At Rise of AI 2017 in Berlin, Hans Uszkoreit, Scientific Director at the German Research Center for Artificial Intelligence, asks not how to build superintelligence but whether it can exist at all, and if so, where. He argues AI already shows superhuman performance in narrow tasks, citing Deep Blue at chess, IBM Watson at Jeopardy, AlphaGo, and self-driving cars, but distinguishes this from genuine intelligence. He breaks human cognition into innate instincts, acquired explicit knowledge such as facts and symbolic sequences, and acquired implicit knowledge such as language and procedural skills, then maps these onto two separate families of AI systems: symbolic, knowledge-based systems and statistical, learned-behavior systems such as neural networks. He argues these two families are not yet combined: neural networks such as machine translation systems produce results without understanding, lack reusable knowledge, and cannot explain their own decisions. Uszkoreit predicts that combining learned behavior with large, growing knowledge graphs, rather than any single self-conscious system, will drive the next leap in AI capability, and rejects Hollywood scenarios of an uncontrollable evil AI, while expecting eventual benefits including augmented human cognition and longer, safer lives.
Key takeaways
- Uszkoreit distinguishes superhuman task performance, already achieved by Deep Blue, Watson, AlphaGo and self-driving cars, from genuine intelligence.
- He argues neural networks and symbolic knowledge systems remain two separate, unconnected families in current AI.
- He calls Google's machine translation a mechanical parrot that translates without understanding any word.
- He predicts the next AI leap will come from combining learned behavior with runtime access to large knowledge graphs like DBpedia.
- He rejects Hollywood-style evil AI scenarios, since current systems lack self-models and genuine drive.
Detailed notes
Superhuman performance versus superintelligence
Uszkoreit opens by reframing the debate: instead of explaining how to build superintelligence, he asks whether superintelligence is possible and where its limits lie. He references Vernor Vinge's 1993 prediction that within 30 years humanity would have the technological means to create superhuman intelligence, soon after which the human era would end, and notes Stephen Hawking's warning that robots could evolve faster than humans and become hard to stop, and Elon Musk's frequently cited warnings about AI. He argues AI has already achieved superhuman performance on narrow tasks: Deep Blue beating human chess players, IBM Watson defeating human Jeopardy champions on a fair, internet-disconnected setup, AlphaGo beating a world champion, and Google's self-driving cars reaching a driving standard US authorities judged better than the average human driver. He frames these as proof of superhuman performance, but asks whether they reflect real intelligence or simply raw computing power, using Deep Blue's move calculation as an example.
Three P's and the limits of the brain analogy
For companies, he frames AI adoption around three P's: propaganda, where systems like AlphaGo and IBM Watson first appear as demonstrations; products, where technologies like Watson become core offerings; and processes, where AI improves cost savings, automation, customer relationship management, logistics and maintenance even for companies that never sell an AI product directly. He warns against the common comparison between artificial neural nets and biological neurons, saying the simple units in neural networks do not resemble the highly differentiated neurons of the human brain, and that this analogy overstates how close AI is to the brain.
Two families of AI systems
Uszkoreit breaks human cognition into innate instincts inherited through evolution, acquired explicit knowledge such as facts, theories, methods and symbolic sequences, and acquired implicit knowledge such as a first language and procedural skills like swimming, which are learned through repetition rather than explanation. He maps these onto two separate families of AI system: symbolic, knowledge-based systems built on rules, such as an old expert system used by American Express for credit checks, and statistical, learned-behavior systems such as neural networks, which resemble the reactive, non-symbolic side of human cognition. He argues these two families remain largely unconnected in current systems, though some research labs are already working to combine them, which he calls the coming revolution rather than incremental gains on benchmark scores.
Why learned behavior lacks knowledge
He argues machine-learning systems have no reusable knowledge and cannot explain their own decisions because their outputs are not grounded in explicit facts. Google's machine translation, he says, does not understand a single word but produces translations because it has been trained on millions of human-translated sentence pairs, calling it a very good mechanical parrot. He compares this to a hypothetical doctor trained only on pairs of symptoms and diagnoses, never allowed to study medicine, who could develop intuition but no explainable knowledge. He notes a branch of machine learning tries to extract explicit rules from trained neural networks, which he calls a fruitful but currently unsuccessful direction, and says understanding a problem is close to solving it, while systems without grounded knowledge hit a barrier at this point.
Knowledge graphs, predictions and the absence of an evil AI
Uszkoreit points to fast-growing knowledge graphs, comparing Wikipedia's scale to the Encyclopedia Britannica and citing DBpedia, a project he works with, as holding more than a billion facts, alongside Google's knowledge graph. He argues the next leap will come from systems that access such large external knowledge bases at runtime rather than trying to store everything inside a single neural network. He predicts that superhuman decision-making support will arrive quickly, but rejects Hollywood scenarios of an evil AI looking for an off switch, since current systems lack self-models, reflection and genuine drive. Unlike human behavioral knowledge, which dies with the individual, machine-learned knowledge can be copied and implanted across many devices. He closes by predicting that combining learned behavior with grounded knowledge will unlock major economic reserves, augment rather than replace human cognition, liberate people from alienated labor and lead to a longer, safer life.
Quotes
Google machine translation does not understand a single sentence, not even a word, but it translates because it has been trained on millions of sentence pairs translated by humans.
Hans Uszkoreit
If we have understood a difficult problem, we have almost solved it.
Hans Uszkoreit
Questions answered in this talk
Is current AI superintelligent?
According to Hans Uszkoreit at Rise of AI 2017, AI already shows superhuman performance on narrow tasks such as chess, Jeopardy and driving, but current systems lack grounded knowledge, self-reflection and genuine drive, so he does not consider them intelligent in the human sense.
Will AI become an uncontrollable evil AI?
According to Hans Uszkoreit at Rise of AI 2017, he rejects Hollywood scenarios of an evil AI searching for an off switch, arguing current systems have no self-consciousness, no models of themselves or the world, and no drive of their own.
Why can't neural networks explain their decisions?
According to Hans Uszkoreit at Rise of AI 2017, neural networks such as machine translation systems produce outputs learned from statistical patterns rather than grounded knowledge, so they cannot explain why they made a given decision.