At Rise of AI 2018 in Berlin, Christian T. Schäfer, head of the Deloitte Analytics Institute, argues that AI is a double-edged promise: enormous business value alongside real risks around trust, liability, security and control. He frames AI as a wide array of technologies rather than one thing, distinguishes today's narrow AI from a hoped-for general AI, and traces a hardware-driven history from the web and cloud to falling storage costs and specialized chips like Google's TPUs. He lays out risks including autonomous weapons, an AI arms race, loss of control over robust self-preserving systems, and data privacy, alongside potential upside such as recursive self-improvement, help with war, disease and poverty, and automation combined with universal basic income. He explains deep learning as a subset of machine learning that handles messy real-world data well, cites Deloitte's client demand data across five use-case areas, and closes by urging that a human-centered front end and process design, not just algorithms, determine business outcomes.
Key takeaways
- AI is not one technology but a wide array of methods, and today's systems are narrow AI, not the general AI researchers are pursuing.
- Storage costs fell by a factor of about 17 million over 35 years, and Google's TPUs are roughly 30 times more performant than GPUs for machine learning.
- Autonomous weapons and the risk of losing control over robust, self-preserving AI systems are already present risks even with narrow AI.
- Only 0.5% of all data is ever analyzed or used, despite a projected 44 trillion gigabytes of global data production by 2020.
- Deloitte's top five client AI use-case demands are NLP with OCR, early warning systems, knowledge management, computer vision, and production quality optimization.
Detailed notes
What AI actually is
Schäfer opens by framing the buzz around AI as distracting: there is no single definition of AI, only a wide array of technologies combined to make machines act with apparent human-like intelligence. Today's AI is narrow AI, capable of specific tasks like object recognition, facial recognition or driving a car, while researchers are working toward a general AI that could match humans across many cognitive tasks at once, something current systems cannot do.
A hardware-driven history
He describes AI's history as a roller coaster of optimism after breakthroughs and disappointment at impasses, citing the web in the 1990s and cloud computing in the 2000s as inflection points. Storage costs fell by a factor of about 17 million over 35 years, enabling much larger unstructured data sets. Specialized computing hardware also drove progress: GPUs outperform CPUs for the matrix multiplications neural networks rely on, and Google's TPUs are roughly 30 times more performant than GPUs for machine learning while being more energy efficient. The underlying algorithms, he notes, have existed for decades; what changed is the hardware and the scale of training data.
The double-edged promise: risks and upside
Schäfer lists concerns around trust, liability, security and control, including autonomous weapons as AI systems built to kill, the risk of an AI arms race, and the danger that self-preserving, resource-acquiring systems built to be robust against attack could become impossible for people to control, a risk he says is already present with narrow AI. On data privacy he stresses the need to control and verify what data is collected and to instill and negotiate values with machines. On the positive side, he points to AI already automating and augmenting business processes and running critical infrastructure like power grids, and speculates that recursive self-improvement could produce a superintelligence that helps eradicate war, disease and poverty, with automation combined with universal basic income easing the transition for human labor.
Deep learning and dirty data
He describes deep learning as a subset of machine learning, using layered models that learn data representations without hand-designed features, where depth relates to problem complexity rather than model complexity. A key strength, he says, is robustness to dirty data, badly parsed, spotty or duplicate records, which machine learning can process with far greater accuracy than traditional methods, and which scales well as data volumes grow. He cites that only 0.5% of all data is ever analyzed or used, against a projection that by 2020 each person would produce 1.7 megabytes of data per second, totaling 44 trillion gigabytes.
Making AI work in the enterprise
Drawing on Deloitte's 106-person analytics team and hundreds of client use cases, Schäfer lists the top five AI demand areas from recent client requests: natural language processing with OCR, early warning systems, knowledge management, computer vision, and production quality optimization. He describes clients moving from rule-based, judgment-driven systems toward data-centric architectures, where cataloged, well-described data becomes a usable asset, a concept he calls a data asset inventory. He closes by arguing that the human interface and human-machine collaboration design, not the algorithm alone, determine whether AI delivers real business and customer value.
Quotes
The tale of AI is kind of a roller coaster ride, its optimism with every major breakthrough and disappointment when reaching an impasse.
Christian T. Schäfer
Only 0.5% of all data is ever analyzed or used.
Christian T. Schäfer
A journey towards AI starts from the front end, from the human interface.
Christian T. Schäfer
Questions answered in this talk
What is the difference between narrow AI and general AI?
According to Christian T. Schäfer at Rise of AI 2018, narrow AI performs specific tasks like object recognition, facial recognition or driving, while general AI, still a research goal, would match human-level performance across many cognitive tasks at once.
How much of the world's data is actually used?
According to Christian T. Schäfer at Rise of AI 2018, only about 0.5% of all data is ever analyzed or used, even as global data production was projected to reach 44 trillion gigabytes by 2020.
What are the main risks of AI according to this talk?
According to Christian T. Schäfer at Rise of AI 2018, the key risks are around trust, liability, security and control, including autonomous weapons, an AI arms race, and the danger of losing control over robust, self-preserving AI systems.