Artificial intelligence is approaching a fundamental turning point: moving from systems that work within existing knowledge to systems capable of expanding that knowledge through genuine discovery. This transition from AI agents to what experts call "AI innovators" represents the next major inflection point in the technology's evolution, with profound implications for biomedical research and drug development.

What's the Difference Between AI Agents and AI Innovators?

The distinction matters enormously. Current AI agents, no matter how sophisticated, operate within the boundaries of existing evidence. They can synthesize complex information, expose inconsistencies, and improve decision-making by analyzing what is already known. But they cannot generate the results of experiments that nature has yet to run. AI innovators, by contrast, would be capable of originating novel and falsifiable hypotheses, identifying which experiments are most likely to distinguish between competing explanations, and learning from negative evidence to generate knowledge that survives independent validation.

To illustrate this gap, researchers recently asked two teams of AI agents to revisit 20 high-stakes biopharma decisions using only the evidence available at each historical decision point. In some cases, the agents arrived at better decisions than those made at the time. In others, they reproduced the same failures because the evidence required to make the correct choice did not yet exist. This exercise revealed both the remarkable power and the present ceiling of agents: they can synthesize complex evidence and improve decisions, yet analysis alone cannot generate novel experimental results.

Why Did AI Breakthroughs Happen First in Math and Software?

To understand where biomedical AI innovation might go, it helps to examine where AI has already achieved superhuman capability. Advanced versions of AI models have achieved gold-medal performance at the International Mathematical Olympiad, while other systems have improved solutions to open mathematical problems and optimized algorithms deployed in major computing infrastructure.

These domains moved first for structural reasons. Their objects can be represented in machine-readable form, candidate solutions can be evaluated rapidly, and the evaluators themselves are often objective. Code either compiles or fails, tests pass or fail, proofs can be checked, and algorithms can be scored millions of times inside closed computational loops. The key insight is that when a domain provides adequate representations, rapid feedback, and objective evaluators, AI can progress from producing plausible answers to discovering better solutions.

Biology presents almost the opposite structure. Its state is only partially observed and often imprecisely measured in terms of molecules, cells, tissues, organs, and complete patient histories. Experiments may take weeks or years, cost millions of dollars, and yield noisy or ambiguous results. An intervention changes the very system a model is attempting to predict, and at the clinical scale, the most meaningful evaluator may be a patient outcome observed years later.

What Five Requirements Must Be Met for Biomedical AI Innovation?

Biomedical superintelligence, defined as domain-specific AI systems that generate novel biomedical knowledge withstanding empirical validation and translate it into better patient decisions than world-leading human experts, will require more than increasingly capable AI models. Instead, five interlocking requirements must be satisfied:

• Biological Representation: Before a model can reason about biology, it must represent biological state at the resolution at which mechanism operates. Feature extraction must convert raw measurements into representations that preserve information relevant to biological tasks, a process far more difficult for specialized biomedical data types than for language or code.
• Multimodal Data Integration: Gigapixel pathology slides, spatial transcriptomic maps, multiplex imaging, single-cell perturbation data, longitudinal circulating tumor DNA, and continuous physiological signals each provide different views of the same biological system. None in isolation is sufficient to represent complete biological state.
• Rapid Feedback Loops: Unlike software testing, biological experiments may take weeks or years and cost millions of dollars, making rapid iteration and evaluation far more challenging than in domains where AI has already succeeded.
• Objective Evaluators: Biology lacks the clear pass-fail criteria of software or mathematics. Patient outcomes, the most meaningful evaluator at clinical scale, may not be observed for years after an intervention.
• Adequate Experimental Infrastructure: Just as protein design succeeded where other biological domains have lagged, progress requires the foundational infrastructure that mathematics and software possess by default.

Foundation models are beginning to address these challenges. Systems like CONCH have learned visual-language representations from pathology images and text, while TITAN extended representation learning to whole-slide images aligned with pathology reports. Visual-omics models have begun connecting histologic morphology with spatial gene expression, creating the kind of multimodal understanding necessary for genuine biological discovery.

How to Prepare for the AI Innovation Era in Biomedical Research

• Invest in Data Infrastructure: Organizations should prioritize building the representation systems, feedback loops, and evaluators that mathematics and software possess by default. This means creating standardized formats for specialized biomedical data types and establishing rapid assay infrastructure.
• Develop Multimodal Integration Capabilities: Teams should work on connecting disparate data sources, from pathology images to genomic data to clinical records, into unified representations that AI systems can reason about holistically.
• Establish Clear Validation Standards: Before AI systems can function as true innovators, the field must define what constitutes valid biomedical discovery and establish processes for independent validation of AI-generated hypotheses and experimental results.

"Artificial intelligence is approaching its most important transition: from systems that retrieve, reason and execute within an existing evidence base to systems capable of expanding that evidence base through discovery," stated Jorge Reis-Filho, Chief Strategy Officer and Operating Partner at Foresite Labs and Venture Partner at Foresite Capital.

Jorge Reis-Filho, Chief Strategy Officer and Operating Partner at Foresite Labs

The timeline for this transition remains uncertain. Protein design, which sits near the software end of the verification spectrum with fast assays and decades of curated crystallographic data, has already seen scale-first approaches succeed. Whole-organism and clinical biology, where the oracle is priced in years, still await their equivalent breakthrough. But the direction is clear: the next frontier for AI is not better analysis of existing knowledge, but the generation of knowledge that nature has yet to reveal.