Notes

Selected from my LinkedIn posts, in their original French or English.

  1. Music similarity, twenty-four years later

    24 years later! Jean-Julien Aucouturier and I are delighted to receive a Test-of-Time Honorable Mention at ISMIR 2026 for our 2002 paper “Music Similarity Measures: What’s the Use?”…

    Read noteClose note

    24 years later! Jean-Julien Aucouturier and I are delighted to receive a Test-of-Time Honorable Mention at ISMIR 2026 for our 2002 paper “Music Similarity Measures: What’s the Use?”

    Back then, at Sony CSL Paris, we were exploring a simple question: how can we measure whether two pieces of music sound similar? And what can we actually do with such a measure?

    We had great fun recording this conversation, looking back at those early days of music information retrieval, our experiments, and how the field has evolved since.

    A great opportunity to revisit a research adventure that began almost a quarter of a century ago.

    Many thanks to the ISMIR community for this recognition!

    Here's our conversation 👇

  2. Imiter n’est pas encore créer

    Tout admirateur de Proust connaît Pastiches et Mélanges, ce livre préliminaire dans lequel Proust démontre sa maîtrise des grands auteurs qui l'ont précédé : Flaubert, Balzac, Saint-Simon et quelques autres. Une étape sans doute nécessaire …

    Read noteClose note

    Tout admirateur de Proust connaît Pastiches et Mélanges, ce livre préliminaire dans lequel Proust démontre sa maîtrise des grands auteurs qui l'ont précédé : Flaubert, Balzac, Saint-Simon et quelques autres.
    Une étape sans doute nécessaire : avant de créer, il faut peut-être savoir imiter.
    Lors des Assises de l'IA à Caen avec Cédric Villani et Sylvia Serfaty, je me suis permis de prendre cet exemple pour suggérer que nous en sommes collectivement, à ce même stade avec l'intelligence artificielle.
    Nous savons produire des pastiches et des mélanges, avec une virtuosité stupéfiante.
    Mais imiter n'est pas encore créer.

    Nous attendons toujours notre Recherche du temps perdu.

  3. Beyond gradient search

    Now that we master the art of representing symbols as vectors, AI is dominated by gradient search.
    Powerful, but unfortunate as it prevents many new ideas about how to build intelligent machines to take place.

  4. A population needs more than plausible individuals

    Our new paper on synthetic populations is out in Frontiers in Artificial Intelligence.

    Read noteClose note

    Our new paper on synthetic populations is out in Frontiers in Artificial Intelligence.

    Synthetic populations cannot rely on plausible individual personas alone. They also need a solid mathematical foundation ensuring statistical coherence at the population level.
    How can we generate synthetic populations that satisfy hundreds of statistical constraints, including correlations between multiple attributes?

    We propose a method based on maximum entropy, and compare it with existing approaches on real survey statistics.
    This technique is already at work in our technologies at ImagineAllThePeople (IATP), across a wide range of use cases combining mathematical modeling and generative AI.

    Our paper: https://lnkd.in/e-DWB5Pp
    Our technology: https://lnkd.in/eJPduGtx

  5. Learning a musical language, controlling its structure

    I’ve just released MarkovConstraints on npm, together with the TypeScript project vo_bp_regular_ts. Generate sequences from a variable-order Markov model while satisfying global constraints exactly. This is the TypeScript/browser counterpart of…

    Read noteClose note

    I’ve just released MarkovConstraints on npm, together with the TypeScript project vo_bp_regular_ts.
    Generate sequences from a variable-order Markov model while satisfying global constraints exactly.
    This is the TypeScript/browser counterpart of my work on exact regular-constrained sampling for variable-order Markov models, now packaged so it can be used directly in web applications.
    👉 Project + interactive demos:
    https://lnkd.in/eixqzxNz

    And, for something more musical: I used it to build a Guillaume de Machaut melody generator.
    The model learns local melodic patterns from Machaut, while constraints can impose things such as length, pitch range, final note, cadence, phrase structure, etc.
    You can generate new Machaut-like melodies here:
    🎵 https://lnkd.in/e29d2_Hu

    What I find interesting is the separation between what is learned (the local statistical language of the repertoire), and what is constrained (the larger-scale musical structure we want the generated piece to satisfy).
    This is the kind of combination I’ve been interested in for a long time: probabilistic generation + explicit control, rather than either pure sampling or purely rule-based composition.
    The library is open source, and I’d be very interested in other applications where constrained Markov generation could be useful.

