From pharma to AGI hype, and developing AI in finance: Martin Shkreli’s journey
Summary
Shkreli’s highest-conviction AI call is that drug discovery is bottlenecked by choosing the biological target, not generating the molecule. He estimates invention at roughly 10% of the process, human testing at the other 90%, and clinical trials at 50-70% of total cost; brute-force screening of 100 billion molecules is already feasible, while screening 100 trillion or 100 quadrillion may add little. The real opportunity is a “GPT pharmaceuticals” that reads roughly 36-38 million PubMed papers, ranks overlooked targets, and tells humans what to build.
His strongest example is BTK, where the valuable insight sat in published research for 15-20 years before anyone translated it into a major drug strategy. Shkreli argues that an LLM could connect such papers to clinical possibilities across every human protein, addressing the “human intelligence” constraint that forced his team of 18—and Vivek Ramaswamy’s team of roughly 80—to manually scan medical literature. “The intelligence goes in the front end,” while much of the downstream chemistry is comparatively deterministic.
The investable constraint in biopharma remains scarce targets, extreme capital intensity, and ten-year development cycles—not a shortage of molecule-generation tools. Only about 40 drugs receive FDA approval annually by Shkreli’s count, perhaps 20 of them meaningfully innovative, while competitors swarm every promising target. In his first drug company, he says he retained about 10% of the equity; software may reach product-market fit for less than a single $1 million mouse experiment. His ketamine program reinforced the risk: strong Yale data became merely better than SSRIs, not the expected “miracle,” in Phase 3.
Shkreli’s finance startup found millions of dollars in revenue after abandoning forced AI positioning and building a dense, low-latency financial suite for traders. After earlier attempts including TTS and consumer medical AI, the still-unreleased beta “went off the charts”; its intentionally crowded interface is “Vim for finance,” optimized for people spending 8-10 hours daily navigating securities. His near-term AI wedge is breaking-news interpretation: an LLM could have reacted to Hims’s GLP-1 announcement in a second, though any edge may diffuse across Wall Street within six months to a year.
He thinks AI enthusiasm may expand into “one of the bubbles to end all bubbles,” but does not think today’s market yet displays dot-com-style mania. He thinks OpenAI will go public near a $1 trillion market cap, supported by real revenue and productivity gains as companies such as Verizon or Procter & Gamble save a billion dollars here and there. He even calls 100 times sales potentially rational for fast-growing SaaS; the danger signal would be indiscriminate funding, while opaque private marks currently make price discovery difficult to assess.
The failed AI products taught him that technical novelty is not a substitute for distribution, workflow integration, or willingness to pay. TTS may be only a “billion-ish” aggregate market vulnerable to open-source local models, while his AI-doctor product drew usage but little revenue and risked competing directly with Grok and other general foundation models. His conclusion tracks the Nuance precedent: the durable business may be packaging technology into a physician workflow, not owning the base model.
On the price controversy and criminal case, Shkreli offered a categorical self-defense rather than a concession, while Biewald preserved the core moral pushback. Shkreli said he would not let a patient go without the drug, defended million-dollar pricing for a fatal pediatric disease with perhaps 500 patients, and described the fraud prosecution as “lawfare”; he says his hedge fund returned 4-5x capital and no investor lost money, while acknowledging convictions for misleading investors and maintaining innocence. He says he received a seven-year sentence, of which one typically serves roughly five, read 300-400 books, and advised Sam Bankman-Fried—whom he described as facing 25 years—that prison requires cultural flexibility: “You just have to roll with the punches.”
Deep dive
1. Shkreli learned pharma as an asset allocator, then became an operator
Shkreli’s original lens was financial: biopharma looked like a fragmented, modular collection of broadly comparable assets. A Pfizer or Merck might contain 1,000 drugs whose managers inevitably prioritize some programs over others, creating opportunities to license, acquire, fund, or otherwise reposition assets.
Trading eventually felt “tedious” and hair-pulling, while private-equity-style operating let him engage with the products, people, and industry lore. His retrospective judgment is emphatic: “I was a hundred million times better at that than I was at investing or trading.”
Although self-taught rather than formally trained in biology, he says 15 years of immersion took him through discovery, medicinal chemistry, crystallography, licensing, drug development, and public-company management. His standard for a CEO is not being the best chemist, but knowing enough to ask why a low-log-P compound would ever become useful.
2. The pricing dispute turns on access, incentives, and moral intuition
Biewald’s pushback — worth keeping: whatever the capitalist logic, being connected to a patient who cannot afford a life-saving medicine would be “so horrible.” He asked directly whether that had happened after Shkreli raised the price of the disputed drug.
