Curing Every Disease With Al by 2050 | Sam Rodriques, Edison Scientific
Summary
Rodriques’ core thesis is that biology is constrained by scarce scientific talent, while “capital scales, and logistics scales, and talent just does not scale.” AI can relieve that bottleneck through high-throughput reasoning: reading more evidence and testing more hypotheses than any researcher could. It cannot eliminate physical validation, especially when the decisive experiment is a multiyear human clinical trial.
FutureHouse’s first full-loop agent generated a treatment hypothesis for dry age-related macular degeneration that progressed from wet-lab validation into animals and, two days before the interview, publication in Nature. Rodriques calls that May 2025 result the moment “the future is here.” Its successor, Cosmos, has probably been used to generate 20,000-30,000 novel findings, though that figure describes findings proposed rather than approved medicines delivered.
Biewald frames the commercial tension as “progress comes from discovery, but the commercial value is in development.” Rodriques says discoveries pan out infrequently and may not reveal their value for a decade, while faster experiments can move medicines toward patients. AI can help prepare protocols and regulatory documents, coordinate trial sites, and move compounds through the pipeline; molecule-design specialists can simultaneously improve candidates. Rodriques says future pharma companies will be much leaner and pursue many more programs with the same headcount, but human trials remain the binding physical and operational constraint.
Edison Scientific’s proposed moat is specialized scientific reasoning, customer-specific training, and last-mile deployment—not an attempt to outperform frontier labs at everything. Rodriques says small amounts of task data can produce “enormous gains” over general models in niches such as synthetic chemistry, while proprietary pharma data preserves customer differentiation. His hedge matters: specialization may lose its advantage if intelligence reaches task saturation, and coding has not yet shown that a specialized model can beat the strongest generalists.
The platform deliberately exploits the frontier models’ non-overlapping strengths instead of betting on one provider. Rodriques says Anthropic led Edison’s evaluation for reproducing analyses from scientific papers, while only Gemini 2.5 could perform one early Cosmos world-model task. This multi-model architecture also answers pharma’s reluctance to become locked into OpenAI, Anthropic, or Google while leadership keeps changing.
Rodriques pairs enthusiasm for faster experimentation with deep skepticism toward uncontrolled biohacking and a provocative case for clinical-trial reform. Unstudied peptides may not work, may merely produce placebo effects, or may cause untracked harm; “no one is looking” for adverse events among self-dosing communities. Yet onerous trials help create that behavior, so he favors decentralized early-stage approvals like those he describes in Australia and China and, conditionally, loosening efficacy requirements after safety is proven.
Despite declaring that this AI cycle is different, Rodriques sets a hard industry scoreboard at pivotal Phase 3 trials and FDA approvals. AI drug discovery has overpromised since roughly 2012, with AlphaFold the notable exception, and pharma insiders will discount announcements that a model “came up with” a drug. “The proof is in the pudding, and the pudding is approved drugs.”
Deep dive
Not yet available upstream; scheduled sync will retry.