  6. The final piece of a long research puzzle

    Happy to share that my paper, “Exact Regular-Constrained Sampling for Variable-Order Markov Generation,” has been accepted to the NeurIPS 2026 Main Track.

    Read noteClose note

    Happy to share that my paper, “Exact Regular-Constrained Sampling for Variable-Order Markov Generation,” has been accepted to the NeurIPS 2026 Main Track.

    This solves a problem I had been trying to address for a long time: how to combine variable-order Markov models with regular constraints while retaining exact sampling.

    The code is available on github: https://lnkd.in/eVypgFvH

    With this library one can build variable-order systems (e.g. continuators) in a few lines of code.

    This NeurIPS paper feels like the final piece of a rather long research puzzle.
    It closes a chapter on Markov Constraints, a line of research that started fifteen years ago.

    Two other publications complement this story:
    – The paper "Markov constraints revisited: commentary on “Markov constraints: steerable generation of Markov sequences” in Constraints Journal (Springer), looks back at our original 2011 paper, its motivation and what happened to the idea since.
    – A much longer, roughly 110-page chapter, forthcoming in the Springer volume edited by Cédric Villani’s about the proceedings of Festum Pi 2025, where Iwe develop the approach and its broader context in more detail.

    Looking forward to presenting it at NeurIPS.

  7. Leaving the jungle era of AI music

    In 2016, we trained an AI system on the Beatles and released Daddy’s Car with Flow Machines at Sony CSL. Almost nobody objected. Ten years later, AI-generated music has become the centre of a global debate about copyright, training data…

    Read noteClose note

    In 2016, we trained an AI system on the Beatles and released Daddy’s Car with Flow Machines at Sony CSL. Almost nobody objected.
    Ten years later, AI-generated music has become the centre of a global debate about copyright, training data, creativity, authorship and the future of musicians.

    I recently spoke with Joe Sparrow at Music Ally about what has changedn and why I think we are now leaving what I call the “jungle era” of AI music.

    We discussed the early days of Flow Machines, Daddy’s Car, the spectacular progress of systems such as Suno, and the strange situation we have reached today: production quality is extraordinarily convincing, while the deeper problems of musical generation structure, songwriting, control and genuine musical intention remain largely open.
    The interesting question is no longer “Can AI generate music?”
    Clearly, it can.
    The interesting questions are now:
    What kind of music should these systems generate?
    How should musicians control them?
    What does originality mean in this context?
    And what technical ideas do we still need beyond scaling up generative models?

    It was enjoyable to look back at almost ten years of AI music research from the perspective of 2026.

    👉 Music Ally interview:
    https://lnkd.in/eEX9HwBj

    Thanks to Joe Sparrow and Music Ally for the thoughtful conversation.

  8. Rediscovering plans, beyond the next move

    50 years ago, Jacques Pitrat proposed a new idea for chess programming: search through plans, not just moves. Plans can make the effective search space dramatically smaller. Why isn’t this central to chess today? And why isn’t it central to…

    Read noteClose note

    50 years ago, Jacques Pitrat proposed a new idea for chess programming: search through plans, not just moves.
    Plans can make the effective search space dramatically smaller.
    Why isn’t this central to chess today?
    And why isn’t it central to creativity in general? music, writing, design, mathematics, etc.

    Modern generative AI is very good at the next move. Maybe we should rediscover the plan, because it entails finding much deeper representations of the world.

    Jacques Pitrat, A Chess Combination Program Which Uses Plans, Artificial Intelligence 8(3), 1977, 275–321.

  9. From Gesualdo to a string quartet

    This summer, at Festum π in Chania, Crete, I had the great pleasure of presenting a rather unusual concert with the New Cretan Quartet.

    Read noteClose note

    This summer, at Festum π in Chania, Crete, I had the great pleasure of presenting a rather unusual concert with the New Cretan Quartet.

    The idea was simple: take extraordinary vocal music, from Gesualdo to the Beach Boys and Take 6, and ask whether algorithms could transform these complex vocal textures into convincing music for string quartet.

    Not by training a model to imitate a style or asking an LLM to “compose something like Gesualdo”. But by treating the problem as an algorithmic one: preserve as much as possible of the notes, harmony and voice-leading of the original, while finding a playable and meaningful realization for four string instruments.

    Many thanks to the wonderful New Cretan Quartet, and to Festum π for making this experiment possible.

    The full concert is now online:

    https://lnkd.in/e5FyE9Z2

  10. When AI’s early enthusiasts become its opponents

    How come some of the most ardent proponents of AI in the 1990s, when it barely worked, have become some of its most reactionary opponents now that it works beyond expectations?