Shkreli’s categorical answer was, “I’m a humanist. I would never let that happen.” He said replacement medicines were available, the company was unprofitable, he took no salary, and critics emphasized the price multiple rather than an absolute cost still far below some million-dollar orphan drugs.
His orphan-drug arithmetic used a fatal pediatric disorder affecting perhaps 500 people: if development is to exist at all, “you’re going to have to charge a million dollars a patient.” He framed the choice as a wealthy insurer absorbing 500 claims versus children having no medicine, while saying prior price increases had produced no access problem and sometimes enabled broader provision.
Shkreli never withdrew the decision—he said he would “defend [it] to the death”—but expressed regret that the controversy obscured the work. His analogy for drug design captures the attraction: write a molecule as one writes code, then “your body’s kind of the compiler.”
3. Infectious-disease preparedness was the strategy hidden behind the controversy
The disputed medicine was roughly 70 years old, yet nobody had produced a better successor. Shkreli’s team targeted subtle differences between parasitic and human dihydrofolate reductase, seeking a compound that blocked the parasite’s enzyme without making a person sick at high doses.
His broader criticism was that infectious-disease development begins only after “it’s on fire,” despite a roughly ten-year drug-development lead time. Antibiotic resistance, West Nile virus, Chagas disease, and coronaviruses illustrated markets where weak near-term economics could leave society without ready countermeasures.
He argued that a standing infectious-disease company might have had a coronavirus antiviral ready on day one of an outbreak and could hedge even speculative AI-enabled biosecurity threats. That ambition, he said, arrived before its time—and the pricing spectacle helped deepen public scorn for an industry whose cures for hepatitis C and cystic fibrosis he regards as major human triumphs.
4. Target selection, not chemistry, is the scarce input in drug discovery
Shkreli’s central claim is that the major bottleneck is deciding where to point the “clinical trial, medicinal chemistry, toxicology apparatus.” Drug programs generally begin with a protein whose function must be blocked, restored, or augmented; selecting the right protein determines whether everything downstream matters.
Dupixent was his specimen: the breakthrough was choosing the molecular target for diseases including asthma and eczema, not inventing antibody technology. In his telling, making an antibody or basic small-molecule inhibitor has been routine for skilled teams for decades.
He divided the process sharply: actual drug invention may be around 10%, while testing in humans occupies the other 90%; clinical trials alone represent roughly 50-70% of total cost. AI cannot eliminate the time spent observing patients or the analogous requirement to dose animals and await toxicology results.
Nor does he see molecule volume as the constraint. Screening 100 billion candidates serially is already feasible, he argued, while escalating to 100 trillion or 100 quadrillion does not meaningfully improve the medicine: “We’re not gated by that.”
5. A literature-reading LLM could attack the intelligence bottleneck directly
Shkreli traced his AI interest to PubMed, whose corpus he variously put at 38 million and 36 million papers. His team of about 18 and Vivek Ramaswamy’s roughly 80-person team were trying to find overlooked medicines by reading global research and deciding what to license, acquire, or finance.
His proposed loop is simple but enormous: ask GPT, for every human protein, whether inhibiting or otherwise interacting with it could treat disease; then rank the top thousand opportunities. A stronger future system could tell Pfizer or Merck, “Maybe we should make this drug.”
The BTK example carries the thesis. Research in the 1980s implicated Bruton’s tyrosine kinase in autoimmune disease, but 15-20 years passed before someone translated the finding into an inhibitor program. “All that needed to happen is somebody had to read that paper and translate it clinically.”
RNA-expression data may benefit from machine learning because the transcripts are voluminous, but Shkreli still ranks literature comprehension higher. For conditions such as Sjögren’s disease or focal segmental glomerulosclerosis, the crucial question remains what biological intervention could actually change the disease.
6. Biewald pressed the case for AI-native biotech—and Shkreli rejected its assumed moat
Biewald’s challenge was institutional: Isomorphic, Recursion, and other heavily funded, technically sophisticated organizations have convinced serious investors that AI can transform discovery. If their thesis is not empty, where exactly does Shkreli’s model diverge?
His answer was that they think discovery is harder—and more representation-dependent—than he does. Sampling 100,000 Alzheimer’s patients and connecting genetic or blood data is worthwhile, but the decisive signal may emerge from ordinary clustering or even direct inspection, such as identifying a protective APOE allele and considering a mimetic.
Huntington’s disease shows the opposite failure mode: the causal malformed protein is already known, yet delivery remains unsolved. CRISPR can reach liver cells relatively well but not brain cells; the problem is building “a better screwdriver,” perhaps through serendipity rather than extracting hidden features from more candidate molecules.
