At Rise of AI 2026 in Berlin, Prof. Dr. Hans Uszkoreit of DFKI revisits a talk he gave at an early, then little-known edition of Rise of AI nine years earlier, titled 'Superintelligence: Will machines be smarter than us?', and argues that the field has changed completely since then, from narrow, knowledge-engineered AI to large language models trained on text, a shift triggered by the transformer paper published a month after his original talk. He argues today's LLMs have super-knowledge and super-eloquence far beyond any human, but lack autonomy, memory and continual learning, comparing them to a narcoleptic Alzheimer's patient. Using the Cattell-Horn-Carroll model of fluid and crystallized intelligence, he argues that current benchmarks measuring how close GPT-4 and GPT-5 are to human intelligence are flawed because they only test tasks humans can already do. He surveys new memory-architecture research at Google, Stanford, Nvidia and by Sepp Hochreiter as the likely path to closing the remaining gap, calls the resulting systems very dangerous, dismisses claims that LLMs use excessive energy or water as illusions, and argues Europe can still catch up in frontier AI with one to three well-funded, well-staffed labs, but is currently talking rather than acting.
Key takeaways
- Uszkoreit argues today's LLMs have 'super-knowledge' and 'super-eloquence' far beyond any human but lack autonomy, memory and continual learning.
- He argues that GPT-4/GPT-5 benchmarks understate their capability because they only test tasks humans can already do, not superhuman tasks.
- New memory architectures at Google, Stanford/Nvidia and from Sepp Hochreiter's xLSTM work are the likely path toward closing the remaining intelligence gap.
- He calls the resulting technology very dangerous, potentially controllable only by AI itself, and says mastering it is unrelated to political sovereignty.
- He argues Europe could still catch up in frontier AI with just one to three well-funded labs offering top talent real compute for experimentation, not just model training.
Detailed notes
From narrow AI to the transformer era, nine years on
Uszkoreit opens by recalling that nine years earlier, at a then little-known, early Rise of AI conference, he gave a talk titled 'Superintelligence: Will machines be smarter than us?' Back then AI was narrow: no generative AI, no foundation models, no LLMs, and every system could do only one thing because knowledge could not be gotten into machines except through manual knowledge engineering. He notes that one month after that talk, the transformer paper appeared, and even its authors did not know how much potential it held; within a few years, knowledge stopped being the bottleneck.
How humans learn language versus how LLMs learn it
He distinguishes two human learning modes: behavioral/routine skills (walking, dancing, swimming), which require repeated exposure, and explicit knowledge learned through language (school, books, television), which can stick after being told just once. Children acquire the full complexity of grammar before ever entering explicit instruction, learning language through situations with parents where meaning is present alongside words. LLMs, by contrast, learn language behaviorally, through probabilities trained layer over layer across millions or billions of examples, but purely from textual context rather than situational meaning, so they understand only the world of text, indirectly.
Super-knowledge without autonomy or memory
Uszkoreit argues today's AI has 'super-knowledge' exceeding what even 50,000 well-selected humans could match, and 'super-eloquence' writing job applications, novels, poems and haikus in over 100 languages better than any human. But it lacks autonomy and memory, which he compares to a narcoleptic Alzheimer's patient: it knows old knowledge but forgets anything new instantly, and 'falls asleep' after each answer until woken by a new question. Agents compensate by calling databases, retrieval-augmented generation, the internet or other models, and giving a context memory, but he stresses these systems still do not think, plan or learn through interaction, and lack interpersonal social skills, common sense, instincts, drives, emotions, embodiment, sensory input and consciousness.
Redefining superintelligence and measuring the gap
He contrasts Nick Bostrom's definition of superintelligence, an intellect vastly smarter than the best human brain in practically every field including scientific creativity, general wisdom and social skills, which he says is nowhere near being reached, with a narrower definition: an intellect exceeding humans on the dimensions essential to measuring intelligence. Using the Cattell-Horn-Carroll psychometric model, where fluid and crystallized intelligence together account for over 90% of measured intelligence (crystallized alone roughly 80-85%), he shows spider-chart comparisons of GPT-4 and GPT-5 against human performance and criticizes the tests used: they only measured tasks humans can solve, not tasks exceeding human ability, so GPT-5 appears not to reach human level in some social and cultural categories despite far exceeding human knowledge. He identifies memory, specifically working memory, as the most defining missing piece of human-like intelligence in current systems.
New memory architectures as the path forward
He surveys frontier research aimed at adding real, learning memory to transformers: Google's approach adds memory as a neural network that learns continuously while operating, rather than a static context window; Stanford and Nvidia's test-time-training (TTT) research uses partially recurrent networks that learn while being tested; Moonshot AI's Kimi model uses a 'Kimi Delta attention' variant with recurrent components; and Sepp Hochreiter has extended LSTM architectures (citing Hochreiter and Schmidhuber's earlier work) into an xLSTM incorporating transformer-like power, though Hochreiter says he lacks the funding to prove it beats the competition. Uszkoreit guesses Google's approach will arrive first since it builds on existing transformer power, but bets recurrent neural networks will win long-term because they resemble how the human brain works. He also flags catastrophic forgetting/model collapse as the core difficulty in continual learning: a system trained on thousands of repeated occurrences cannot easily update on a single new fact without either ignoring it or destabilizing existing knowledge.
Risk, energy myths and Europe's choice
Asked implicitly how dangerous this path is, Uszkoreit says he does not like to create panic but believes it is very dangerous, dangerous enough that only AI itself may be able to control it, and argues mastering the technology is not about political sovereignty but a matter of not being reckless. He dismisses claims that LLMs consume excessive energy or water as illusions, arguing that concentrating hundreds of millions of users' compute at data centers makes energy use look large, but printing one book per GPT-4/GPT-5 user, or giving each user a single egg, would consume more energy than the system's training. On Europe, he sees modest hope for B2B application uptake but little hope of catching up to the US and China overall, referencing the EU's Next Frontier AI initiative and a German initiative he says is worthwhile but slow (a 24-month program). He argues catching up requires only one to three labs with compute power comparable to Google or Meta, staffed by decently paid local, repatriated or newly attracted top talent, allowed to freely experiment rather than being tasked or supervised; he warns that relying on big industry (citing SAP's shifting AI strategy as an example) or small industry, which gets acquired by American firms before it can scale, will not work, and that current EU compute plans, sized only for training a single model, are insufficient for enabling the experimentation talent needs.
Quotes
These systems are so smart and they have almost nothing.
Hans Uszkoreit
We are closer to that than we think.
Hans Uszkoreit
Questions answered in this talk
Why does Hans Uszkoreit say current AI systems lack real intelligence despite vast knowledge?
According to Hans Uszkoreit at Rise of AI 2026, LLMs have super-knowledge and super-eloquence exceeding any human, but they lack autonomy, memory and continual learning, comparing them to a narcoleptic Alzheimer's patient that forgets everything new and 'falls asleep' between questions.
What new research does Hans Uszkoreit point to as the path toward superintelligence?
According to Hans Uszkoreit at Rise of AI 2026, memory-architecture research at Google (continuously learning memory), Stanford and Nvidia (test-time training with recurrent networks), Moonshot AI's Kimi model, and Sepp Hochreiter's xLSTM work are the leading approaches to giving language models real working memory.
Does Hans Uszkoreit think Europe can still catch up in frontier AI?
According to Hans Uszkoreit at Rise of AI 2026, yes, but only with one to three labs given compute power comparable to Google or Meta and top talent free to experiment, since current EU funding plans are sized only for training a single model, not for open-ended research.