DeepMind, recently rebranded as Google DeepMind after some reshuffling with Google Labs, has had some stunning breakthroughs with their AI AlphaDev (built atop of their prior game-changing and game-playing AlphaZero).
Among the record-breaking, the AI has found faster ways to perform complex calculations, to sort lists, and to do some of the critical steps of cryptography (used for security encoding and for code-breaking). The list-sorting one is quite incredible as its new method is up to 70% faster than the best previous method.
The AlphaZero and AlphaDev AIs are not LLMs of course, but rather a very different kind of AI that allows the AI to continuously learn. It is an AI built for reasoning and problem-solving, with the capability to both experiment and learn, all four of which things are NOT what LLMs do.
The thing is that AIs of this kind, the non-LLM ones, are great at building the algorithms that logic-based AIs use, and can thus also be used to build future LLM training systems that are more efficient.