The financial valuations in AI drug discovery are inflated, but the underlying science is not a bubble at all. That is the distinction Ed Addison draws on this episode of Unscripted, and it comes from someone with unusual standing to make the call. Addison took his first AI course in 1985, has founded five AI companies with three exits, and has spent the last decade and a half specifically inside AI drug discovery, one of the hardest and most capital-intensive corners of the industry. His answer to what's actually working and what's just FOMO is more specific than most takes on the AI investment cycle.

Who is Ed Addison?

Ed Addison has been building AI companies since before most people believed the field existed. He took his first AI course at MIT in 1985, lived through the rule-based expert systems era of the late 1980s, the neural net wave of the 1990s, and the deep learning breakthrough around 2006 that made modern computer vision and speech recognition possible. In 2011 he founded Cloud Pharmaceuticals, one of the earliest AI drug discovery companies, which ran for a decade before being split into three pieces during the pandemic for financial reasons. He has also founded companies in bioinformatics and clinical trial enrollment, and now advises four or five AI drug discovery companies while teaching engineering management. Outside of AI, he plays improvisational jazz piano, flies planes, swims, and recently became a first-time novelist.

What has 40 years in AI actually taught him about what works?

Addison's career maps almost exactly onto AI's major paradigm shifts, and each one taught him the same lesson from a different angle. Rule-based expert systems, the dominant approach when he started, turned out too brittle for real users by the late 1980s. Neural nets in the 1990s topped out around 98 percent accuracy, which sounds impressive until you're processing speech or language at scale. Deep learning arrived around 2006 once computational power and internet-scale data sets caught up with the theory. Generative adversarial networks followed in 2014, the transformer architecture in 2017, and it took until 2022 for that architecture to produce a usable large language model. His conclusion after living through all of it: natural language is best processed statistically, not through rules, and expertise is now something you train into a model rather than program into it by hand.

Why does he say algorithms are a dime a dozen in AI drug discovery?

This is the sharpest reframe in the conversation. Addison argues that the algorithm, the part every AI drug discovery company markets itself around, is the least valuable piece of the business. Data is what matters, and Big Pharma already owns most of it. Beyond data, the only thing that actually moves a company's valuation is a drug that has reached the clinic. Investors aren't pricing these companies on the sophistication of their models. They're pricing them on pipeline. The algorithm might win the first round of venture funding and generate the initial buzz, but it stops being the story the moment a company needs to prove it's worth its valuation.

Is AI drug discovery in a bubble?

Addison's answer splits the question in two. He does not think AI itself is a bubble, generative AI is still early and the underlying science has a long way to go before it's exhausted. But he does think the financial valuation of AI, and AI drug discovery specifically, is inflated by fear of missing out. He points to companies like Xaira, which raised a billion dollars in seed funding, as exactly the kind of valuation an angel investor should avoid: the higher the starting valuation, the lower the probability of a return that justifies the risk. He compares the current moment to the dot-com crash, where valuations fell 80 percent before slowly climbing back, and expects something similar, though likely more muted, in AI drug discovery given the longer timelines involved.

What would actually change the economics of drug development?

According to Addison, the number that matters most isn't speed or cost, it's failure rate. The widely cited $2.6 billion cost to bring a drug to market is calculated by dividing total industry R&D spend by total approvals, meaning every failed project across the industry is baked into that number. If AI genuinely reduces how often drug candidates fail in the clinic, the economics change dramatically. The problem is that no AI-discovered drug has completed FDA approval yet, so the industry is still operating on belief rather than proof. Addison sees two paths where AI could eventually matter more than incremental speed gains: precision medicine, where AI-driven patient stratification and drug repurposing could make multi-drug treatment economically viable in a way it currently isn't for Big Pharma, and a longer-term shift from targeted, on-off therapeutics toward drugs designed as control functions that modulate biology the way an optimal control system would. He puts that second shift ten to fifteen years out.

Why did an AI drug discovery pioneer write a novel about robots gone wrong?

Addison's book, "Probability of Doom," started as an unplanned conversation on a five-hour car ride home from an amusement park with his daughter and grandsons. The story they invented: a fictitious company deploying humanoids worldwide, a disgruntled chief robotics engineer who gets fired after a personal betrayal, and a drunken decision to reprogram every humanoid on the network with guardrail changes that were never meant to be malicious but spiral into something none of the characters intended. Addison wrote the novel over five months with ChatGPT, Claude, and Grammarly as collaborators, drawing on his background in complex systems and the principle of emergence, the idea that a system can be built correctly and still produce behavior nobody predicted. The book was never meant to be a commercial project. He now uses it in his own classroom to teach engineering students why guardrails have to account for insider misuse and emergent behavior, not just external attacks.

FAQ

What is the most valuable asset in AI drug discovery, according to Ed Addison? Data, not algorithms. Addison argues algorithms are replaceable and easy to come by, while Big Pharma's ownership of clinical and biological data is the real competitive advantage in the space.

Is AI drug discovery a bubble? Ed Addison distinguishes between the science and the money: he believes the financial valuations in AI drug discovery are inflated by FOMO, but the underlying science still has a long way to go and is not itself a bubble.

Who is Ed Addison? An AI entrepreneur who took his first AI course in 1985, founded five AI companies with three exits including Cloud Pharmaceuticals, one of the earliest AI drug discovery companies, and now advises multiple drug discovery startups while teaching engineering management.

What is "Probability of Doom" about? A novel Addison co-developed with his family and wrote with AI collaborators, about a disgruntled robotics engineer whose guardrail reprogramming spirals into humanoids worldwide exhibiting dangerous emergent behavior, written as a teaching tool for AI engineers rather than as commercial fiction.

Why does failure rate matter more than speed in drug discovery? Because the industry's commonly cited $2.6 billion average cost per approved drug already includes every failed project's spending. Reducing how often candidates fail in the clinic would lower that number far more than making the discovery process faster or cheaper.

Watch the full conversation and explore more episodes at Unscripted with Jeff Pedowitz.