  11. Peut-on rendre Spinoza exécutable ?

    Peut-on rendre une partie de la philosophie de Spinoza exécutable ?

    Read noteClose note

    Peut-on rendre une partie de la philosophie de Spinoza exécutable ?

    Je publie deux ressources complémentaires autour de Snarky, un nouveau moteur d’inférence symbolique en Python.

    1. Le moteur Snarky
    Snarky s’inspire du langage historique SNARK de Jean-Louis Laurière et du système BOOJUM développé par Jean-Luc Dormoy.
    Il permet notamment de manipuler des règles d’ordre 0, 1 et 2, des propositions imbriquées, des variables en position de relation, ainsi que la provenance et les chaînes de preuve produites par le raisonnement.
    Le projet vise aussi, à terme, à faire coopérer raisonnement symbolique et programmation par contraintes.
    Code, documentation et exemples :
    👉 https://lnkd.in/eazXhnpt

    2. Une application à la partie III de l’Éthique de Spinoza
    J’ai utilisé Snarky pour formaliser les propositions de la partie III de l’Éthique, consacrée à l’origine et à la nature des affects.
    Cette expérimentation s’inscrit dans la continuité des travaux de Michel Gondran, et du rapport de stage de Fabrice Cavarretta consacré à SpinoLog, une première formalisation de l’Éthique à l’aide de SNARK et de BOOJUM.
    Le projet actuel reprend et complète ce travail historique : il étend la formalisation à l’ensemble des 59 propositions de l’Éthique III, aux définitions des affects et à leurs relations, tout en ajoutant des chaînes de preuve, une provenance explicite des faits et des contre-cas permettant de tester les limites de chaque traduction formelle.

    Le site permet d’explorer les textes, les règles correspondantes, les affects, les dépendances entre propositions et les preuves produites par le moteur.
    Il ne s’agit pas de réduire la philosophie à un calcul, mais d’expliciter certaines conséquences du système spinoziste et de rendre discutables et testables les choix effectués lors de sa formalisation.

    Atlas interactif :
    👉 https://lnkd.in/eK-KkuyG
    Commentaires, critiques et contributions sont les bienvenus.

  12. AI music needs a composition stage

    I’m pleased to share this interview with AIhub about music generation with AI. We discuss what deep learning has changed, and what remains fundamentally unresolved. Systems such as Suno and Udio produce astonishing results, but they jump…

    Read noteClose note

    I’m pleased to share this interview with AIhub about music generation with AI.
    We discuss what deep learning has changed, and what remains fundamentally unresolved.
    Systems such as Suno and Udio produce astonishing results, but they jump directly from a prompt to a finished production. What is still missing is a genuine composition stage: forming intentions, exploring possibilities, detecting problems, backtracking, and refining ideas.

    I believe the next generation of creative AI will need to combine the power of statistical sampling with search and explicit reasoning.

    We also discuss why musical quality is so difficult to evaluate, why creativity has no simple loss function, and why researchers should sometimes pursue fascinating, ill-defined problems rather than follow established trends.

    Many thanks to Liliane-Caroline Demers and AIhub for this thoughtful conversation.

    https://lnkd.in/e-85RDxj

  13. Musical harmonization, twenty-five years on

    Twenty-five years after our survey Musical Harmonization with Constraints, Pierre Roy and I are pleased to publish a retrospective commentary in Constraints.

    Read noteClose note

    Twenty-five years after our survey Musical Harmonization with Constraints, Pierre Roy and I are pleased to publish a retrospective commentary in Constraints.

    The central question remains highly relevant: how can we generate music that is not only correct, but also interesting and controllable?

    We revisit the evolution from classical constraint programming to hybrid approaches combining constraints, statistical learning and neural models, an intellectual path that also led to projects such as the Continuator, Markov Constraints, FlowComposer and DeepBach.

    Many thanks to Eugene Freuder for inviting this commentary.
    Open-access link:
    https://rdcu.be/ftvXD

  14. Convincing is not necessarily interesting

    Why does AI-generated music often sound convincing, yet not necessarily interesting? I had the pleasure of discussing this question on the Creative Difference podcast, in an interview about AI, music, creativity, and the strange gap between…

    Read noteClose note

    Why does AI-generated music often sound convincing, yet not necessarily interesting?
    I had the pleasure of discussing this question on the Creative Difference podcast, in an interview about AI, music, creativity, and the strange gap between producing something that resembles music and producing something that truly captures our attention.
    We talked about what current generative systems do well, where they still fall short, and why “interestingness” may be one of the hardest dimensions of creativity to formalize.
    Thanks to Maxime Gabella for the conversation!

