Is AI Now Good Enough To Cure Cancer? | YC Paper Club
Y Combinator
Turing test chat 0:05
The host opens with a live test of the Turing test. People in the room scan a link on their phones and are paired either with another person or with an AI agent for a two-minute chat. The result is messy. Some humans are mistaken for AI, and the host says the score is only about 30%, even after trying to give the AI a convincing persona and context.
Why bio needs AI 3:30
The talk then turns to AI for biology. The host says more data than ever is now available, from huge protein and genome models to drug work that is already moving into patients. He argues that biology can be treated like any other task that can be checked, sampled, tested, and retrained on, even if the real-world loop costs time and money. He also introduces the speakers who will talk about cancer, cell states, foundation models, and bio design.
Rare tumors and real time 16:30
Rare tumors often leave clinicians with too little data, so the team clusters a patient’s samples across many datasets and looks at the active pathways instead of the label alone. In one brain tumor case, that showed a closer match to mesenchymal tumors, which opened therapy options that the literature had not made clear. The same approach also helped after a patient had already failed several lines of treatment.
Faster models and trials 20:31
The speaker says future work will build proteins and delivery tools in real time, cutting design timelines by years. He points to Google DeepMind’s co-scientist as an early sign that self-improving biology models can save thousands of research hours. During questions, he says organoids are scientifically feasible but hospital tissue handling often blocks them, and that clinicians need clear reports because they still make the final treatment call.
Virtual cell models 32:00
Virtual cell research aims to model biology at the cell, tissue, and organism level. The speaker says sequence models for protein and DNA have advanced fast, but the missing step is cell state. A useful model would map healthy and diseased states, then point to interventions. He links this to poor clinical efficacy and the weak ability to predict what will work in people.
Looped discovery systems 36:01
He describes a universal cell embedding trained on 35 million cells from gene expression data, then says the field found the public data too narrow, dominated by immune and T cells. His answer is to shift from scale-first data gathering to AI-guided experimentation. He argues that discovery loops, like those in automated ML research, can improve data generation, search, and noisy evaluation.
Compute and harness 46:31
The work on long-running agents centers on keeping track of progress over long horizons. It uses judge agents and other agents to rank ideas, account for the cost of experiments, and spend GPU compute well. When asked about the chart, the answer is that the comparison is about GPU hours, not token cost, and every system was run in the same sandbox with the same hardware.
DNA as language 52:01
The speaker then turns to DNA as a language model problem. The model is trained on 9 trillion base pairs, with data filtered to one species sample each and enriched around genes, RNAs, and conserved regions. It predicts mutations, matches disease-linked loss of function, and exposes learned features like phage signals, which can be used to scan genomes and design new phage-like systems.
Prompting genome models 1:03:00
The talk shifts from pretraining to what comes next. The speaker says these genome models can be used as tools for scientists, or joined with other data types in new training setups. Questions from the audience focus on tokenization, mixing DNA with text, and whether the pretraining mix can be tuned more carefully as evaluations improve. The answers stay practical. Nucleotide-level tokenization seems to help mutation prediction, but other levels may work too.
Design through examples 1:10:01
The next speaker lays out a new design method based on prompting a genome language model. Instead of fine-tuning on large labeled sets, the model is given a few examples in context and asked to generate new sequences with similar properties. The same pattern is shown for proteins, antibodies, CRISPR-Cas systems, phage genomes, and cell-type enhancers. The enhancer work is then improved with GRPO reinforcement learning and tested in mouse brain, where the designed sequence shows the expected activity.
Enhancer specificity sources 1:19:30
The discussion turns to enhancer design. The speaker says the model’s success comes from both what it learned in its weights and what it picks up in context. Prompt examples may show motifs that signal cell-type specificity, and the model preserves and recombines those motifs.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.

