I started doing this too after reading The Artists Way by Julia Cameron. She prescribes these “morning pages” as one of principal tools for overcoming internal creativity blocks. I also really enjoy drinking my coffee and writing these pages with my pilot fountain pens.
I like this paper and it appears to be one of the best in the literature so far for AI for materials. Even DFT is not really scalable for this, computing the ground state of even a dozen unit cells requires many many CPU-hours. They themselves in fact relax the proposed structures by minimizing the energy of psuedopotentials, for even DFT is too expensive for that step. I said already I think improving DFT itself is the most potentially impactful application of AI in this space, in my opinion. Of course approximations are always necessary, I’m not at all against that, but DFT ignores or approximates correlations by design so there is an inherent limitation there, which means, if you train your models to predict that, it will have the same limitation. It’s just like with LLMs, only imagine training principally on synthetic data. Obviously LLMs have succeeded with limitation sources of synthetic data, but they are principally trained with “real” data.
I’m glad they actually tried synthesizing one of the materials their model predicted. Looks like they succeeded in synthesizing only 1 out of 4 of the materials for which they tried. The 20% accurate property claim appears to be for bulk modulus. I’m still seeing little value for this technology for designing electronic properties, mainly because density functional theory which provides the training data is not reliable. Their code looks nice and clean and well organized, perhaps I’ll give it a try.
My biggest problem with this application of AI is trying to approximate DFT, which itself is an unreliable approximation. The claim is it lets you amortize the expensive DFT to search the space, but it’s also true that especially for inorganic materials, training sets do not appear to promote strong generalization. So you embark on an expensive task to wind up back with unreliable DFT. I think perhaps the best goal would be to try to make DFT itself better, and I have seen impressive albeit computationally expensive approaches, e.g. FermiNet by DeepMind.
These cyclists unfortunately need to just point their headlight downward--included in the installation instructions for almost all bike headlights is the direction to point the light below horizontal. That is also the only difference between automotive high and low-beams: the angle at which they are directed.
This is great work, but HumanEval is an extremely limited benchmark and I don’t think you can seriously claim to beat GPT-4 at coding based only on that metric.
I work on Copilot now, and I can say this guy is totally full of himself. Just because he contributed to the first prototype does not entitle him to claim creation. He wasn’t even the first person inside Microsoft to use LLMs in an IDE completion experience. The work from prototype to product is an order of magnitude more than prototype alone.
Very nice. For those interested, you can get similar information using SourceGraph and LSIF in a standardized, language agnostic form: https://lsif.dev/. It still generally requires build information for each project/language, unfortunately.
You may be interested in this paper: https://arxiv.org/abs/2105.09352
As far as I know it was the first to train a model on commit diffs to generate code mutations, in this case for bug fixing.