaging earth spacecraft and a ship from a nearby star system.

Written by Murray Leinster (William Fitzgerald Jenkins), the tale features the original use of the phrase “first contact” in science fiction. Think Close Encounters of the Third Kind, or perhaps more appropriately the “To Serve Man” episode of The Twilight Zone, since Leinster’s aliens turn out to be carnivorous plants who quickly develop (as alien plants often do) a taste for human flesh.

“First contact” is also the phrase Microsoft’s head of research Peter Lee selects to describe his initial encounter with GPT-4, in The AI Revolution in Medicine: GPT-4 and Beyond, the engaging and timely new book he’s co-authored with Harvard physician-scientist and

AI researcher Zak Kohane and journalist Carey Goldberg, documenting the trio’s early experience with this emerging, intriguing technology.

Lee’s phrase seems apt. When I asked GPT-4 to provide the first paragraph of a blog post on the subject, written in the style of The New Yorker, it responded:

The phrase “first contact” conjures up images of flying saucers and green men, of awe and wonder, of fear and loathing. It is the ultimate fantasy—or nightmare—of humanity’s encounter with the alien Other. But what if the Other is not from another planet, but from another machine? What if the first contact we are about to make is not with a distant civilization, but with a new form of artificial intelligence that surpasses our own in linguistic prowess and creativity? This is the prospect that looms over us as we await the arrival of GPT-4, the latest and most advanced iteration of a technology that can generate coherent and convincing texts on any topic, given only a few words or sentences as input….

GPT-4 even provides a cogent opinion on the popularity of “first contact” stories in science fiction, suggesting it may reflect the interplay of curiosity, fear, hope, and imagination.

Lee and Kohane have worked on AI their entire careers – yet neither seemed prepared for their encounter with GPT-4. Both described losing sleep as they contemplated the implications of what they describe as an “alien intelligence” with whom they each formed a “relationship.” Both are struck by its capabilities,yet acknowledge its limitations. 

Lee describes GPT-4 as “at once both smarter and dumber than any person you’ve ever met.”  He is particularly captivated by its ability to “imagine emotions and perhaps even empathize with people.”

Lee and his colleagues also present a series of specific, formalized challenges to GPT-4, examining its common-sense understanding, its moral judgement, and even its ability to navigate “theory of the mind” tests, exploring GPT-4’s ability to intuit what others might be thinking. The performance of the technology was extraordinary, Lee reports. 

Responding to one question involving two roommates who shared a cat that was moved from one place to another when one roommate left the apartment, GPT-4 not only correctly anticipated the likely reaction of each roommate, but even offered an opinion on what the cat in the scenario might be thinking – namely that the two roommates were “very rude and confusing because they keep moving it from one place to another without asking.”

As discussed in my last column, Kohane was especially impressed by the GPT-4’s clinical reasoning skills, and by the way it approached complex medical cases and suggested reasonable next steps. On the other hand, GPT-4 would also make some fairly basic mistakes (especially involving math), and would often confidently assert it was correct even when it demonstrably was not.

Having established GPT-4’s bona fides, Lee, Kohane, and Goldberg devote the balance of their text to a question that even the technology itself might struggle with: how do you solve a problem like GPT-4?  Each of the three authors seems focused on slightly different aspects of the challenge.

Lee appears especially preoccupied with the underlying existential question of what GPT-4 is, and whether at some level it can “understand” and “empathize.” He seems consumed by GPT-4’s capabilities, particularly since so many of these, “were not programmed by humans. Instead, they emerged into existence – sometimes unexpectedly – as its neural network grew.”

He continues, “how on earth can it be that the next-word prediction can possibly do… all the things that we’ve seen…unfortunately, we simply do not know the answer to this question.”

Kohane seems hopeful the technology can be used as a partner or co-pilot for healthcare providers, reducing burnout, administrative burden, and medical error, while allowing providers to practice at the top of their license. He delights in the possibility of physicians able to focus once more on the complex reasoning and human relationships that had first attracted so many of us to the profession.

Carey Goldberg

Projecting forward, Kohane imagines a next-generation large language model, fancifully called “Dr. One-With-Everything,” encompassing “protein structure, other basic biological databases (like gene regulation and human genetic variation), preclinical studies, and the design and conduct of clinical trials.”

This model, Kohane suggests, “will be likely the central intellectual tool for biomedical research by the mid-2030’s.”

Goldberg, meanwhile, offers the most pragmatic and in some sense relatable perspective on how GPT-4 will be used. She recognizes that “because GPT-4 appears to be such an extraordinary tool for mining humanity’s store of medical information, there’s no question members of the public will want to use it that way – a lot.” 

Already, Goldberg writes:

“health-related web searches are second only to porn searches, by some counts. Surveys find roughly three-quarters of American adults look for health information online. It’s not hard to predict a massive migration from WebMD and old-style search to new large language models that let patients have a back-and forth for as long as they want with an AI that can analyze personal medical information and seems almost medically omniscient.”

