The prestigious ‘MIT Technology Review’, associated with the famous tech college of the same name, has published a list of groundbreaking technologies for 2021. Last year the tech magazine stated that it did not want to go for hypes, but actually wanted to recognize revolutionary technologies. The prognoses for 2020 were: a non-hackable internet, digital money wars, new drugs thanks to AI, the struggle for quantum domination. These technologies still need some time, in some cases the pandemic is likely to have set new priorities in their places. Nevertheless, we have again picked three out of the new tech trends.
GPT-3: King of the natural language models
Large natural language computer models that learn to write and speak are a big step in the direction of artificial intelligence that can better understand the world and interact with it, writes the Technology Review‘. GPT-3 is by far the largest and most linguistically gifted model to date. It has been trained on the text from thousands of books and most of the (filtered) internet. GPT-3 can mimic human-written text with eerie – and sometimes bizarre – realism, making it the most impressive language model machine learning has ever created, tech magazine praises.
But bizarre also means that some of the results seem nonsensical. After all, GPT-3 does not understand what it is writing . In addition, the Internet, with which the model was trained, is full of fakes and prejudices, some of which flow into the training data: The result is distortion. At the same time, the model consumes enormous computing power, data and money to train. That leaves a huge CO2 footprint and limits the model to use in research institutions, the ‘MIT Technology Review’ puts its praise in perspective.
The AI project comes from the legendary non-profit organization OpenAI. ETH professor Benjamin Grewe stated about the model in the summer: “According to the developers, GPT3 is already reaching the limits of what is possible with the amount of data. This also shows that highly specialized learning algorithms with even more data will no longer significantly improve machine learning.”
Multi-talented AI: see, feel, hear
The road to general Artificial Intelligence is long and the goal is not really known. According to the MIT Technology Review, AI is still stupid when it comes to identifying new problems or finding your way in unfamiliar environments. Artificial intelligence ultimately lacks the ability that small children already have: to recognize and learn how the world works and to apply this general knowledge to new situations.
A promising approach to improving AI capabilities is to expand its “senses”. At present, AI can perceive things with computer vision or audio recognition, but not “speak” about what it sees and hears with natural language algorithms. But what if you combined these skills into a single AI system? Could these systems begin to gain human-like intelligence? Could a robot that can see, feel, hear, and communicate be a more productive human assistant? Questions that drive the ‘MIT Technology Review’.
OpenAI also offers an approach. In January, the organization presented a model that can convert language into images : Dall · E could put together different ideas through input with natural language and generate free images. However, it was still a long way from replacing photographers and artists with their own style and creative vision, it was said at the time. This is because machines at this stage have difficulty developing and implementing really new ideas.
Data Trusts: The data trustee model
Tech companies have proven to be poor stewards of personal information. Our data has been leaked, hacked, sold and resold more often than most of us can count, writes the ‘MIT Technology Review’ and asks: “Maybe the problem is not with us, but with the model of privacy, which we have long worked on have held? ” A model in which we ourselves are responsible for protecting our privacy.
Data trusts offer an alternative approach here that some governments are beginning to explore. In the data trustee model, a legal entity is responsible for collecting and managing people’s personal information on their behalf. Although the organization and function of these trusts have yet to be defined and many questions remain unanswered, data trusts would offer a potential solution to long-standing problems in the area of data protection and security, the tech magazine hopes.
Whether this will also benefit users is an open question: In Switzerland, anonymization and storage in data centers of the state administration were promised for the corona tracing app. Nevertheless, there were great fears among the population that their data would not be in safe hands . However, the approach is not very far advanced either.