An investment firm with zero biology background tried to cure scorpion stings with Claude and the outcome says more about AI hype cycles than medicine.
The Summary
- Citrini Research partnered with Anthropic to test if Claude could develop scorpion antivenom from scratch, despite admitting they "couldn't claim to know anything about AI as it related to biology"
- The experiment failed to produce anything resembling an actual drug, yielding only "promising starting points" according to experts
- This is a textbook case of vibe-coding meets capital markets: smart money chasing AI capabilities that don't exist yet
The Signal
Investment firms are now running wet lab experiments to prove AI thesisdecks. Citrini Research, working with an unreleased version of Claude, tried to develop scorpion antivenom without anyone on the team having relevant biological expertise. They published their results as if they'd accomplished something meaningful. Experts looked at the output and said: this is not a drug.
This is not about Anthropic building bad models. Claude is a capable reasoning engine for text-based tasks. The problem is the category error. Drug development requires experimental validation, protein folding simulation, toxicity screening, and iterative wet lab work. An LLM can suggest molecular structures. It cannot replace the entire pharmaceutical development pipeline.
"We wanted to use Claude to create a new drug. Here's what happened: we didn't create a drug."
The real story is why an investment research firm thought this was worth attempting, and more importantly, worth publicizing. Because the AI narrative in 2025 has detached from capability reality. Every frontier model release comes with breathless claims about scientific breakthroughs, reasoning improvements, and agent autonomy. Capital flows to companies that can tell the best story about what AI will do next quarter.
Citrini likely ran this experiment to establish credibility in AI-driven biotech investing. The logic: if we can show AI making progress on hard biology problems, our investment thesis looks prescient. The execution revealed the gap between AI's current state and the marketed vision.
Key failure modes this experiment exposed:
- Confusing text generation with scientific validation
- Assuming domain expertise can be replaced by prompt engineering
- Publishing negative results as if they were progress
The drug development process exists for reasons. Clinical trials catch the things in silico models miss. Regulatory review prevents well-intentioned amateurs from poisoning people. The entire apparatus of pharmaceutical R&D is designed to navigate the gap between "this molecule looks interesting" and "this medicine won't kill you."
The Implication
If you're building in the agent economy, watch what happens when AI capabilities get oversold to people with capital but no domain knowledge. The correction will be sharp and it will splash onto adjacent markets. Real AI progress in scientific domains is happening at places like DeepMind (AlphaFold) and Recursion Pharmaceuticals, where years of domain-specific training data meet actual experimental validation.
The gap between what AI can do and what people think it can do is widening. That gap is where capital gets destroyed and narratives collapse. Know which side you're on.