    🎧 Listen here: https://lnkd.in/ex82tcwp

  15. Reviving MusicSpace in the browser

    Reviving MusicSpace in the browser: constraint propagation meets spatial audio and musical control. Sources, listeners, trajectories, and constraints become a live interface for audio, MIDI-style generators, and synthesis parameters.

    Read noteClose note

    Reviving MusicSpace in the browser: constraint propagation meets spatial audio and musical control.
    Sources, listeners, trajectories, and constraints become a live interface for audio, MIDI-style generators, and synthesis parameters.

    3 video examples:

    Cycloid Percussion:
    Three bundled percussion loops are spatialized by nested cycloid-style motion.
    https://fpachet.github.io/musicspace/assets/videos/linkedin/musicspace-cycloid-percussion-linkedin.mp4

    Faust Control Study:
    Constraint-driven source motion controls oscillator, filter, resonance, and gain parameters.
    https://fpachet.github.io/musicspace/assets/videos/linkedin/musicspace-faust-control-study-linkedin.mp4

    Granular Cloud Study:
    Compound trajectories and constraints control a small granular synthesis cloud.
    https://fpachet.github.io/musicspace/assets/videos/linkedin/musicspace-granular-cloud-study-linkedin.mp4

    Live demo: https://fpachet.github.io/musicspace/
    Code: https://github.com/fpachet/musicspace

  16. Specifications are worth more than code

    We used to write specs to make code. Now we make code to write specs.
    Specs, precise, complete, clear are worth more than code.

  17. Finding the tonal logic of a chord sequence

    Analyzing musical sequences, specially jazz ones, is not trivial. It is usually seen as a minimization problem: minimize the number of modulations. Dynamic programming does minimize modulations, but can't minimize also the total number of…

    Read noteClose note

    Analyzing musical sequences, specially jazz ones, is not trivial. It is usually seen as a minimization problem: minimize the number of modulations.
    Dynamic programming does minimize modulations, but can't minimize also the total number of tonalities which result in diverging analysis.
    In this paper, I do both, and very fast (though it's NP-Hard in general) with thorough evaluations on thousands of sequences.
    Then you can compose/improvize with the right scales 🙂
    https://lnkd.in/eDmQNiyf

  18. LLMs and humanity’s zone of proximal development

    LLMs explore exactly the "zone of proximal development" of humanity, as conceptualized by Vygotsky.

  19. Plausibility is not enough: generating under constraints

    Very happy to announce that our paper “Constraining Generative Models: A Survey from the Constraint Programming Perspective”, by Bonlarron, Pachet, Roy and Régin, has been accepted at IJCAI 2026 for the survey track. This is another strong signal…

    Read noteClose note

    Very happy to announce that our paper “Constraining Generative Models: A Survey from the Constraint Programming Perspective”, by Bonlarron, Pachet, Roy and Régin, has been accepted at IJCAI 2026 for the survey track.
    This is another strong signal that the combination of generative probabilistic models and constraints is a central question for the next generation of AI systems.
    Generative models are impressive because they can produce plausible continuations, images, texts, melodies, programs, designs. But in most serious applications, plausibility is not enough.
    We want the generated object to satisfy constraints. It must respect a structure, a style, a scenario, a budget, a grammar, a musical form, a legal requirement, a user intention, a set of partial specifications, or a collection of hard and soft constraints.
    In short: generation is constrained sampling. In fact, creativity is probably only about that!
    This is a question we started exploring back in 2011 in the context of music generation, combining Markov models and constraint solving. It is fascinating to see how relevant this issue has become today, now that generative AI is everywhere.
    Constraint Programming has much to contribute here: modeling, propagation, search, global constraints, optimization, and a deep culture of explicit control.
    I hope this survey helps clarify the landscape and encourages more bridges between the CP and generative AI communities.

  20. The world has become one huge Media Lab

    With Claude and Codex, the world has become one huge Media Lab.

  21. A Continuator that respects musical constraints

    Continuator is now available as a Hugging Face Space at: https://lnkd.in/ebE4S5ya…

    Read noteClose note

    Continuator is now available as a Hugging Face Space at:
    https://lnkd.in/ebE4S5ya

    Continuator is a real-time interactive music generation system: you play a phrase, and the system listens, learns, and continues in the same musical style. The new web interface makes it much easier to demonstrate and use, with MIDI input/output and an interactive client-server architecture.