His historical analogy was Human Genome Sciences: owning a celebrated platform did not remove the requirement to make medicines that work in people. Antibodies similarly became an industry-wide modality within years, not a durable strategic monopoly; AI may follow the same path. “The tail’s wagging the dog” when chemistry tooling precedes the clinical strategy.
7. Biotech’s economics pushed him toward software
Shkreli says a regulatory and legal barrier bars him from the pharmaceutical industry, though he has considered developing residual ideas in Europe and later backing into the United States. Even without that barrier, he now finds software more enjoyable, less dilutive, and more rewarding.
Retaining roughly 10% of his first drug company felt like an exceptional outcome to biotech peers but absurdly low to software founders. A single mouse experiment can cost $1 million, while some software businesses reach product-market fit and profitability on similar resources—and founders may retain nearly all the equity.
Scarce targets intensify the capital problem. Shkreli estimates roughly 40 FDA approvals per year, perhaps only 20 “truly interesting” advances; after Keytruda and Opdivo validated PD-1, nearly every large drugmaker pursued the target, with similar herding around less successful oncology ideas such as LAG-3.
Ketamine taught him how biological noise defeats confident forecasts. Early Yale depression data looked so powerful that Phase 3 appeared easy; later results landed “a little bit south” of the miracle and somewhat north of SSRIs. In software one changes the code; with a failed molecule, “you start from scratch.”
8. His account separates the drug-price scandal from the fraud conviction
Shkreli says livestreaming gave him a route around hostile media: instead of relying on media interviews and coverage, “I’m just going to show you who I am.” Technologists who watched him program, analyze financial statements, and teach entrepreneurship became high-conviction defenders, outweighing millions of low-conviction negative impressions in his view.
The criminal case concerned fraud, not the medicine’s price. Shkreli says his hedge fund returned 4-5x capital, no investor lost money, and he personally made nothing; he was acquitted of much of the alleged conduct but convicted on charges including misleading fund investors, none of whom he says were upset.
He still maintains complete innocence and interprets the prosecution as “lawfare” after fights with politicians, including Hillary Clinton, over corporate pricing power. That is expressly his version: he also acknowledges that charges stuck and describes entering a courtroom where he believed “everything is stacked against you.”
9. Prison replaced computing with books and forced a different kind of adaptation
Shkreli’s strongest prison observation was socioeconomic rather than carceral: among thousands of people, he recalls almost no one from a stable two-parent household. He encountered severe, apparently undiagnosed learning and mental-health problems, including people unable to perform basic addition or subtraction.
With computers prohibited, he read roughly 300-400 books, sometimes one per day, including SICP—the MIT “wizard book.” On an evidence-review machine, he found a browser developer console and amused other prisoners by altering HTML to display giant jokes: “Good old JavaScript.”
He called loss of internet access almost “corporal punishment,” yet conceded the silver lining: since returning to startup life, he may have finished only one book. He rated the prospect of prison an eight or nine on his anxiety scale; his father’s death was a ten. “You only do two days, your first and your last,” went the prison saying he found partly true.
His advice to Sam Bankman-Fried was to respect unfamiliar cultures, become malleable, communicate, and “roll with the punches.” Shkreli thinks Bankman-Fried deserved punishment but not 25-30 years; his concern was that Bankman-Fried’s “anti-charisma” and determination to nerd out unchanged could make adaptation harder.
10. TTS exposed the gap between impressive models and durable markets
Shkreli’s company first explored LLMs and text-to-speech but could not find strong product-market fit. Microsoft’s Nuance acquisition shaped his conclusion: Nuance succeeded by packaging speech technology into products—especially physician workflows—not simply by possessing TTS or ASR technology.
He estimated the whole TTS market at only “a billion-ish” in revenue, much of it residing inside IBM, AWS, GCP, and Azure. Even high-quality specialists face the prospect that an excellent Hugging Face model eventually runs locally as compute and data become cheaper.
The remaining technical challenge is long-range context: sentence-level chunks may miss whether a name belongs to a villain or lover and therefore choose the wrong prosody. Shkreli still imagines personalized Disney characters or a daily SpongeBob episode, but could not convince himself that the market would become sufficiently large.
The technical immersion was real: the team rewrote parts of Torch, Shkreli submitted a pull request to Microsoft’s DeepSpeed, and they studied open-source TTS work. Greg Brockman’s maxim—that teaching software engineers AI is easier than teaching AI specialists good engineering—proved difficult in practice.
