@FigaroVox@Aurelie_JEAN@Aurelie_JEAN, il ne faut pas trop écouter Dario Amodei qui est, parmi les dirigeants des entreprises d'IA, celui qui laisse le plus de place à l'éventualité d'une conscience des machines
If you ever feel like you're late to the game, consider that in the 1890s many scientists thought physics as a field was completely solved (quote below is from Albert Michelson in 1894).
On the front of intelligence science, it feels more like the 1870s. For the first time we have something that is starting to really work (however primitive it may be), which we can use as a springboard for the next few decades of discoveries.
@babgi Ce sont aussi des travaux qui amènent une perspective très intéressante sur l’apparition des émotions et de la conscience à partir de l’évolution du vivant et pas à partir du traitement de l’information. C’est si évident quand on lit Antonio Damasio.
Stéphane Mallat nous fait avancer dans la compréhension théorique des modèles de réseaux de neurones, avec rigueur et modestie, et une vision passionnante et intelligente de l'IA qui se démarque de ce dont on se contente souvent dans la tech
En alliant abstraction théorique et retombées concrètes, Stéphane Mallat, lauréat 2025 de la médaille d’or du CNRS, a marqué de son empreinte les mathématiques appliquées à l’informatique. Du format de compression d’images JPEG 2000 aux fondements ... https://t.co/8Q67v9lNis
This corresponds to the limits of combining symbolic and connectionist approaches, where it's impossible to create an ontology of the whole world and it's impossible to automatically generate all the explicit concepts we need for reasoning.
World models are clearly missing in current LLMs as @ylecun says but it’s unclear how future model architectures will balance useful emergent representations of the world with explicit concepts that are understandable to humans.
https://t.co/9DUTl9fSrL
Interesting that Adam Brown anticipates LLMs will have Einstein-level intelligence in 10 years, but explains that at the moment physicists (like most of us) mainly use LLMs to search for and explain information and theories that are well known in the literature
New episode w Adam Brown, a lead of Blueshift at DeepMind & theoretical physicist at Stanford.
Stupefying, terrifying, & absolutely fascinating.
On destroying the light cone with vacuum decay, mining black holes, holographic principle,
& path to LLMs which make Einsteinian conceptual breakthroughs.
Enjoy! Links below.
0:00:00 - Changing the laws of physics
0:26:53 - Why our universe is the way it is
0:38:18 - Making Einstein level AGI
1:01:19 - Physics stagnation & particle colliders
1:11:58 - Hitchhiking
1:29:48 - Nagasaki
1:37:07 - Adam’s career
1:44:13 - Mining black holes
2:00:30 - Holographic principle
2:24:13 - Infinities
2:32:30 - Engineering constraints for future civilizations
Excellent thread by @fchollet about the performances of the new model by OpenAI on ARC-AGI. Evaluation dataset creation is an interesting challenge and is also a never-ending task: my guess is that we will always be able to design problems easy for humans and hard for machines.
Today OpenAI announced o3, its next-gen reasoning model. We've worked with OpenAI to test it on ARC-AGI, and we believe it represents a significant breakthrough in getting AI to adapt to novel tasks.
It scores 75.7% on the semi-private eval in low-compute mode (for $20 per task in compute ) and 87.5% in high-compute mode (thousands of $ per task). It's very expensive, but it's not just brute -- these capabilities are new territory and they demand serious scientific attention.
Tester l'intégration d'IA générative par Qwant dans ses résultats de recherche - résumés de réponses, réponses détaillées, toujours en citant les sources d'information
✨ L’IA DE QWANT EST EN OPEN WEEK ✨
Notre IA est disponible à toutes et tous pendant une semaine ! Plus besoin d’avoir un compte (même si c’est gratuit 👀) pour l’utiliser.
Elle répond à toutes vos questions et requêtes en un clin d’oeil.
Plus d’excuses pour tester ;)
"Information is a matter of questions and answers, it's not an objective thing, information doesn't just sit there." ... "it entails a relationship between a subject and an object", brilliant remarks by @Mark_Solms https://t.co/1NHYDYYUBl
@rheimann Good analysis, but you ignore an argument I’ve sometimes heard, that our mind also does something like predicting the next word in a sentence, and so we could expect LLMs to have some intelligence similar to ours. The toughest arguments are often the most reductionist.