    What makes this version especially interesting is its positional constraint mechanism. To our knowledge, Continuator is currently the only interactive music generation system able to satisfy positional constraints "exactly" during generation. For example, one can require that a generated phrase ends on a terminal note, or impose constraints on notes at specific positions in the sequence. The system generates continuations that satisfy these constraints and samples them with the right probabilities, rather than hoping to approximate them.

    This is important because constraints are everywhere in music: endings, cadences, rhythms, melodic targets, formal structures. Yet most generative AI systems are weak precisely when one asks them to obey such constraints reliably. This is the subject of this recent paper about hidden biases:
    https://lnkd.in/e6dQeRV9

    More soon, including a “constraint studio” to make this capability visible and playable and new engines that are more creative, yet do not require huge training sets and are explainable.

  22. Est-ce que cette musique touche les gens ?

    “Est-ce qu’il y a de l’IA ?” devient une question de moins en moins intéressante. La vraie question est : est-ce que cette musique touche les gens ? RTL m’a interrogé sur Willylancien, artiste IA dont le titre Magique vient d’entrer en playlist…

    Read noteClose note

    “Est-ce qu’il y a de l’IA ?” devient une question de moins en moins intéressante.
    La vraie question est : est-ce que cette musique touche les gens ?
    RTL m’a interrogé sur Willylancien, artiste IA dont le titre Magique vient d’entrer en playlist radio.
    Mon point de vue : l’IA ne crée pas “à partir de rien”. Il y a toujours des choix humains, des prompts, des sélections, des ajustements. Mais il faut reconnaître que certaines productions IA atteignent déjà le niveau d’une grande partie de la musique humaine actuelle.
    Elles ne cassent pas encore les codes, ne produisent pas encore de grands chocs esthétiques. Mais elles fonctionnent, circulent et elles touchent.
    Les auditeurs, notamment les plus jeunes, se poseront sans doute de moins en moins la question de savoir “qui est derrière”.
    Le débat sur l’avenir de la musique ne fait que commencer.

    https://lnkd.in/e64g2ixK

  23. Génération musicale : quel problème résout-on ?

    Heureux d’annoncer la parution de mon article « Génération musicale et intelligence artificielle : quel est le problème ? » dans le dernier numéro des Cahiers de sémiotique des cultures, consacré à IA et culture / AI and Culture, chez Classiques…

    Read noteClose note

    Heureux d’annoncer la parution de mon article « Génération musicale et intelligence artificielle : quel est le problème ? » dans le dernier numéro des Cahiers de sémiotique des cultures, consacré à IA et culture / AI and Culture, chez Classiques Garnier, et dirigé par François Rastier.

    La question que j’y pose : lorsqu’une IA “génère” de la musique, quel problème est-elle censée résoudre ?

    Dans beaucoup d'applications du deep learning, le problème est bien posé : on dispose d’une entrée, d’une sortie attendue, d’une fonction d’erreur, et donc d’un objectif à optimiser. Mais en musique, et plus largement dans les artefacts culturels, cette évidence disparaît. Il n’existe pas de fonction de coût consensuelle pour définir une “bonne” musique. La popularité n’est pas un signal fiable : elle dépend de conditions sociales, de dynamiques de diffusion, d’effets d’âge, de contexte. Les données de streaming sont massives, mais profondément ambiguës : réduire les skips ne signifie pas nécessairement produire une musique plus intéressante.

    J’essaie de montrer que les systèmes actuels de génération musicale ne souffrent pas seulement de limites techniques. Le problème est plus profond : ils optimisent souvent des approximations de qualité; ressemblance stylistique, adéquation au prompt, continuité, absence d’événements perturbateurs, qui ne coïncident pas avec ce que nous appelons, faute de mieux, l’intérêt esthétique.

    Une autre piste me semble féconde : penser l’IA musicale non comme une machine à produire automatiquement des morceaux “finis”, mais comme un espace d’action. Un bon système créatif permet peut-être de faire "avec", et non pas de "faire mieux" : explorer, contrôler, modifier, comprendre, s’approprier.

