Key Takeaways

  • Dario Amodei predicts artificial intelligence will cure most human diseases in the coming years, while Martin Shkreli predicts zero clinically meaningful AI-discovered drugs will reach the market over the next decade.
  • Heavy computation in biotech is not new: Shkreli built a Cray supercomputer cluster years ago to analyze drug molecules and now runs models on GPUs, yet the physical constraints remain unchanged.
  • Shkreli tested natural language processing at scale in 2015 by building a company around IBM Watson to index 40 million published medical research papers, finding that computational reading failed to solve drug discovery.
  • The FDA regulatory pipeline requires physical proof of safety and biological efficacy that algorithmic modeling cannot short-circuit.

The Disagreement

Silicon Valley treats drug discovery as an information problem. Shkreli treats it as a physical biology problem that software cannot compress.

Dario Amodei, the CEO of Anthropic, argues that artificial intelligence will soon cure most human diseases. Tech founders view biology as code: map the protein target, optimize the molecular binding affinity on a computer screen, and synthesize the pill.

Shkreli rejects that premise entirely. “AI does not help drug discovery,” Shkreli said. “If you've ever built drugs, and raise your hand if you've made a drug, gotten FDA approval for a drug. Okay, I'm pretty sure that's it. You know, it's pretty hard.”

His pushback centers on a specific timeline prediction. While Amodei promises rapid cures, Shkreli takes the short side: “He also said that we'd cure all diseases. I don't think there'll be a single AI relevant drug in the next 10 years.”

To Shkreli, software executives confuse computational power with biological reality. “We've had computers for a long time,” he pointed out. “I built a Cray cluster to analyze drugs. Now I can use GPUs. It's great, but there isn't much AI can help us with in drug discovery.” Faster chips do not eliminate wet-lab biology, animal models, toxicity screens, and multi-phase FDA clinical trials that eat up 90 percent of a drug's capital and time.

Who's Right (and When They're Wrong)

Shkreli has the regulatory and biological reality on his side, while the AI optimists have the early-stage pipeline argument.

Software founders consistently underestimate the sheer friction of physical biology. Shkreli learned this barrier firsthand in 2015 when he attempted to automate biomedical research workflows. “Our biggest bandwidth problem was reading these papers,” Shkreli explained. “So I started a small NLP company on the side in 2015. We had Watson. It was great. We tried to read every paper that was published in medicine. It was 40 million papers. And having humans do that is a really tough thing.” Ingesting medical literature did not translate into creating a viable molecule that survives human metabolic clearance and passes FDA scrutiny.

Where Shkreli is right: The next decade belongs to physical laboratory constraints. An algorithm can generate ten thousand theoretical binding targets in ten seconds, but researchers still have to physically synthesize those compounds, test them in cell cultures, inject them into animal models, check for liver toxicity, and recruit human patient cohorts. If an FDA Phase 2 clinical trial requires three years to measure patient survival, GPU clusters do not accelerate that calendar by one hour.

Where Shkreli is wrong: He assumes the boundary between computational screening and physical biology is static. Machine learning will not replace human clinical trials by 2035, but it can narrow the candidate funnel before wet-lab testing starts. If software cuts early-stage synthesis failures by twenty percent, pipeline efficiency improves, even if the drug still takes seven years to clear the FDA.

What to Do With This

If you run a startup targeting a physical or regulated sector, audit your development pipeline this week. Identify every milestone that depends on a physical gate like wet labs, hardware certifications, or regulatory boards. Calculate the exact cycle time of those physical processes, and eliminate any pitch deck assumption that software speed alone can compress that calendar.