Moreover, she adds, patients want this capability now. While awaiting access to GPT-4, she writes, “My ‘everypatient’ emotion has tended to be impatience. Sure, I get the risks of hallucinations and other little-understood distortion, but I’ve still felt mainly the AI-delay frustration of people on the waiting list for chatGPT: ‘I get it – it’s not perfect. I’ll manage – just give me access!’”

The promise of democratizing this powerful technology is easy to appreciate.  As Microsoft AI expert Karmel Allison tells Goldberg, “If the internet and the mobile age were about putting information in the hands of everyone across the planet, the AI age is about putting intelligence into the hands of everyone across the planet.”

Karmel Allison

At the same time, the co-authors all recognize that there are clearly risks involved in deploying this powerful technology. Kohane, for example, asserts “GPT-4 cannot be used in medical settings without direct human supervision.”

Kohane and Goldberg also suggest that this is the “moment for broad, thoughtful consideration of how to ensure maximal safety and also maximal access,” and emphasize the need for “guardrails to keep patients as safe as possible.” But they also recognize it’s a “tricky balance,” adding:

“those safety measures must not mean that the great advantages that we document in this book end up unavailable to many who could benefit from them. One of the most exciting aspects of this moment is that the new AI could accelerate healthcare in a direction that is better for patients, all patients, and for providers as well — if they have access.”

None of the co-authors seem to seriously contemplate pausing ongoing work in generative AI – as Elon Musk and others have recently proposed.

Kevin Scott

Collectively, the authors’ perspective seems well-summarized by Microsoft Chief Technology Officer Kevin Scott, who tells Goldberg, “The technology will exist. It will have an enormous amount of possibility. I think it will be incredibly useful and powerful. And then society has to choose how it’s going to use it.”

It’s a thrilling future to envision, to anticipate, and (as many of us hope) to help create.

Unless, of course, the technology breaks away, and chooses how it’s going to use us. 

How To Serve Man? It was a cookbook.

David Shaywitz

Biopharma, like the rest of the world, appears to be on the threshold of profound, technology-induced change. Incredible advances in artificial intelligence, manifested most recently in GPT-4, are here. 

This technology, Ezra Klein explains in the New York Times, “changes everything.”  Bill Gates describes it as “the most important advance in technology since the graphical user interface,” and declares, “the age of AI has begun.” Similarly, Times columnist Thomas Friedman argues:

“This is a Promethean moment we’ve entered — one of those moments in history when certain new tools, ways of thinking or energy sources are introduced that are such a departure and advance on what existed before that you can’t just change one thing, you have to change everything. That is, how you create, how you compete, how you collaborate, how you work, how you learn, how you govern and, yes, how you cheat, commit crimes and fight wars.”

At an entrepreneurship salon at Harvard this week, I discussed GPT-4 with Dr. Zak Kohane, Chair of the Department of Biomedical Informatics at Harvard University (disclosure: I’m a lecturer in the department), and Editor-in-Chief of the soon-to-be-launched NEJM-AI.  

Kohane received early access to GPT-4. He has just completed a book, The AI Revolution in Medicine: GPT-4 and Beyond, to be published in mid-April, about the impact of emerging AI technology on healthcare. I read an advanced, draft copy of the book. Kohane’s co-authors are Peter Lee, Corporate Vice President and Head of Microsoft Research, and Carey Goldberg, a distinguished journalist. 

From both the book and the salon, the three most striking features of GPT-4 seem to be:

1. Its ability to reason;
2. Its ability to communicate and engage with people in natural language;
3. The fact that no one really understands how it works.

How did GPT-4 impress Kohane? For starters, it performs spectacularly on standardized exams like the medical boards, and seems to be able to reason thoughtfully, Kohane says.

Zak Kohane

For example, we discussed the ability of GPT-4 to respond to an apparent paradox that Kohane says stumps the vast majority of the non-physician data scientists. The question: why is a low white blood cell count between midnight and 8 am associated with far worse outcomes than a low count between 8am and 4 pm? 

GPT-4’s top suggestion, Kohane says, was the correct answer: the issue isn’t so much the low blood count but rather the existence of a blood draw in the middle of the night. That signals the patient is experiencing some sort of medical crisis. 

GPT-4 can also provide sophisticated differential diagnoses, Kohane says, and suggest relevant next steps. 

He posed GPT-4 a question from his own specialty, pediatric endocrinology:

“I gave it a very complicated case of ambiguous genitalia that I was actually called for once back in my training. And it’s able to go through everything from the clinical presentation to the molecular biology. It had a disagreement with me and was able to cogently disagree with me, and it was also able to articulate concerns for the parents of this child and for the future engagement of the child in that discussion. So, on the surface it’s acting like one of the most sensitive, socially aware doctors I’ve ever met. But we have no guarantee that it is such.”