    Lien vers le numéro :
    https://lnkd.in/eNk6Seg9

  24. Détecter l’IA, ou savoir ce que l’on veut valoriser ?

    Peut-on reconnaître un morceau produit par intelligence artificielle ? Aujourd’hui, parfois. Demain, probablement beaucoup moins. Les indices actuels, voix trop parfaite, production trop lisse, paroles génériques, absence d’aspérités, vont…

    Read noteClose note

    Peut-on reconnaître un morceau produit par intelligence artificielle ? Aujourd’hui, parfois. Demain, probablement beaucoup moins.
    Les indices actuels, voix trop parfaite, production trop lisse, paroles génériques, absence d’aspérités, vont s’effacer. Le vrai enjeu ne sera pas de “détecter” l’IA, mais de comprendre ce que nous voulons encore valoriser dans la musique : le son produit, l’intention, la trajectoire, la signature, la prise de risque ?

    L’IA peut générer des morceaux très convaincants. Mais une musique intéressante n’est pas seulement une musique qui ressemble à de la musique. C’est souvent une solution singulière à un problème singulier.

    https://lnkd.in/egPWCKYP

  25. The hidden biases of constrained generation

    LLMs are often asked to do things like rhyme, obey a fixed meter, fill in blanks, or satisfy positional constraints. They sometimes do surprisingly well. But there is a catch: these procedures are not exact conditioning of the underlying…

    Read noteClose note

    LLMs are often asked to do things like rhyme, obey a fixed meter, fill in blanks, or satisfy positional constraints. They sometimes do surprisingly well.
    But there is a catch: these procedures are not exact conditioning of the underlying autoregressive model. That creates a hidden inferential bias.

    We show that while local autoregressive sampling is easy, several exact constrained inference tasks are computationally hard:
    exact sentence-level MAP decoding is NP-hard.

    This remains true under unary and metrical constraints: exact conditioned normalization is #P-hard even for regular constraints such as fixed-length terminal events.

    This helps explain why language and music models often get close to rhyme, syllable count, infilling, or positional constraints, but not *exactly* right. Large subsets of actual solutions are probably never produced.

    Hidden Biases in Conditioning Autoregressive Models
    https://lnkd.in/edGHqAW7
    Comments very welcome.

  26. The mathematics behind synthetic populations

    Our paper, “Maximum Entropy Relaxation of Multi-Way Cardinality Constraints for Synthetic Population Generation” is out on arxiv. It tackles a hard problem at the heart of synthetic populations: generating realistic individuals from complex …

    Read noteClose note

    Our paper, “Maximum Entropy Relaxation of Multi-Way Cardinality Constraints for Synthetic Population Generation” is out on arxiv. It tackles a hard problem at the heart of synthetic populations: generating realistic individuals from complex statistics, including multi-way constraints that quickly defeat traditional exact methods. The paper shows that our MaxEnt approach scales much better as complexity rises, and on difficult 40-variable ternary benchmarks it reduces error by 50% to 80% versus generalized raking.

    For us, this is not just a paper. It is a strong validation of the science behind Imagine All The People and of our ambition to build the reference platform for synthetic populations and population-based design.

    The traction of the company is awesome, and this result gives us even more confidence in the road ahead.

    arXiv: https://lnkd.in/eb6GH5bb
    Great summary: https://lnkd.in/eEyEZw_H

  27. La fluidité n’est pas la vérité

    Les LLM sont des Google Maps du langage: ils trouvent toujours un chemin, même quand il ne mène nulle part. Ils produisent des continuités parfaites, des liens élégants, des raisonnements fluides. Mais fluidité n’est pas vérité! Dans ma dernière…

    Read noteClose note

    Les LLM sont des Google Maps du langage: ils trouvent toujours un chemin, même quand il ne mène nulle part.
    Ils produisent des continuités parfaites, des liens élégants, des raisonnements fluides. Mais fluidité n’est pas vérité!
    Dans ma dernière tribune aux Echos, j’explore cette illusion de cohérence, à travers le concept d’apophénie, bien connu des psychiatres, et ce qu’elle change dans notre manière de penser avec les machines.
    https://lnkd.in/eQ5wpgaB

  28. L’IA se porte-t-elle en pendentif ?

    L’IA se porte-t-elle en pendentif ?

    Nous avons en nous des mécanismes d'attachement qui échappent à notre raison. Dans RTL, on parle de “Friend”, cet ami virtuel intégré dans un collier, qui provoque la curiosité mais aussi des questions de vie privée, de dépendance émotionnelle et … d’attachement.
    👉 À lire ici : https://lnkd.in/e44AhPtT

  29. Au-delà des sondages : explorer des populations entières

    Les sondages, c'est fini. Continuer à gouverner avec des moyennes statistiques dans des sociétés fragmentées, c’est piloter un avion avec un seul cadran. Dans cette tribune des Échos, je défends l'idée que l’IA ne doit pas servir à mieux prédire…

    Read noteClose note

    Les sondages, c'est fini.
    Continuer à gouverner avec des moyennes statistiques dans des sociétés fragmentées, c’est piloter un avion avec un seul cadran.
    Dans cette tribune des Échos, je défends l'idée que l’IA ne doit pas servir à mieux prédire l’opinion, mais à explorer des populations entières, leurs contradictions, leurs minorités, leurs points de rupture.
    Moins de certitudes. Plus de scénarios.

