Well I think the issue is more of if you’re Genentech and you need ML people and can’t afford to pay them you’re probably better off retraining chemistry PhDs.
I think you missed my point. Genentech, AFAIK, was not doing research on machine learning as in the principles of how machine learning works and how to make it better. They do biotech research which uses applied machine learning. You don't need a PhD in ML to apply things that are already known
As a PhD student working on core ML methods with applications in chemistry, I second this. During my PhD, I read very few papers by chemists that were exciting from a ML perspective. Some work very well, but the chemists don't even seem to always understand why they made the right choice for a specific problem.
I don't claim that the opposite is easy either. Chemistry is really difficult, and I understand very little.
Genentech has several ML groups that do mostly applied work, but some do fairly deep research into the model design itself, rather than just applying off-the-shelf systems. For example, they acquired Prescient Design which builds fairly sophisticated protein models (https://nips.cc/Conferences/2022/ScheduleMultitrack?event=59...) and one of the coauthors is the head of Genentech Research (which itself is very similar to Google Research/Brain/DeepMind), and came from the Broad Institute having done ML for decades ('before it was cool').
I can't say I know anybody there who is doing what I would describe as truly pure research into ML; it's not in the DNA of the company (so to speak) to do that.
Sadly a lot of foundational ML research works for single-label image classification and not much else. ImageNet is a niche problem and way too much ML research is over-indexed on it. If you can make your problem look like ImageNet, you're going to do OK, but if not you effectively need to re-invent the wheel...
The HEAR benchmark is a great eye-opener. They have basically three classes of audio tasks, and find very different models excel in each, with the best overall models being kinda-mindless ensembles of the ones that do well on particular problems.
So if you've got something that works well for text... it'll take a couple years and maybe an entire new branch of research (diffusion!) to work well for image generation. I have no idea what generative models for chemistry will look like, but will happily bet that it takes some significant specialized effort.
1. The era of ML benchmarks is ending. New models have to be and will be evaluated the same way human experts are evaluated.
2. Foundational models are becoming multi modal. There will be no separation of text and image generation. Sure, different methods will be used for each, but the learned representations of visual and textual objects in models like Stable Diffusion already live in the same conceptual space.
I don’t think there will be specialized generative models for chemistry two years from now. There will be GPT-5 (and similar competitors) which will be used to perform all kinds of research, including chemistry.
For example, AlphaFold just fundamentally isn't a language model, but is fundamentally useful. We'll still need these models that Do Stuff in many areas, and that will still involve benchmarks... Even if we're able to ask GPT-N+1 to design the next version of the model for us.
