Y Combinator1h 22m
Is AI Now Good Enough To Cure Cancer? | YC Paper Club
Alfredo Gonzalez, Yusuf Roohani, Garyk Brixi, Divya Nori
Conference Presentation, Keynote
- Predicts a greater than 70% chance of AI interactions being identified by humans, failing to pass the Turing test as defined by the year 2000 projection.
- States there is a 0% chance that a code project vibe-coded within the last 12 hours will function correctly, though an attempt is planned.
- Anticipates a shift from data analysis to real-time biological design by 2030, compressing traditional 10 to 20-year grant and recruitment timelines into immediate patient outcomes.
- Expects the ability to sign proteins in real time to become standard practice within 50 to 80 years, enabling the design of delivery mechanisms and peptides for individual patients.
- Forecasts self-improving biology models emerging in the near future to save thousands of hours of research time through autonomous system enhancement.
- Plans to begin a dose escalation for ground-up biological designs to ensure safety and mitigate unintended consequences, mirroring current clinical trial methodologies.
- Predicts new preprints regarding prompting and steering EVO will be released next week to address mechanistic interpretability limitations and demonstrate post-training pipelines.
- Expects next week's preprints to cover in-context learning and post-training mechanisms for enhancer design, including wet lab validation in mouse brain models.
- Anticipates future models will integrate multi-omic and population data to fill gaps left by foundation models trained on a single genome per organism.
- Believes few-shot prompting in biological sequence design will enable lab-in-the-loop styles where initial prompts are continuously updated based on lab tests.
- Predicts that inference time techniques and harness engineering from the LLM field will be applied to biological design settings to improve efficiency.
- Identifies the burstiness of genomic data distribution as a primary driver for the emergence of in-context learning capabilities in next token prediction models.
- Expects reinforcement learning loops using sequence-to-function models as reward functions to post-train models for leveraging in-context examples in specific tasks like enhancer design.
- Plans to combine a generative model of DNA with classifiers and function prediction to design new phix-like bacteriophage.
- Expects the bidirectional model Minerva to be used for more accurate coevolution prediction, enabling rapid scanning of genomes for non-coding RNA and regulatory DNA structures.
- Anticipates future work will involve patching issues like distal enhancer elements by adjusting data composition to boost model performance.
- Predicts frontier labs are testing text-DNA combined approaches versus separate embeddings to achieve human-interpretable insights in models trained on text.
- Expects next token prediction on nucleotide-level tokenization to remain the preferred method for mutation prediction benchmarks despite potential shifts to byte-level tokenization.
- Plans to use agents to mine data in real time for personalized research labs, aiming to shorten patient outcome timelines from years to real time.
- Expects clinical trials and drug repurposing projects to increasingly rely on AI to identify targets modulating specific human phenotypes rather than designing molecules for known targets.
- Anticipates scaling AI-guided experimentation rather than just data generation to build high-utility virtual cell models.
- Expects autonomous systems optimized for ML post-training to show significant improvement over coding agents as AI R&D compute scales.
- Expects virtual cell research to map cell states to identify interventions transitioning diseased cells to healthier states, addressing failures in current clinical efficacy predictions.
- Plans for physical labs to connect with AI loops to optimally generate data for specific questions, representing an accelerated approach to virtual cells.
- Expects post-training to be necessary to bake fast retrieval capabilities when harness engineering alone is insufficient for specific retrieval tasks.
- Anticipates human-specific data will be integrated into future models to complement fixed token budget approaches focused on diverse conserved signals.