    📄 Tribune : https://lnkd.in/eRsy7Z3Z

  30. Les outils de création musicale ont encore un avenir

    Une interview dans Billboard:
    https://lnkd.in/edzASSU6

    L’IA ne va pas faire disparaître les outils de création musicale.
    Les logiciels historiques restent indispensables pour structurer, contrôler et travailler la musique dans le temps. L’IA s’y intégrera peut être, ou pas. Peu importe.

  31. Pourquoi tant de haine contre l’IA ?

    Nouvelle tribune dans Les Échos, un peu à rebours de l’air du temps.

    Read noteClose note

    Nouvelle tribune dans Les Échos, un peu à rebours de l’air du temps.

    👉 « Pourquoi tant de haine contre l’IA ? »
    https://lnkd.in/eRaVam73

    Je suis frappé par la violence émotionnelle de certains discours sur l’IA.
    Pas seulement des critiques mais une forme de rejet quasi réflexe, parfois indifférent aux faits, aux usages réels, ou même à la recherche elle-même.

    Et si une partie de cette hostilité relevait moins d’un danger que d’un vertige ?
    Je ne nie ni les risques, ni les abus, ni les enjeux politiques.
    Mais la peur déguisée en certitude morale est un mauvais guide pour comprendre le présent, et anticiper l’avenir.

  32. Penser, ce n’est pas simplement évoquer un sujet

    Un trait de notre langue qui bousille littéralement le monde intellectuel français depuis des années est l'usage du mot "question" pour signifier "sujet". On entend "Cette histoire soulève la question du temps" (remplacer temps par sujet…

    Read noteClose note

    Un trait de notre langue qui bousille littéralement le monde intellectuel français depuis des années est l'usage du mot "question" pour signifier "sujet".
    On entend "Cette histoire soulève la question du temps" (remplacer temps par sujet, politique, climat, alimentation, déplacement, altérité, musique, petit-déjeuner, vacances, etc.).
    En évoquant une "question" qui n'en est pas une (pouvez-vous répéter la question ?) on se convainc (et aussi beaucoup d'interlocuteurs) qu'on est en train de dire quelque chose alors qu'on ne fait qu'effleurer un sujet avec des mots.
    Dire ce n'est pas évoquer. Penser, ce n'est pas prononcer des mots. C'est faire des propositions, qui seront forcément criticables.

  33. Des idées, sans l’excipient de l’intrigue ?

    Dans un roman, un film ou une série, il y a en général une “idée“ de départ (“les dossiers oubliés”, par exemple, self-contained, ou bien “Saving private Ryan”, sur le débarquement).

    Read noteClose note

    Dans un roman, un film ou une série, il y a en général une “idée“ de départ (“les dossiers oubliés”, par exemple, self-contained, ou bien “Saving private Ryan”, sur le débarquement).

    Mais il faut toujours rajouter à cette idée des personnages complexes, des intrigues amoureuses, des trucs en plus, pour que l’idée soit digestible.

    Comme si le spectateur, lecteur, n’était pas capable d’apprécier une idée directement, comme s'il fallait toujours l’enrober dans un excipient, pour garder son attention.

    Et si on essayait autre chose, une littérature sans histoire, des films sans intrigue, de la musique sans production, juste avec des idées intéressantes?

  34. Realization without a plan

    End-to-end generation systems are increasingly impressive. They generate content that becomes undistinguishable from human-produced ones, in music, images, videos, etc.

    Read noteClose note

    End-to-end generation systems are increasingly impressive. They generate content that becomes undistinguishable from human-produced ones, in music, images, videos, etc.

    But they do something crazy: realization without plan.

    Imagine building a house from scratch, with no plan, building a wall, then deciding to add a room, painting walls before knowing where are windows and roofs.

    Imagine writing a book by typesetting it along the way.

    Imagine producing a music title by adding reverb, compression, etc. before melody, harmony and all instruments are known to the composer.

    Imagine a world with realization and no plan.

    This is generative end-to-end AI.