11. AI medicine attracted users but collided with reimbursement and foundation models
Shkreli saw medicine as a natural automation target because human labor constitutes a large share of healthcare costs and many protocols resemble flowcharts. Hypertension treatment, for example, can progress through specified drugs according to systolic pressure and prior response.
A Middle Eastern country told his team it could cut costs by compelling use of an AI doctor, but US deployment looked harder: clinicians defend reimbursement, insurers would need convincing, and political resistance would be substantial. He variously guessed that Medicare and Medicaid account for roughly half to three-quarters of the budget deficit, while explicitly saying he was unsure.
The consumer product generated usage but not revenue. More fundamentally, Grok or another general LLM could provide the “second opinion” itself; as Cursor can pipe into foundation models rather than train a wholly separate coding model, Shkreli doubted that a dedicated medical LLM would retain a defensible advantage.
12. The finance terminal found product-market fit by serving its founder’s own workflow
Shkreli returned to programming after his arrest, learning JavaScript and building small financial tools that accumulated into a large suite. After the AI pivots, lifelong investing supplied the market he understood well enough to obsess over every detail.
Biewald described the demo as a “monster BI tool” full of graphs, tables, and metrics, then asked whether it aided decisions or merely fed trader anxiety. Shkreli’s answer: a professional desk already looks that way, and someone working 8-10 hours needs command-line speed, not a polished SaaS page that pauses to hydrate.
The guiding phrase is “Vim for finance.” Its scattered, slightly ugly interface is a feature because switching from Apple to another security should take a millisecond; the deeper moat includes negotiating data from Frankfurt, Osaka, Tokyo, and other venues, then detecting late or broken feeds.
Earlier products ranged from no connection to usage without revenue and modest TTS traction. The finance beta, still not formally launched, has reached millions of dollars in revenue and “gone off the charts.” Team members are encouraged to invest for even ten minutes daily so they feel the user’s pain firsthand.
13. LLMs can trade news faster, but any informational edge will commoditize
Shkreli expects AI to strengthen quantitative trading and further diminish discretionary humans. People retain a small advantage in combining scattered clues—perhaps inflation signals from several companies—with market psychology, but high-frequency markets already belong to computers.
His immediate product idea is an LLM that reads breaking news, determines whether the information is real, and acts before the market finishes interpreting it. When Hims announced it would carry a GLP-1 product, he said the market took minutes to digest news an LLM could classify immediately: “Yeah, buy that.”
The caveat is brutally financial: once the technique works, everyone may have it within six months or a year. Shkreli wants to make such tools available to smaller traders too, while acknowledging that human stock picking increasingly resembles Kasparov seeing Deep Blue’s trajectory.
He rejects the Robinhood pattern of depositing each paycheck and guessing whether a stock rises or falls. The broader terminal could instead expose quantitative tools and graph recurring non-security data—weather, shipments, or other structured series—while helping users recognize when they have no real edge.
14. The AI bubble may grow much larger because the technology is already useful
Shkreli’s position sounds contradictory only because it is conditional: this could become “one of the bubbles to end all bubbles,” yet current behavior still looks relatively rational. Like the dot-com cycle, an apparent plateau might precede another large upward gap.
His specific forecast is that OpenAI goes public around a $1 trillion valuation. Unlike many dot-com companies, today’s leaders have revenue, working products, and immediate enterprise impact; Verizon, Procter & Gamble, and peers might each save a billion dollars here and there through productivity improvements.
He does not treat 100 times sales as automatically irrational for a fast-growing SaaS startup. The real warning would be indiscriminate mania—such as a pre-product robotics company raising at $100 billion—although even that example produces hesitation because replacing human labor might support an enormous eventual market.
Private capital now clouds the diagnosis: giant venture and growth funds can assign “crazy prices” while leaving price discovery unclear, whereas public markets impose a reality check. A nominal $5 billion company can disappear when sponsors stop funding it, revealing that the valuation was “just a mark.” His honest horizon: the bubble will be easier to identify five years later.
15. The Wu-Tang purchase was a thesis about patronage, not just a trophy
Shkreli framed the Wu-Tang album as an experiment in paying musicians to make uncompromised art, analogous to historical patronage. Rather than another Picasso or summer house, a founder could commission Taylor Swift or Drake to create for one patron without optimizing commercial output.
His other imagined trophies—a $40 million dinosaur skeleton, early Apple computers, historical letters—share the same principle: wealth should fund objects meaningful to the buyer. Biewald closed on Wu-Tang’s pull among technologists; Shkreli added that one analysis placed the group’s vocabulary “many sigma” beyond most rappers, alongside unusually dense metaphor and mythology.