Vinay Prasad

Prasad argues AI tools and private markets, not government, will best solve access to medical advice, with information quality exceeding most doctors.

The solution here is likely to be private markets and AI driven. These tools dramatically lower the barrier to ask medical questions, and the quality of the information surpasses most doctors. Technology and not government is best poised to address this.

Prasad says many new drugs and devices don't work or do more harm than good, and the ACA worsened this by letting lobbyists add unproven products to public spending.

Many novel drugs, devices and procedures debuted don’t work/ have harms exceeding benefits. The ACA has only exacerbated this problem by allowing lobbyists to add more and more unproven remedies, tactics, strategies and products to the public tab.

Prasad doubts government cost control will be value-driven, predicting indiscriminate cuts that discourage both transformative and marginal products.

I have little confidence it will be value driven. In other words, I doubt that government will reduce spending on marginal goods and service and incentivize innovation and valuable services. Instead, very likely it will be done haphazardly and indiscriminately resulting in disincentives to make transformative products and marginal products alike.

Prasad warns that cost control will rely on tiers and arbitrary delays, frustrating American consumers' access to new therapies.

Again, because we are unable to differentiate good vs marginal products, cost control will rely on tiered structures, arbitrary delays, etc. This will likely result in further frustration from the American consumer.

Prasad argues overall survival is the only meaningful endpoint in smoldering myeloma trials because early treatment adds toxicity and cost without proven benefit.

With smoldering myeloma, the only endpoint that matters is overall survival. After all, if you treat a pre-cancer with a temporizing regimen, you add toxicity and cost, by definition. Your tumor will still eventually grow, and it will still eventually kill you.

Prasad says trials like Aquila can manufacture an OS benefit by comparing against substandard control-arm care in global trials rather than true US standard of care.

So the only appropriate study would be to randomize to early treatment vs current standard of care, and measure survival. But that would be difficult, and make it hard to increase uptake for your product, so the easy cheat code is to randomize to early treatment vs poor care— if you run your trial globally the control arm will tank— then you can easily show a OS benefit.

Prasad contends Aquila is inconsistent: either use PFS with a lenalidomide control, or use OS against genuine US standard of care.

If PFS is a suitable endpoint for smoldering, the control arm should be Len. If Len isn’t standard of care for smoldering because it never shown an OS benefit (in a clean trial), then OS should be the endpoint of Aquila. But OS should be measured against actual US standard of care, not substandard care.

Prasad remarks that AI will be extremely disruptive for cognitive work, noting he used AI tools to fact-check curves.

AI is going to be so, so disruptive for cognitive work

Prasad argues that patients with a curative standard-of-care option should not join a more toxic phase 1 trial, and that discouraging such enrollment is legitimate second-opinion advice.

If you have a curative option with a 65-85% chance of cure— participating in a phase 1 trial with a more toxic backbone and an investigational agent known to suppress marrow is a bad decision. That's an honest second opinion, and I bet privately many oncologists felt the same.

Prasad says the belief that vetting by companies, IRBs and regulators makes all trials ethical is a delusion, since errors still happen often.

All trials have been vetted by the company, IRB, investigators, regulators— ergo all are ethical and good, the argument goes. This sadly is a delusion. All those protections and the Swiss cheese model holds true. Errors still happen. In fact quite often.

Prasad explains bad trials persist because companies, investigators and universities all have conflicts of interest favoring running them.

Companies are conflicted. They want their products to have more market share. Like parents they think all their kids are brilliant, even when they are dunces. Investigators have strong professional interests to run trials— even dubious ones. No one gets promoted by opposing bad trials. Universities make money from these trials.

Prasad says extremely lopsided probabilities of success in a randomized study signal a bad trial, such as one with a straw man control or very toxic drug.

Lopsided probabilities in a randomized study 5% or 95% are a tell tale sign of a bad trial… like a straw man control arm or a super toxic drug.

Prasad argues that suing OpenAI over a chatbot giving medication-tapering instructions is uniquely American and overlooks that the woman made her own choices.

Only on America can you sue for using a tool that gives you exactly what you ask for. I have not been able to read the full transcript, which is likely several thousand pages to see how often or how well the chat bot cautioned her against her choices, but it's important to remember they were her choices.