  35. L’IA, une discipline pour polymathes

    Devant le déluge de postures sur l'IA venant de personnes manifestement incompétentes et en quête de notoriété facile, il faut rappeler que L’IA est avant tout, dans sa version "noble" du moins, une discipline pour polymathes. On dit parfois que…

    Read noteClose note

    Devant le déluge de postures sur l'IA venant de personnes manifestement incompétentes et en quête de notoriété facile, il faut rappeler que L’IA est avant tout, dans sa version "noble" du moins, une discipline pour polymathes.
    On dit parfois que l’IA n’est pas vraiment une science.
    C’est oublier qu’elle mobilise simultanément :
    – des mathématiques et de la statistique (algèbre liénaire, géométrie, Bayes)
    – de la physique (statistique notamment, belief propagation et max entropy),
    – de la théorie de l’information (l'entropie et ses variantes),
    – du signal (Fourier, convolution, filtrage)
    – de l’algorithmique et de la complexité ("sentir" qu'un algo est polynomial),
    – de la théorie des graphes (les problèmes NP et les P)
    – de la logique (les limites de l'inférence d'ordre 1)
    – de l’informatique au sens large (la pile et le tas).
    L’IA est, par nature, plus ouverte que chacune de ces disciplines : elle oblige à circuler entre elles, à comprendre leurs liens, à construire des ponts. Un bon chercheur en IA n’est pas un spécialiste étroit : c’est un polymathe, capable d’articuler plusieurs visions du monde pour résoudre un même problème.
    C’est cette transversalité féconde qui fait de l’IA une discipline à part entière.

  36. Pachet’s conjecture

    Pachet's conjecture:
    All predictions about the future of AI have been, are and will be wrong.
    Including this one.

    Elaborated variant:
    AI evolves faster than our frameworks for understanding it. Consequently, all forecasts about its future are invalid the moment they are uttered — including this forecast itself.

    A consequence:
    Anyone predicting where AI is headed is wrong. I say this with full awareness.

  37. Inventer des mots nous aide à penser

    Interview sur Zoom zoom zen sur musique et IA: ça commence par ce terme "slop", qui souligne l'inventivité lexicale des anglo-saxons.
    Hélas, nous sommes trop souvent suiveurs sur ce terrain. Pourtant, inventer des mots nous aide à penser!

    https://lnkd.in/e_4VPmdk

  38. Understanding takes time

    What does "understanding" means ?

    Deep question, and there are many levels of understanding. One thing is certain: the time it take to understand deeply something is not compressible.

    Understanding takes time, there is no free lunch for acquiring knowledge.

  39. Attention, surprise and conformism

    The idea that attention needs a subtle mix of surprise and conformism is not new. It was at the origin of automatic playlist generation. However, 26 years later, I am not sure we gained a lot of deep understanding on the matter.

  40. Flow Machines: what did the machine do, and what did the artist do?

    At SONY CSL Research Laboratory, we have created two entire pop songs composed with Artificial Intelligence, thanks to Flow Machines. How is it possible? That’s because our Flow Machines software learns music styles from a huge database of songs…

    Read noteClose note

    At SONY CSL Research Laboratory, we have created two entire pop songs composed with Artificial Intelligence, thanks to Flow Machines. How is it possible? That’s because our Flow Machines software learns music styles from a huge database of songs. Then, exploiting unique combinations of style transfer, optimization and interaction techniques, it can compose in any style.

    For example, listen to “Daddy’s Car”, a pop song in the style of The Beatles:

    Our "composer-in-residence" Benoît Carré arranged and produced the songs, and wrote the lyrics.

    Here’s another example: a composition in the style of American songwriters such as Irving Berlin, Duke Ellington, George Gershwin and Cole Porter. The title is “The Ballad of Mr Shadow”:

    How is it done? What did the machines do, and what did the artist?

    1) We set up a database called LSDB. It contains about 13000 leadsheet from a lot of different styles and composers (mainly jazz and pop about also a lot of Brazilian, Broadway and other music styles).

    2) The human composer (in this case Benoît Carré, but we are experimenting with other musicians as well) selected a style and generated a leadsheet (melody + harmony) with a system called FlowComposer. For Daddy’s Car, Carré selected as style “the Beatles” and for Mr. Shadow he selected a style that we call “American songwriters” (which contains songs by composers like Cole Porter, Gershwin, Duke Ellington, etc).

    3) With yet another system called Rechord the human musician matched some audio chunks from audio recordings of other songs to the generated leadsheets.

    4) Then the human musician finished the production and mixing.

    Daddy’s Car is an excerpt of an AI pop album to be released in 2017.