Many years ago (13?), I was around when Amazon moved SABLE from RAM to SSDs. A whole rack came from a single batch, and something like 128 disks went out at once.
I was an intern but everyone seemed very stressed.
Shameless self promotion: I wrote one of the more cited papers in the field [0], back in 2016.
A key challenge: very few labs have enough data.
Something I view as a key insight: a lot of labs are doing absurdly labor intensive exploratory synthesis without clear hypotheses guiding their work. One of our more useful tasks turned out to be interactively helping scientists refine their experiments before running them.
Another was helping scientists develop hypotheses for _why_ reactions were occuring, because they hadn't been able to build principled models that predicted which properties were predictive of reaction formation.
Going all the way to synthesis is nice, but there's a lot of lower hanging fruit involved in making scientists more effective.
At Amazon I set up an evaluation approach based on whether the system completed the desired task (in that context it was "did the search result using the speech recognition return the same set of items to buy as the transcript.)
I'm not with the lab any more, but last I heard they were soliciting data from nearby but related subfields (our lab was doing mostly vanadium templated tellurides, but there are a lot of other things to explore in the space of inorganic–organic hybrids, and there are labs working on them). The problem really is data sharing; the most active collaborators were willing to give us lab note-books as long as we paid to digitize them; it was very hard to convince other labs to digitize their own notebooks. TBH I think that's the hard part of a lot of this (collecting the data).
We did something pretty similar in a nearby field of materials science, using support vector machines rather than linear regression and focusing on a less specific target:
LSTMs are very different. Think of it this way. LSTMs store information about the current problem you're solving. DNCs store information about the world.
LSTMs are designed to capture long range dependencies, e.g., "this word at the start of the sentence interacts with this word at the end of the sentence."
DNCs are designed to incorporate outside information, e.g., "i happen to know (from background knowledge) that these two people in this sentence are married"
As an alternative perspective: I worked in a lab that was relying heavily on intuition and heuristics for exploratory synthesis. By digitizing all of the lab's data and doing some straightforward machine learning, we were able to sufficiently improve the efficiency of the lab, compared to using intuition, and help the chemists develop a deeper understanding of the system they were studying:
Agreed. I published a paper on exactly that ( augmenting scientists searching for novel materials, rather than replacing scientists entirely). One of the benefits is that you can often model simpler problems which are more tractable.
Walmart could fill a hole in the ground with $3 billion dollars and burn it, and not care. [A bit of an exaggeration, as that's half their annual profit]
Here's one from Deepmind:
https://arxiv.org/abs/2509.10147