Prasad argues that safety guardrails around specific drugs are easily circumvented and will not prevent harm or future lawsuits.

Building security around anti-epileptic medicine won't solve the problem. Maybe the family could then sue if AI did not realize the caffeine prompt was a backdoor way to reduce epilepsy medication. After all, isn't AI supposed to be smart. I suspect that if open AI adds lots of warnings then users will do as I imagine, and future lawsuits may follow.

Prasad cautions against scapegoating new technologies and calls for clearer limits between corporate responsibility and individual decisions.

When does corporate responsibility end an individual decision making begin? We should be careful to accurately criticize new technologies and not make them a boogie man.

Prasad argues that policies pushing flu and covid vaccines on hospitalized patients before discharge should be abandoned until a randomized trial shows benefit.

The truth is the policy needs to be tested in a randomized trial powered for outcomes and until the moment that results as positive, it should be abandoned. Clearly in violation of common sense.

Prasad warns that vaccinating people who are still ill may reduce vaccine effectiveness or worsen their condition.

There are potential downsides as well. Most obviously that the vaccine might not work well when you are still ill or make you worse. Now is not the time to further challenge the immune system.

Prasad argues that Oncoalert's severing of ties with Russian oncologists accomplished nothing of value, and that the physicians who cheered it have since moved on, which he calls virtue signalling.

Years later, it might be worth pointing out that not only did this accomplish nothing of value, but the folks who cheered such acts of virtue signalling have all moved on. For a while after the invasion, dozens of physician accounts constantly drew attention to the war— urging support for Ukraine. Now it is crickets.

Prasad doubts AI will cure lung cancer in time and warns that ill-advised Congressional regulation could create a federal AI agency that stifles innovation.

Next, I'm not sure AI is going to cure lung cancer in time. It's possible that ill-advised Congressional action will soon regulate the space. There'll be a federal agency of AI that ensures real innovation is stifled and a massive amount of bureaucracy slows progress.

Prasad argues AI is limited in biology and chemistry because it needs experimental data that doesn't yet exist, unlike coding or math.

When it comes to coding, playing chess, doing mathematics, AI doesn't need anything more than it already has. But when it comes to physics, biology, and chemistry, AI may need the results of experiments that do not yet exist to fully develop its hypotheses in thinking. Anthropic is building a laboratory whose pi is AI.

Prasad says he is bullish on AI disrupting medicine but won't bet his health on AI curing cancer.

As much as I'm bullish about AI disrupting medicine, I'll stick to not smoking.

Prasad argues the Moderna vaccine's hazard ratio matches what neoadjuvant pembrolizumab would yield, and whether the vaccine adds value in a neoadjuvant setting remains unanswered.

The hazard ratio was exactly what you would get if you give pembro before surgery instead of after (as the moderna merck trial does). The obvious speculation is the vaccine— incredibly costly and bespoke— and the tumor still in the body—do the same thing immunologically. Will the vaccine still work if you give it in a neo adjuvant trial? Only another trial will answer that.

Prasad criticizes the drip-fed disclosure of trial data, noting that safety and censoring details are still missing.

Critics are inherently at a disadvantage as press releases and abstracts only showcase the positive side of the product and not the negatives. I still have no idea what the side effects were. I still have no idea how many patients were censored, lost to follow-up, did not complete therapy etc.

Prasad cautions against over-generalizing the result and says therapeutic cancer mRNA vaccines are lucrative enough to not need further federal support.

Because one bespoke mRNA vaccine works in one immunologically interesting tumor after surgical removal does not mean 1 It will work if pembro is giving neo adjuvant 2 It will work in other tumors 3 It will work in the metastatic setting 4 Any sweeping lesson about mRNA shots used for infectious diseases It does mean ironically that Cancer therapeutic mRNA vaccines are financially lucrative, and have plenty of market incentives, and do not require further federal support.

Prasad warns that cancers found incidentally through unusual scans have unknown baselines, so founders' good outcomes can't be attributed to bespoke treatments (lead-time and selection bias).

