Why does biology remain exceptionally difficult to master? This compelling question cuts to the core of the strange dynamic between large language models—often viewed as a form of digital consciousness—and the unpredictable, tissue-based architecture of the human brain. While early machine learning treated digital networks as entirely distinct from biological systems, today's landscape reveals a blurring of those lines.

During a panel discussion at the Next Endeavor event held September 14-15 in Mountain View, California, industry experts explored the friction of applying artificial intelligence to biological sciences. Lisa Dolan of Link Ventures moderated the conversation, speaking with Surya Ganguli, Krithik Ramesh, and Ali Khalatpour about the immense hurdles in bridging tech and biology.

Ramesh noted that he entered the field rather than pursuing finance, emphasizing that biological systems operate under fundamentally different rules than human language. Despite massive advancements, pre-training models on protein sequences does not guarantee accurate characterization. Instead, researchers must focus on scaling paradigms that remain biologically faithful through rigorous, multi-stage therapeutic development.

The Deep Complexities of Biological Data

Addressing the audience, Ganguli upended conventional wisdom regarding academic difficulty. He pointed out that while mathematics was once considered the ultimate hurdle for students, AI has largely mastered math. True complexity now lies in biology, hampered by scarce and noisy data, partial observation, and a distinct lack of foundational theoretical frameworks.

Successful AI-driven life science companies must forge unique moats around data acquisition, fusing advanced algorithms with proprietary collection methods. For instance, combining novel magnetic field sensors with signal processing and artificial intelligence allows researchers to dramatically improve the signal-to-noise ratio in non-invasive brain-computer interfaces.

Unlocking Diseases of Consciousness

Delving further into neurology, Ganguli highlighted that the human brain represents the most intricate biological system known, yet remains the least accessible. Researchers are now tackling complex disorders of consciousness—ranging from schizophrenia-induced hallucinations to ketamine-triggered out-of-body phenomena—by building predictive AI models of brain activity. By perturbing these artificial systems, scientists can simulate and eventually target neurological disorders that alter our fundamental sense of self.

The conversation then shifted toward hardware innovation as Khalatpour detailed the development of advanced space-rated lasers funded by NASA. Overcoming immense thermal, power, and compute constraints requires translating scientific breakthroughs into functional systems. Furthermore, Khalatpour drew a World Cup analogy to illustrate the critical bottleneck of data movement across GPUs, highlighting systemic inefficiencies in traditional data center hierarchies.

From Electrons to Algorithms

The panel underscored the stark efficiency gap between biological and artificial intelligence. While the human brain operates on a mere 20 watts, an incandescent bulb consumes 100 watts, and a single H100 GPU demands roughly 700 watts. Scaling systems down from millions of watts to biological levels necessitates a complete redesign of the technology stack, stretching from raw electrons up to high-level algorithms.

Concluding the session, Dolan asked the panelists what bold bets they wished founders would pursue. Ganguli delivered a striking perspective, arguing that foundational approaches to AI instruction remain fundamentally flawed due to a failure to match the sample efficiency seen in natural brains. Ramesh advocated for greater humility and a focus on practical, underdeveloped problems within biology, while Khalatpour emphasized the urgent need for hardware and software architectures that enable true continuous learning.