Why AI 2.0 is the next evolutionary step and what it means
Many companies have recognized the potential of artificial intelligence, yet only few fully make use of it. At the same time, new technologies that will take AI to an even higher level are surfacing rapidly.
Kai-Fu Lee, author of "AI Superpowers", recognized four waves of the development of AI technologies. While the first phase, “Internet AI”, refers to applications being based on big data from the Internet and the improvement of the user experience, mainly at large Internet and e-commerce companies, “AI 2.0” is all about “business AI”.
AI 2.0 in business
In the AI 2.0 phase, companies of all industries and sizes use AI technologies to increase efficiency and discover and create new business models. This is achieved by analyses and predictions of existing company data and other sources.
PWC estimates the gigantic economic growth potential of AI 2.0 at 15.7 trillion dollars. I further discuss this topic in my dataiku egg-on-air video.
AI 2.0 infrastructures
In terms of concepts and infrastructures, the focus is on the ability to develop scalable and production-ready AI applications.
Platforms and templates are used to implement data and AI products. To transfer them into production in an agile way, MLOps processes come into play. But alongside, the implementation of the necessary organizational structures and processes in the company is crucial. Therefore, data strategies must be developed, data governance concepts implemented and roles and responsibilities defined.
AI 2.0 technologies
A recent Forrester report lists the elements AI 2.0 technologies are based on:
- Transformer networks
- Synthetic data
- Reinforcement Learning
- Federated Learning
- Causal Inference
AI 1.0 focused on pattern recognition, task-specific models, and centralized training of models and their execution. AI 2.0 on the other hand is defined by the establishment of models to generate language, images and other data, as well as the universal applicability of AI, centrally or locally - at the Edge.
Let's further examine these 5 core technologies:
Transformer
Transformer networks can handle tasks with a time or context element, such as natural language processing and generation. This allows the training of large models that can execute multiple tasks at once with higher accuracy and less data – compared to individual applications working indepently from each other.
The widest know example of this is GPT-3 by OpenAI, a really powerful model.
Synthetic data
One of the biggest challenges in building AI models is the availability of a sufficiently large, usable training data set. Synthetic data can solve this problem and improve the accuracy, robustness and generalizability of models. This can be used in technologies for object recognition, in the fields of autonomous driving, healthcare, and many others.
Reinforcement Learning
Reinforcement learning isn’t new as a concept, but it will be used more widely. AI applications can use reinforcement learning to quickly respond to changes in data by learning from interaction with a real or simulated environment through trial and error.
Federated Learning
One obstacle to training AI models is the need to transfer data from multiple sources to a central data store. This is quite costly, difficult, and often risky in terms of security, privacy or competitiveness. With Federated learning, AI models can be trained in a distributed manner directly on IOT endpoints, for example, and data can be leveraged in different locations.
Causal Inference
Causal inference identifies cause-and-effect relationships between attributes of a data set and analyzes correlations. Therefore, non-optimal business decisions based of spurious correlations can be avoided.
To summarize, AI 2.0 can be understood as an attempt to get closer to natural intelligence by means of imagination, trial and error, exchange of experience and understanding of modes of action. Just like in nature, companies leveraging the resulting advantages are fitter for survival.
The analyses of the possibilities and potential AI 2.0 can bring to a company should be performed right at the start of any AI transformation. This means that potential ?killer applications“ for the company's own business model can be implemented at an early stage.
AI 2.0 in Europe
An important factor in Europe in the development and application of technologies during the AI 2.0 era are values and quality standards. Ethical issues must be discussed, clarified and transfered into regulations – while at the same time not restricting the innovation power and the economical potential of AI.
Regulations must be defined with a sense of proportion, focusing on specific application scenarios. Already existing measures need to be taken into account and new regulations must be based on a transparent and precise risk assessment.
This is the key for Europe to level up to AI nations like China and the US and to build our digital sovereignty.
AI Robotics for the Real World
3 年Hi Alexander Thamm I agree that AI is leaving labs and entering the “real world” now. To add on your reinforcement learning part: on my first two days @Covariant I have seen robots performing picking and placing tasks at human level in distribution centers - even with objects they have never seen before. Exciting times ahead!
Great article and overview - yes we definitely should strengthen our digital sovereignty Vanessa Cann Daniel Abbou Dat Tran Andreas Weiss