Google’s AI research lab, DeepMind, continues to push new progress in the realm of AI. The latest step comes in creating a Google-powered medical AI that is actually somehow better than actual doctors. The goal was to create an AI that could help shoulder some of the load that doctors experience each day, and based on new reports, Google’s medical AI might be able to do just that, new studies suggest.
The model appears to be built off of what Google is already doing with its Gemini models. Med-Gemini is built differently than other medical AIs, though, because Google created it as self-training. It does this by relying on web search capabilities (which is slightly worrying because of how much nonsense is out there).
Whether or not that will actually make Google’s medical AI effective at training itself is a bit iffy. But, the company says it should at least allow for a fine-tuning loop. So far, the Med-Gemini model has been tested on 14 different medical benchmarks, and it has established a new state-of-the-art (SoTA) performance on 10, surpassing the scores seen by GPT-4 models on every benchmark where a comparison existed.
That, of course, isn’t as big of news as the other news that Google revealed, which claims that Med-Gemini actually outperformed actual doctors and its previous medical LLMs, like Med-PaLM 2. Google says Med-Gemini scored a 91.1% accuracy score using its uncertainty-guided search functionality.
This is, of course, very exciting news for Google’s medical AI, though it certainly raises some questions about how soon we’ll see AI like this actually working in doctor’s offices. It also raises questions about how people will react to such a thing. But you can at least rest easily knowing that AI like this shouldn’t replace doctors. Instead, it’s being designed to assist them and make their job easier so they can focus on the patients that need them most.