Aldo Gangemi — Face It! (ULiège)

Aldo Gangemi — Face It! (ULiège)

by Aldo Gangemi · View the channel on PSI

Formal & Physical Sciences AI — News & Applications UYComputer scienceUYQArtificial intelligence (AI)
🎙 Aldo Gangemi 👥 178 📅 February 27, 2023 ⏱ 39 min 👁 47 📄 expert opinion 🧭 2026-10-09
Available in: English (current) Français

Keywords

groundingmultimodalGPT-3ChatGPTDALL-E

Summary

Aldo Gangemi, professor at the University of Bologna, gives a lecture on the varieties of multimodal grounding in AI. He shows how recent generative models like GPT-3, ChatGPT, DALL-E, and Stable Diffusion, while impressive in many tasks, fail on simple spatial and topological reasoning. Using concrete examples, he illustrates that these systems often lack understanding of basic relations such as ‘on top of’ or ’left of’, and they struggle with inferential knowledge that humans find trivial. He also discusses the Moravec paradox, where AI excels at tasks considered difficult and fails at easy ones. Gangemi argues that a hybrid approach combining symbolic and sub-symbolic methods is needed to improve grounding. He contrasts proprietary models with open-source alternatives like Stable Diffusion, and highlights the importance of provenance in AI outputs, mentioning Perplexity AI as a better option in this regard. The talk concludes by suggesting that AI models are far from achieving human-like understanding and that foundational issues remain unresolved.

159 words

Critical Evaluation

Value of the Information & Strength of the Argument

The lecture provides valuable insights into the limitations of current AI models through numerous concrete examples, making the argument more convincing. Gangemi effectively demonstrates the gap between statistical pattern matching and genuine semantic understanding. He illustrates how models like ChatGPT can produce fluent text but fail on simple inferential tasks, such as inferring that Adolph Sax, the inventor of the saxophone, must have played it. The argumentation is structured around case studies that highlight the lack of topological and ontological grounding. He also brings in the Moravec paradox to contextualize these failures, suggesting that AI’s competence does not align with human intuitions about task difficulty. However, the examples are anecdotal and not systematically quantified, and the speaker acknowledges that the field is evolving rapidly, which somewhat weakens the long-term validity of his claims.

Scientific Rigor, Source Quality, Title Accuracy

Gangemi references several sources and figures, including Gary Marcus as a critic of deep learning, and mentions the Moravec paradox. He also points to Perplexity AI as a tool that provides provenance. However, he does not provide specific citations or data to support his claims, relying instead on his own tests and observations. The talk is a perspective from a leading expert, but lacks formal empirical rigor. The title ‘Face It!’ is a reference to the symposium’s theme, but the actual content is more about general AI grounding, which may be slightly misleading. The speaker himself apologizes for not tailoring the talk to the event’s focus on faces. Overall, the sources are mentioned but not thoroughly documented, and the title-content alignment could be improved.

272 words

Title / Content Match

The title refers to the symposium name 'Face It!' but the content is broader on AI grounding. There is some mismatch, but the talk is part of the event and the subtitle clarifies the topic.

Quality & Reliability

7/10

The lecture is given by a recognized researcher in AI and semantics, but the claims are based on personal tests and examples rather than systematic empirical data. The author openly acknowledges the ongoing evolution of the models.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

  • Gary Marcus — Mentioned as a relentless critic of deep learning and a source of examples of GPT-3 failures.
  • Moravec's paradox — Referenced to explain why AI can be good at complex tasks but fail at simple ones.
  • Perplexity AI — Mentioned as a useful AI tool that provides provenance for its answers, unlike ChatGPT.

Contribution & Novelties

Gangemi brings a fresh perspective by systematically testing current generative models on simple spatial and topological problems, highlighting the gap between their impressive language abilities and lack of grounding. He shares concrete examples from his own research, showing that these models often produce plausible but incorrect results. He also suggests that a hybrid symbolic-sub-symbolic approach is necessary to overcome these limitations, which is a valuable position in the ongoing debate.

Pour aller plus loin :

  • Moravec’s paradox — Reference for the paradox that AI fails at tasks humans find easy.
  • Symbol grounding problem — Foundational issue of how symbols acquire meaning, directly related to the lecture’s theme.
  • Conceptual blending — A theory of cognition that Gangemi mentions in relation to AI’s ability to blend concepts creatively.

126 words

Radar Profile

The radar profile shows moderate scores across all dimensions, with a slight emphasis on quantitative and qualitative information, but a lower technical level. This reflects a lecture that is informative but not deeply technical, and relies on expert opinion rather than rigorous data.

Reliability 7/10