I’ve been watching the AI drug discovery narrative unfold with growing skepticism. When Citrini Research and Anthropic announced they’d used Claude to design proteins for scorpion antivenom, the headlines screamed revolution. The reality, as reported by The Verge, tells a much more cautious story that should concern anyone building or investing in AI-powered biotech tools.
Let’s be clear: the experiment failed. Citrini set out to create a new drug and ended up with some protein candidates that bind to toxins. That’s not a drug. It’s not even close. Yet the firm celebrated it as a success, and Anthropic, preparing for its IPO, seemed happy to amplify the narrative. This is exactly the kind of strategic misrepresentation that erodes trust in AI development.
The Gap Between Discovery and Medicine
The technical experts quoted in The Verge’s reporting make the key distinction crystal clear. Andreas Hougaard Laustsen-Kiel from Technical University of Denmark noted that finding binders against toxins is “easy-ish,” but “the hard part is to translate it into a medicine.” This isn’t a minor distinction. It’s the entire distance between a promising experiment and a viable treatment.
When you’re designing systems for AI and machine learning, this gap matters enormously. Claude generated nearly 320,000 protein candidates, which sounds impressive until you realize that’s just the beginning. Those candidates need to be tested for safety, stability, and efficacy in biological systems. They need to work reliably across different individuals. They need manufacturability. They need regulatory approval. Only 17 of the proteins even passed initial binding tests, and none bound to more than a single toxin.
That’s not a 5% success rate worth celebrating. That’s a reminder that computational discovery, however sophisticated, remains separated from real-world medicine by valleys of complexity that AI hasn’t yet bridged.
Why Investment Firms Push Narratives
Here’s what bothers me most: Citrini isn’t a biotech startup trying to solve real problems. It’s an investment research firm with financial incentives to overhype AI’s capabilities. The firm previously published fictional accounts of AI-induced economic catastrophe that moved markets. Now they’re making bullish cases about drug discovery, with specifics reserved for paying subscribers.
And let’s not ignore the enabler here. Anthropic, preparing for its IPO, has reasons to showcase AI’s potential in life sciences. It provided unreleased models, covered nearly $50,000 in compute costs, and gave the team access to its offices. When you’re about to go public, nothing helps stock valuations like headlines suggesting your technology is revolutionizing drug development.
This is the intersection where AI hype meets financial incentives. As developers and engineers, we need to be especially vigilant about these narratives.
What Actually Happened Here
The experiment demonstrated something real but limited: AI can rapidly explore chemical and protein design spaces in ways that would take humans much longer. That’s valuable. High-throughput screening and computational protein design aren’t new, but orchestrating them via AI agents is genuinely novel.
But here’s the thing: the novel part isn’t the biology. It’s the orchestration. The autonomous discovery process is interesting. The proteins it found? That’s ordinary screening yielding ordinary results. Experts described the findings as “technically promising but premature,” and “far from a medicine and hyped.”
The testing alone cost $150,000 and required external labs. That’s after Anthropic’s subsidized compute budget. The entire endeavor felt designed to generate headlines and investor interest rather than to actually solve the antivenom problem.
What This Means for Developers
If you’re building AI systems for scientific discovery, understand the limits of what you’re optimizing for. LLMs and AI agents are excellent at searching solution spaces and proposing candidates, but they’re not architects of causality. They can’t guarantee that a protein will behave predictably in a living system. They can’t navigate regulatory frameworks or manufacturing constraints.
The future of AI in drug discovery probably involves AI as a powerful tool within a much longer process, not as a replacement for the process itself. The firms that will win are those honest about this limitation while still leveraging AI’s real strengths.
The uncomfortable truth emerging from Citrini’s experiment is that groundbreaking results require actual groundbreaking work, not just impressive compute budgets and press releases. When a company claims success but won’t actually develop its findings further, that should raise immediate red flags about whether the success was ever real.