Decoding AI — April 2026
Pharma doesn't trust one expert opinion. Why are we trusting one AI model?
Decoding AI — Karpathy's LLM Council pattern, and why multi-model deliberation maps onto how pharma already governs hard decisions.
Andrej Karpathy released a weekend project late last year called LLM Council. The architecture is simple. A query goes to multiple frontier models in parallel. Each writes a response. They peer-review each other anonymously, ranking blind. A designated Chairman model synthesises the final output.
Andrej Karpathy released a weekend project late last year called LLM Council. The architecture is simple. A query goes to multiple frontier models in parallel. Each writes a response. They peer-review each other anonymously, ranking blind. A designated Chairman model synthesises the final output.
It's the same logic pharma has run for decades.
Medical advisory boards. Scientific committees. Adjudication panels for regulatory submissions. We don't accept a single voice on an indication strategy or a benefit-risk profile. We triangulate, we challenge, we record disagreement, we synthesise.
Now run that thought through your AI roadmap.
If you're using Claude or GPT or Gemini as a single oracle for regulatory interpretation, competitive intelligence, or evidence synthesis, you're inheriting one model's blind spots, one training corpus, one reasoning style. MIT research on debating LLMs shows accuracy gains of 7 to 45% when models critique each other versus answering alone. That's not a marginal upgrade, that's a different category of output.
Three places the pattern earns its cost in pharma: Regulatory grey-zone interpretation, where vendor concentration risk is real and audit trails matter. Strategic option generation in commercial planning, where averaged single-model answers strip out the contrast you actually need. Medical writing QC, where anonymised peer review across models surfaces issues a single model would confidently miss.
The interesting shift is governance. The Council pattern formalises disagreement as signal. That's exactly what compliance, risk, and quality functions have been waiting for from generative AI.
Single-model AI was the prototype. Multi-model deliberation is the production system.