🚀 Nouvel épisode de "Monde Numérique" ! 🎙️ L'IA va-t-elle remplacer les développeurs ? Pas demain la veille, selon Laurent Ach @ach3d 🤖💡 #MondeNumérique#IA#Tech https://t.co/hV6BedxKgJ
I had the pleasure participate in an interesting talk with @jonsvt and @brucel, about developing technologies like web search engines and web browsers in Europe
🎙️ On a réalisé un petit podcast avec nos amis de chez @vivaldibrowser !
Notre CTO, @ach3d et Jon von Tetzchner, CEO chez Vivaldi, discutent de la manière dont il est possible de concevoir des technos respectueuses de la vie privée en ligne 🔐
https://t.co/xfmuTFICyF
Interesting thoughts on the Turing test by @MelMitchell1
https://t.co/HZcPDmtJot
It was once believed that beating a human at chess required general intelligence. The story goes on with AI mastering one task after another, without any intelligence
"Attempts to erase and devalue the most humane parts of our existence are nothing new; AI is just a new excuse to do it.” Great thoughts by @ShannonVallor on the usual comment "You don’t think that your brain is a machine?" this time by Yoshua Bengio https://t.co/1H96eIL1mW
@ylecun information and computation are related to models of the world that can only exist at some particular scales in particular scientific and philosophical contexts. It’s naive to think that we can come up with a model that would explain everything.
The question of whether LLMs can reason is, in many ways, the wrong question. The more interesting question is whether they are limited to memorization / interpolative retrieval, or whether they can adapt to novelty beyond what they know. (They can't, at least until you start doing active inference, or using them in a search loop, etc.)
There are two distinct things you can call "reasoning", and no benchmark aside from ARC-AGI makes any attempt to distinguish between the two.
First, there is memorizing & retrieving program templates to tackle known tasks, such as "solve ax+b=c" -- you probably memorized the "algorithm" for finding x when you were in school. LLMs *can* do this! In fact, this is *most* of what they do. However, they are notoriously bad at it, because their memorized programs are vector functions fitted to training data, that generalize via interpolation. This is a very suboptimal approach for representing any kind of discrete symbolic program. This is why LLMs on their own still struggle with digit addition, for instance -- they need to be trained on millions of examples of digit addition, but they only achieve ~70% accuracy on new numbers.
This way of doing "reasoning" is not fundamentally different from purely memorizing the answers to a set of questions (e.g. 3x+5=2, 2x+3=6, etc.) -- it's just a higher order version of the same. It's still memorization and retrieval -- applied to templates rather than pointwise answers.
The other way you can define reasoning is as the ability to *synthesize* new programs (from existing parts) in order to solve tasks you've never seen before. Like, solving ax+b=c without having ever learned to do it, while only knowing about addition, subtraction, multiplication and division. That's how you can adapt to novelty. LLMs *cannot* do this, at least not on their own. They can however be incorporated into a program search process capable of this kind of reasoning.
This second definition is by far the more valuable form of reasoning. This is the difference between the smart kids in the back of the class that aren't paying attention but ace tests by improvisation, and the studious kids that spend their time doing homework and get medium-good grades, but are actually complete idiots that can't deviate one bit from what they've memorized. Which one would you hire?
LLMs cannot do this because they are very much limited to retrieval of memorized programs. They're static program stores. However, can display some amount of adaptability, because not only are the stored programs capable of generalization via interpolation, the *program store itself* is interpolative: you can interpolate between programs, or otherwise "move around" in continuous program space. But this only yields local generalization, not any real ability to make sense of new situations.
This is why LLMs need to be trained on enormous amounts of data: the only way to make them somewhat useful is to expose them to a *dense sampling* of absolutely everything there is to know and everything there is to do. Humans don't work like this -- even the really dumb ones are still vastly more intelligent than LLMs, despite having far less knowledge.