• Understanding AI: start with an answer you can check
    Jun 15 2026
    Imagine an associate preparing a memo. An assistant produces fluent text with an appealing reference. She opens the source and cannot find the claimed passage. This is an invented teaching scenario, not a report about a real firm. In our exercise, the associate separates the document’s contents from the assistant’s additions. She checks the reference and retains uncertainties. Begin with a specific, checkable task rather than a promise to replace a professional. Which result could you verify independently in your next trial? Revised and re-recorded 29 October 2026. Fictional teaching scenarios; synthetic voices. Episode text and sources: https://univ-ai.net/podcasts/en/episodes/s1e1-quest-ce-que-ia/ Sources: https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf
    Show More Show Less
    1 min
  • Learning from which examples?
    Feb 16 2026
    A fictional team wants to classify documents. It has labelled examples, but several categories conflict. Another group wants to discover clusters without reference answers. These are different questions. For our exercise, define success in advance without claiming a universal percentage. Which example would expose a weakness missed by the usual demonstration? Revised and re-recorded 29 October 2026. Fictional teaching scenarios; synthetic voices. Episode text and sources: https://univ-ai.net/podcasts/en/episodes/s1e2-familles-ia-apprentissage/ Sources: https://www.deeplearningbook.org/
    Show More Show Less
    1 min
  • Choose a concrete use before the tool
    Feb 16 2026
    A fictional firm considers summarising a file, sorting documents and drafting a letter. Instead of announcing a gain, it chooses one task with a checkable result. Which task could you stop without disrupting the firm if the test disappoints? Revised and re-recorded 29 October 2026. Fictional teaching scenarios; synthetic voices. Episode text and sources: https://univ-ai.net/podcasts/en/episodes/s1e3-applications-professionnelles/ Sources: https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10
    Show More Show Less
    1 min
  • A correct answer, but for which context?
    Feb 16 2026
    In our fictional scenario, a memo correctly answers an earlier contract version. The reviewer is working on a different version. The problem is not merely a false sentence; it is the wrong reference document. In the workshop, ask a second reviewer to repeat the check. Do they have the same documents? Can they locate the cited passage? The key question is: exactly what supports our conclusion? Revised and re-recorded 29 October 2026. Fictional teaching scenarios; synthetic voices. Episode text and sources: https://univ-ai.net/podcasts/en/episodes/s1e4-limites-responsabilite/ Sources: https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10
    Show More Show Less
    1 min
  • AI history can be checked in original research
    Feb 16 2026
    In June 2017, Ashish Vaswani and seven coauthors submitted Attention Is All You Need. Their paper presents the Transformer architecture and reports translation experiments. It describes specific research, not the entire history of AI. Spectacular claims are removed when precise evidence is missing. Collaboration, experiments and limitations also make this history interesting. Which discovery would you like to understand through its original work rather than a legend? Revised and re-recorded 29 October 2026. Fictional teaching scenarios; synthetic voices. Episode text and sources: https://univ-ai.net/podcasts/en/episodes/special-grands-mathematiciens-ia/ Sources: https://arxiv.org/abs/1706.03762
    Show More Show Less
    1 min
  • Before sending a file, follow the data path
    Feb 16 2026
    A fictional professional removes a name but leaves an exact date, a location and a rare situation. She is about to send the document to an assistant. The question is whether those details still identify someone. Request evidence for each answer: configuration, contract, log or technical description. A country, logo or marketing statement does not replace that record. Can you explain where each data item goes before clicking? Revised and re-recorded 29 October 2026. Fictional teaching scenarios; synthetic voices. Episode text and sources: https://univ-ai.net/podcasts/en/episodes/special-secret-professionnel-ia-cloud/ Sources: https://www.cnil.fr/fr/technologies/lanonymisation-de-donnees-personnelles
    Show More Show Less
    1 min
  • The right measure for the right question
    Feb 16 2026
    A fictional supplier advertises 95% accuracy. We do not yet know what was counted, on which documents or at which threshold. The number belongs to the scenario, not an existing product measurement. Our exercise is to draw the results table and state what it cannot establish. Which missing measure would tell you whether the test actually addresses your need? Revised and re-recorded 29 October 2026. Fictional teaching scenarios; synthetic voices. Episode text and sources: https://univ-ai.net/podcasts/en/episodes/s2e1-pourquoi-maths-essentielles/ Sources: https://scikit-learn.org/stable/modules/calibration.html
    Show More Show Less
    1 min
  • Document similarity: what does the score actually measure?
    Feb 16 2026
    Two fictional contracts concern the same service. One permits termination; the other forbids it. A tool ranks them close together: similar wording hides a decisive negation. This invented example invites us to look beyond the ranking. For our test, prepare contrasting pairs: similar vocabulary with different meanings, then different vocabulary with similar meanings. Review the obligation, exceptions and context. The useful question is: which important differences does the ranking miss? Revised and re-recorded 29 October 2026. Fictional teaching scenarios; synthetic voices. Episode text and sources: https://univ-ai.net/podcasts/en/episodes/s2e2-algebre-lineaire-donnees/ Sources: https://scikit-learn.org/stable/modules/generated/sklearn.metrics.pairwise.cosine_similarity.html
    Show More Show Less
    2 mins