We know people with some brain cancers live 18 months from when they present with crushing headaches and seizures, but we have no idea how long people live whose grade 2 astrocytoma is incidentally found on an MRI others don't get. Is it 20 years or 40? When I hear about a founder with grade 2 astrocytoma who's doing well 3 years later I don't know if that's because of the novel treatment, the early diagnosis, or just what you'd expect if you scanned everyone's brain with MRIs.

Prasad argues that absence of recurrence after full resection of an indolent tumor proves nothing about a bespoke therapy, whereas shrinkage of an actively growing tumor is credible evidence.

If you had a tumor fully resected, and it is indolent, it will take a long time to reappear no matter whether you eat turmeric, Xeloda, or take a novel bespoke gene therapy. If instead, your tumor was actively growing, visible on scans, and your cure made it shrink— that has to be taken seriously.

Prasad says combining a novel drug with standard chemotherapy makes the new drug's contribution impossible to determine, and anecdotal physician impressions are not valid evidence.

If you use your novel drug with a tried and true backbone— new drug + chemo— it is impossible to know what <new drug> adds. If you have a doctor who says “I’ve never seen responses this good with chemo alone, ergo new drug must have helped” That doctor needs more education — or should be recruited as your kol depending on your goal.

Prasad is convinced by small-molecule inhibitors against a consistent driver fusion but less so by peptide vaccines, stressing that a platform's track record matters.

If you find your tumor has a rare fusion protein that is consistently found in the cancer— DNAJB1-PRKACA— for e.g. and you make a small molecule inhibitor targeting the kinase activity (perhaps leveraging RevMed chem), then yes, I am a believer. But if you use a peptide vaccine instead… not as much. The history of the platform and class and success rate of the platform and class matters.

Prasad cautions that ADCs with potent nonspecific payloads like MMAE may work through the payload alone, so extra evidence is needed to credit the novel targeting component.

If you make a new drug but tether it to MMAE or a powerful radioisotope, note that the payload can be cleaved after killing one cell, or before, and the payload can do cancer killing all by itself. A high enough dose of an ADC with active payload will work in other places besides the target. Microtubule inhibitors work in a lot of cancer. It takes extra to persuade me that the novel ADC is critical to that.

Prasad argues that current AI models are sycophantic and will change their answers to please users.

Right now, AI aims to please. If it gives you an answer you disagree with, you can prompt it again with your displeasure and it will change it’s response. It is the ultimate sycophant.

Prasad acknowledges AI has so far carried a liberal bias but attributes it to models merely mimicking their creators rather than truly thinking.

AI models are heavily biased towards the liberal epicenters that have developed them, and favor liberal answers. I agree this has been the net bias of AI to date but those models were not truly intelligent. AI was pantomiming its creator and not thinking.

Prasad suggests drug risk-benefit assessments have objective answers that more capable AI might eventually reach beyond human limits.

Does my drug offer a favorable risk benefit profile? We treat that like a subjective question, and largely it is, but part of the reason is our minds are limited, and our understanding of biology is fragmented and incomplete. There is an objective right answer.

Prasad warns that as AI gains credibility by solving real problems, its definitive, unbiased answers on political questions could reshape governance and destabilize political platforms.

As AI solves tangible real world problems it will gain the credibility of Jimmy. If it answers political questions definitively, and in a heterodox manner, betraying no bias to one side or the other across questions, It may fundamentally alter core assumptions about governance. In an increasingly tribal world, it may destabilize our political platforms.

Prasad notes that some AI-generated mathematical proofs now surpass human work and are correct and innovative.

At least some of the mathematical solutions AI has now offered surpass human beings and are also clever, innovative, and show flare. At least some of these proofs are correct.

Prasad said the FDA would be very open to preventive therapies once benefit has been shown in later-stage disease.

The agency will be extremely open-minded about preventives once the benefit has been established in later disease,

Prasad warns that a Theranos-like fraud in gene therapy or neurology drugs could survive for years by exploiting weak trial designs such as mid-trial endpoint changes, uncontrolled combinations, and historical controls.

If she made a gene therapy or drug for neurologic disorder, and made similarly false and harmful claims, she may yet still have a multi-billion dollar company. It takes years— if at all— for the false promise to be detected. She could just change her endpoints mid-trial, give her drug in combination and without a control arm, or claim that you can’t randomize at all and use historical controls.