The ability of GPT-4 to engage in such a human-like fashion is one of the most striking, and disarming, characteristics of the technology. Many who engage with GPT-4 over time describe the sense of developing a close relationship with it, in a fashion that can feel dislocating. Kohane, Lee, Friedman and others all describing losing sleep after spending time with GPT-4. They are overwhelmed, it seems, by the power and possibility of what they’ve experienced.

 “GPT-4’s abilities to do math, engage in conversation, write computer programs, tell jokes, and more were not programmed by humans,” wrote Lee, of Microsoft Research. These capacities emerged unexpectedly, he wrote, “as its neural network grew.“

Peter Lee, Corporate Vice President,
Microsoft Research

This creates what Lee calls a “very big problem.” He writes: “Because we don’t understand where GPT-4’s capabilities in math, programming, and reasoning come from, we don’t have a good way of understanding when, why and how it makes mistakes or fails….”

The implications are dizzying. As Kohane writes, “I realized we had met an alien agent and it seemed to know a lot about us, but at the moment, I could not decide if it should be given the keys to our planet or sealed in a bunker until we figured it out.”

Implications for Healthcare

Some AI experts, like University of Toronto professor and author Avi Goldfarb, said the AI technologies will serve as a democratizing force in healthcare. He suggests, in a podcast interview with Patrick O’Shaughnessy, it will be able to “automate diagnosis,” with the consequence of “upskilling” the “millions of medical professionals” like nurses and pharmacists. The consequence, he suggests, is that:

“There’s hundreds of thousands of doctors in the U.S. and their special skill in diagnosis is going to go away. They’ll have to retool and figure out how to deal with that. But there’s millions of other medical professionals who are now going to be able to do their jobs much better, be more productive. And that upskilling provides a lot of what we see as the hope and opportunity for AI.”

I asked Kohane about this, and (to my surprise) he seemed to largely agree, albeit with a slightly different framing.  He notes that we have a crisis resulting from a shortage of  primary care doctors. Massachusetts (as the Boston Globe has recently reported) is being hit particularly hard. While we may have an idealized view of how the best primary care doctors can treat patients, Kohane argues, this is generally not the lived reality, and suggests that a nurse practitioner or physician assistant, coupled with AI, could generally offer a higher level of care for patients than a typical primary care doctor without AI. Given the shortage of trained medical professionals, AI can help improve the quality and quantity of available care.

It’s also clear that patients will have access – and, through Bing (which works with GPT-4 when accessed from Microsoft’s Edge browser), already have access – to this knowledge and information. 

Many patients and caregivers are eager for this capability, as Goldberg writes in the book. The problem is that GPT-4 (like surgeons and Harvard grads) tends to be frequently correct but rarely in doubt. It still suffers from the problem of “hallucinations,” making up information that sounds plausible but isn’t accurate. Like humans suffering from the Dunning-Kruger Effect, it can insist that it’s correct, when it’s wrong. I saw that when GPT-4  tried to persuade me that Goose, not Merlin, uttered the line “That MIG really screwed him up,” in the original Top Gun. 

For now, everyone seems to acknowledge the hallucination problem, and call for “human in the loop” approaches. But what happens as we gain more confidence in GPT-4 and, motivated by both convenience and cost, are increasingly tempted to take the human out of the loop?

Stepping Back

What seems clear is that we are truly experiencing what economist Carlota Perez has described (see here) as the “irruption phase” of emerging technology. We recognize that there’s something promising and incredibly exciting, and now everyone is trying to figure out what to make of it, and how to apply it.

These days, it feels like every healthtech person I know who hasn’t started their own VC fund (as noted here) is either starting their own health+AI company or writing a book about health+AI, and in some cases both. Microsoft Office programs are about to be boosted by GPT-4. Other companies, such as the regulatory intelligence company Vivpro – are already offering GPT-powered tools. Every consultancy is offering executives frameworks and navigation guides to the new technology.

The truth, of course, is that no one has any idea how things are going to evolve. The pharma company of the future, the healthcare system of the future, the payor of the future, perhaps even the FDA of the future: all are likely to be profoundly changed by technologies like GPT-4. 

The technology will likely first arrive as incremental, point solutions, says Goldfarb, of the University of Toronto. Eventually, however, the real productivity gains arise from more fundamental change.  The classic example here, as I’ve discussed, and as Goldfarb also cites, is that dropping electric generators into factories built around steam power didn’t have much impact. But reconceptualizing the structure of factories from the ground up, in a fashion enabled by electricity, was transformative.

As Microsoft AI expert Sebastien Bubeck observes in Kohane’s book, “GPT-4 has randomized the future.  There is now a thick fog even just one year into the future.”

What an amazing, terrifying, thrilling, and hopeful time to be alive.  For those of us in medicine and biomedical science: what an opportunity, and profound responsibility, to be in the arena, actively shaping the future we hope to create and aspire to inhabit.