What are the key developments in the field of artificial intelligence? On the other side, taking into account the ethics of AI becomes crucial.
In our opening program for the 2020s, we have laid the foundations for evaluating all the technologies under the generic term of “artificial intelligence”. We will now use it to refer to some key developments in this area, starting with hardware.
The key takeaway here is that the proliferation of machine learning workloads has driven the use of GPUs, previously used primarily for gaming while spawning a whole new lineup of manufacturers. Nvidia, which has come to dominate the AI chip market, has had a very productive year.
First of all, in unveiling its new Ampere architecture, Nvidia claims that it has made about a 20-fold improvement over Volt, its previous architecture. Then, in September, Nvidia announced the acquisition of Arm, another chipmaker. As we noted then, the acquisition of Arm by Nvidia strengthens its ecosystem and brings economies of scale in the cloud and expansion to the edge.
As others have noted, however, the acquisition is subject to regulatory review. The field of AI chips deserves further analysis, which we will embark on soon. However, honorable mentions can already be given to Graphcore, for raising more capital and seeing the chips deployed in the cloud and on-premises, to Cerebra’s, for unveiling its second generation of AI chips, and to Blaize, for having released new hardware and software products.
MLOps, a major themeThe software side has been just as hectic, if not more. As the State of AI 2020 report indicates, MLops has been a major theme. MLOps, short for Machine Learning Operations, is the equivalent of DevOps for ML models: moving them from development to production and managing their lifecycle in terms of improvements, fixes, redeployments, etc.
Some of the most popular and fastest-growing Github projects in 2020 are related to MLOps. Streamlight, which helps deploy applications based on machine learning models, and Dask, which improves Python performance and is made operational by Saturn cloud, are just two of many examples. Explainable AI, the ability to inform decisions made by ML models, may not be as operational, but it is also gaining ground.
Another key theme was the use of machine learning in biology and healthcare. AlphaFold, the DeepMind system that has successfully solved one of the world's toughest computational challenges of predicting how protein molecules will fold is a prime example. Other examples of AI impacting biology and healthcare are already available in the works.
The side effects of AI: debates on ethicsBut what we think should be at the top of the list is not a technical achievement. This is what we call the ethics of AI, that is, the side effects of its use.
Google recently "thanked" TimnitGebru, a prominent and highly respected researcher in AI ethics research and formerly in charge of the team working on ethics issues on Google artificial intelligence.
TimnitGebru was dismissed for uncovering embarrassing truths. In addition to the prejudice and discrimination which TimnitGebru says are not just side effects of datasets reflecting prejudice in the real world, there is another aspect of his work that deserves to be highlighted the disastrous environmental consequence of focusing on ever larger and more resource-intensive AI models. The rejection of the issue by DeepMind speaks volumes about the priorities of the industry.