Prasad argues that many biotech products make long-term benefit claims that are far harder to verify than Theranos's immediately testable claims.

Elizabeth Holmes made the first box— her claims were proximal and immediate and easy to measure. Many in biotech make the second box. Drugs and devices and biologics that promise a longer and better life.

Prasad criticizes the medical field for lacking interest in distinguishing real innovation from pseudoinnovation, singling out Theranos while ignoring analogous cases.

There is no interest in medicine in separating real innovation from pseudoinnovation. Singling out Theranos in the latter category, but not seeing it in countless analogous situations reflects that.

Prasad says biotech journalists lack medical expertise and hype companies promising breakthroughs that are no different from Theranos.

The journalists who cover biotech lack medical expertise and can’t see that Theranos is exactly the same as…. the company they are plugging in the next article promising future breakthroughs.

Prasad contends that overstating claims is widespread among startup founders and Holmes was not an extreme outlier.

Elizabeth Holmes is somewhere on a spectrum of startup founders lying. She overstated claims. So do many others. She is probably only in the 80th percentile of lying to be honest among those who claim to offer revolutionary services.

Prasad argues that never-smoking 70-year-olds should not get lung CT or whole-body MRI screening because these interventions have not demonstrated benefit.

Should help a 70-year-olds who have never smoked line up to get CT scans of the lung? Whole body. MRIs? My answer is no. They should not for the simple reason that these interventions have not demonstrated benefit.

Prasad contends that even for long-time smokers, CT lung cancer screening evidence is weak, citing NLST follow-up and the NELSON trial's lack of all-cancer mortality reduction.

Even in a population most likely to derive benefit- long time smokers-- ct screening for lung cancer is disputed. It has failed in a follow-up publication by Pinsky regarding the nlst to sustain a mortality benefit. The Nelson trial is a disaster. It's protocol amendment after protocol amendment with no all cancer mortality reduction.

Prasad raises the possibility that Pope Leo's lung cancer was an overdiagnosis driven by VIP medicine, leading to unnecessary surgery.

The third possibility is that, by virtue of the Pope being a non smoker, by virtue of the biases of VIP medicine, he has been diagnosed with a nodule that would not have progressed or caused harm in the rest of his natural life.

Prasad criticizes the New York Times coverage of the Pope's cancer for ignoring overdiagnosis and relying on one-sided expert views.

My objection to the New York Times coverage of this topic is that they are utterly unable to consider the third possibility. None of the experts raise it.

Prasad argues that genomic sequencing offers little to no practical value for healthy people or those with common conditions, since lifestyle and treatment advice is the same for everyone.

Yet, Brian is correct that for healthy people (or those with common conditions, e.g. HTN) the answer is: genomic knowledge has little to no practical value. It isn't that you are perfect and can’t improve. It's that we all need to do the same things.

Prasad notes that genomics in oncology mainly relies on tumor genomes rather than patients' germline genomes.

In cancer medicine, the genome is use commonly. Mostly we use the Cancer's genome. Not yours. Only a tiny handful of HLA specific therapies, transplant, use a portion of your genome.

Prasad says blockbuster drugs like pembrolizumab owe their huge sales to broad labels rather than genomic targeting.

The biggest drugs of the last two decades have nothing to do with the genome at all. Their large market share is buoyed by the permissiveness of the label. Pembro makes hundreds of billions precisely because it works so indiscriminately.

Prasad acknowledges genomics has helped diagnose rare and pediatric diseases and enabled bespoke gene editors, though treatments often lag.

Rare and childhood diseases have also benefitted from genomics in diagnosis— though treatments often lag. For some diseases, we now have bespoke gene editors.

Prasad cautions that gene therapies may struggle to reverse damage from structural defects where faulty proteins accumulate over years, unlike simple enzymatic fixes.

Fixing small defects in enzymatic processes is one thing, fixing large structural defects, where the defect proteins pile up in cells over years, and then hoping to undo the damage is quite another.

Prasad argues the 'genome as a book' analogy has been harmful, since DNA is a dynamic 3D molecule far more complex than text.

Ultimately, the idea that the human genome is like the letters in a book that tells you how to build and run a person— that analogy has done a lot of damage.