The items below are statements from research papers Sergey Edunov co-authored and things Edunov said in an interview about machine learning for drug discovery. — Ching, Edunov, Kabiljo, Logothetis and Muthukrishnan, “One Trillion Edges: Graph Processing at Facebook-Scale” (VLDB 2015).

Visual summary of operating lessons from Sergey Edunov.

Part 1: Large-Scale Graph Processing and Infrastructure

  1. On Industry-Scale Graphs: In the paper on graph processing, Edunov and co-authors wrote: "Industry graphs can be two orders of magnitude larger - hundreds of billions or up to one trillion edges". — Ching, Edunov, Kabiljo, Logothetis and Muthukrishnan, “One Trillion Edges: Graph Processing at Facebook-Scale” (VLDB 2015).

Part 2: Scaling Neural Machine Translation

  1. On Reduced Precision: In the paper on scaling translation, Edunov and co-authors wrote: "reduced precision and large batch training can speedup training by nearly 5x on a single 8- GPU machine with careful tuning and implementation". — Ott, Edunov, Grangier and Auli, “Scaling Neural Machine Translation” (2018).

Part 3: Data Augmentation and Model Training Techniques

  1. On Sampling: In the paper on back-translation, Edunov and co-authors wrote: "sampling or noisy synthetic data gives a much stronger training signal than data generated by beam or greedy search". — Edunov, Ott, Auli and Grangier, “Understanding Back-Translation at Scale” (2018).
  2. On Effective Back-Translations: In the paper on back-translation, Edunov and co-authors wrote: "back-translations obtained via sampling or noised beam outputs are most effective". — Edunov, Ott, Auli and Grangier, “Understanding Back-Translation at Scale” (2018).

Part 4: Open Source Foundation Models

  1. On Architectures: In a podcast interview about AI for drug discovery, Edunov said: "I wouldn't be that surprised by the model architecture". — Sergey Edunov interview on The Cognitive Revolution, auto-generated captions.
  2. On No Consensus: In a podcast interview about AI for drug discovery, Edunov said: "there is no consensus whatsoever". — Sergey Edunov on The Cognitive Revolution podcast about AI for drug discovery, auto-generated captions
  3. On Opportunity for Innovation: In a podcast interview about AI for drug discovery, Edunov said: "the opportunity for innovation here is much much bigger". — Sergey Edunov on The Cognitive Revolution podcast about AI for drug discovery, auto-generated captions

Part 5: Physics, Machine Learning, and Molecular Biology

  1. On Data: In a podcast interview about AI for drug discovery, Edunov said: "AI for drug discovery is not fundamentally different from LMS". — Sergey Edunov interview on The Cognitive Revolution, auto-generated captions.
  2. On Selectivity: In a podcast interview about AI for drug discovery, Edunov said: "we don't only need to find the molecule that binds". — Sergey Edunov interview on The Cognitive Revolution, auto-generated captions.
  3. On Useful Models: In a podcast interview about AI for drug discovery, Edunov said: "it makes those models actually useful in real drug programs". — Sergey Edunov interview on The Cognitive Revolution, auto-generated captions.
  4. On Binders: In a podcast interview about AI for drug discovery, Edunov said: "protein binders themselves is not a therapeutic modality". — Sergey Edunov interview on The Cognitive Revolution, auto-generated captions.
  5. On a Long Pipeline: In a podcast interview about AI for drug discovery, Edunov said: "the whole drug discovery pipeline is very long and very complex". — Sergey Edunov interview on The Cognitive Revolution, auto-generated captions.
  6. On Agents: In a podcast interview about AI for drug discovery, Edunov said: "having all of this abstracted away and driven by agent is potentially a super powerful innovation". — Sergey Edunov interview on The Cognitive Revolution, auto-generated captions.
  7. On Not Just Binding: In a podcast interview about AI for drug discovery, Edunov said: "just finding something that binds to a target is clearly not enough". — Sergey Edunov on The Cognitive Revolution podcast about AI for drug discovery, auto-generated captions
  8. On Qualitative Shifts: In a podcast interview about AI for drug discovery, Edunov said: "more of a qualitative shift than quantitative shift". — Sergey Edunov on The Cognitive Revolution podcast about AI for drug discovery, auto-generated captions
  9. On Scaling Differently: In a podcast interview about AI for drug discovery, Edunov said: "we can't just scale the way LLMs were able to scale". — Sergey Edunov on The Cognitive Revolution podcast about AI for drug discovery, auto-generated captions
  10. On Inductive Biases: In a podcast interview about AI for drug discovery, Edunov said: "we need to build inductive biases into the models themselves". — Sergey Edunov on The Cognitive Revolution podcast about AI for drug discovery, auto-generated captions
  11. On Learning Every Tool: In a podcast interview about AI for drug discovery, Edunov said: "learning every single model and tool that is available on the market is very hard". — Sergey Edunov on The Cognitive Revolution podcast about AI for drug discovery, auto-generated captions
  12. On Changing Tools: In a podcast interview about AI for drug discovery, Edunov said: "those tools change all the time". — Sergey Edunov on The Cognitive Revolution podcast about AI for drug discovery, auto-generated captions
  13. On Starting From the Ground Up: In a podcast interview about AI for drug discovery, Edunov said: "rather than trying to rethink things from the ground up". — Sergey Edunov on The Cognitive Revolution podcast about AI for drug discovery, auto-generated captions
  14. On Human Taste: In a podcast interview about AI for drug discovery, Edunov said: "human taste is still very very important in this field". — Sergey Edunov on The Cognitive Revolution podcast about AI for drug discovery, auto-generated captions