MONTREAL ♡ AI YEARS IN THE MAKING. BUILT FOR THE SINGULARITY ERA. THE MISSION: BUILD MONTREAL’S LARGEST AI BUSINESS. UNICORN STATUS IS THE FIRST TARGET. montreal.ai MONTREAL.AI → $1B+ → ∞ #MontrealAI

THE NEXT STEP IN RECURSIVE SELF-IMPROVEMENT MAY NOT BE “AI REWRITING ITSELF.” It may be AI learning how to search better for its own next improvement. A remarkable new paper from researchers at Google, Google DeepMind, the University of Maryland and the University of Virginia: DREAM-RSI Recursive Self-Improvement through Evolving Worlds Tong Zheng, Xidong Wu, Zheng Zhang, Zhankui He, Chaoyi Zhang, Benjamin Coleman, Ruoqiao Wei, Di Bai, Haolin Liu, Rui Liu, Xue Wang, Yue Zhuan, Wang-Cheng Kang, Renkai Xiang, Heng Huang, Xinwu Cheng & Yunsong Guo. The core idea is exceptionally powerful: HISTORY BECOMES A WORLD. Long-horizon discovery is expensive. An agent explores branches. Tests hypotheses. Refines candidates. Hits dead ends. Allocates parallel workers. Eventually discovers something useful. Normally, much of that search history becomes context, logs—or discarded compute. Dream-RSI does something deeper. It turns the accumulated discovery tree into a REPLAY SIMULATOR. Alternative exploration policies can then “dream” inside that recorded world: Which branch should have been explored first? How deep should the search have gone? When should another direction have been opened? What should have run in parallel? When should exploration have stopped? Because the outcomes are already stored, these alternative strategies can be evaluated without repeatedly paying for the underlying coding-agent executions. Then the best exploration policy goes back online. It discovers more. That new discovery expands the replay world. The larger world trains a better explorer. The better explorer discovers a larger world. EXPLORE → RECORD → REPLAY → IMPROVE THE EXPLORATION POLICY → REDEPLOY → EXPAND THE WORLD → REPEAT That is the recursive loop. And an important scientific detail: The underlying coding agent, evaluator and execution interfaces remain fixed. What changes recursively is the EXPLORATION POLICY—the mechanism deciding where to search, how broadly, how deeply, how much to parallelize and when to stop. This is not vague “self-improvement.” It is an explicit meta-optimization target. The system is not merely searching for better solutions. It is improving HOW FUTURE SOLUTIONS ARE SEARCHED FOR. The reported results are substantial. In Lasso solver discovery, Dream-RSI with Gemini-3.1 Pro uses 317 discovery-agent calls versus 550 for the controlled fixed-exploration baseline while producing a better average downstream runtime. In GPU-kernel discovery: VGG16 → comparable performance with 2.43× fewer generations. LayerNorm → comparable performance with 1.79× fewer generations. ConvDiv → 2.09× higher performance at comparable budget. ConvMax → 1.44× higher performance at comparable budget. But one of the most important results may be less obvious: Simply compressing previous experience into high-level semantic “guidance” can perform WORSE. Why? Because a summary is not the search process. The structured history preserves branches, failures, costs, ordering, recovery opportunities and unrealized alternatives. In other words: DO NOT JUST REMEMBER THE LESSON. PRESERVE ENOUGH OF THE WORLD TO REHEARSE DIFFERENT DECISIONS. That is a profound idea for self-improving AI. Failure stops being merely wasted compute. A dead end becomes information about the topology of the search space. A failed implementation can remain distinguishable from a failed idea. A previous discovery run becomes reusable infrastructure for improving the next discovery run. The map becomes a simulator. The simulator improves the navigator. The navigator discovers a larger map. There is a striking connection here to another problem we have been working on with SUCCESSOR Ω. This connection is my synthesis—not a claim or endorsement by the Dream-RSI authors. Dream-RSI asks: How can an exploration process recursively improve by turning accumulated experience into better future search? SUCCESSOR Ω asks a complementary institutional question: Once intelligence keeps learning, changing and generating successors, how do we preserve what was learned WITHOUT allowing the learning process to certify itself or silently inherit real-world authority? Our answer is a governed succession architecture. SUCCESSOR Ω separates: LEARNING from SERVING. MANUFACTURING from PROOF. CAPABILITY from AUTHORITY. MEMORY from inherited permission. A frozen serving release can continue operating inside its proven envelope while a separate shadow successor learns. The descendant can inherit: evidence, provenance, skills, failures, program lineage, and the Live Chronicle. But it inherits neither proof nor authority. The successor must freeze. Then fresh evidence challenges it outside the claimant’s control. Only independently demonstrated capability can be considered for bounded authority. The conceptual progression becomes: Dream-RSI: Experience → Replay World → Better Exploration → New Experience SUCCESSOR Ω: Experience → Candidate → Freeze → Independent Proof → Bounded Admission → Chronicle → Successor These are different systems addressing different layers. But together they point toward an important frontier: RECURSIVE IMPROVEMENT NEEDS BOTH A MEMORY OF SEARCH AND A CONSTITUTION FOR SUCCESSION. One helps intelligence improve how it discovers. The other asks what must happen before those improvements become trusted institutional capability. That distinction may become increasingly important as self-improving agents move from code benchmarks toward science, infrastructure, industrial systems and consequential decision-making. The frontier is shifting: From models that answer. To agents that search. To systems that improve their search. To institutions that preserve what survives reality. Full credit to Tong Zheng and the Dream-RSI team for an elegant, concrete contribution to recursive improvement at the META-EXPLORATION layer. Dream-RSI paper: arxiv.org/abs/2609.14858 Dream-RSI: dream-rsi.com Code: github.com/zhengkid/Dream-RS… For our separate work on governed recursive succession: NEURAL-SYMBOLIC SUCCESSOR Ω Customer-Owned Mission Intelligence Overview: nitter.cf/Montreal_AI/status/209… White paper: montrealai.github.io/neurals… Vincent Boucher President, MONTREAL.AI & QUEBEC.AI THE NEXT GREAT DISCOVERY MAY BE A BETTER WAY TO DISCOVER. THE NEXT GREAT INSTITUTION MAY BE THE ONE THAT REMEMBERS WHAT WAS DISCOVERED—WHILE REQUIRING EVERY NEW GENERATION TO EARN ITS POWER AGAIN. #RecursiveSelfImprovement #RSI #Successor #AGI #AIAgents #ScientificDiscovery #NeuralSymbolicAI #MontrealAI

MONTRÉAL → THE NEXT FRONTIER. ⚜️ Before the factory, the experiment. Before the commitment, the evidence. SUCCESSOR Ω — ANABASIS Ω Discover what’s possible. Prove what’s worth pursuing. Build what comes next. 🎬 The Expansion Institution. Synthetic research. solana:tWKHzXd5PRmxTF5cMfJkm2Ua3TcjwNNoSRUqx6Apump #MontrealAI #SUCCESSOR

MONTREAL ♡ AI YEARS IN THE MAKING. BUILT FOR THE SINGULARITY ERA. THE MISSION: BUILD MONTREAL’S LARGEST AI BUSINESS. UNICORN STATUS IS THE FIRST TARGET. montreal.ai MONTREAL.AI → $1B+ → ∞ #MontrealAI

Un modèle se remplace. Les acquis de votre organisation ne devraient pas disparaître avec lui. SUCCESSOR Ω — THE BOOK L’intelligence de mission qui vous appartient Vincent Boucher · MONTREAL.AI · QUEBEC.AI La proposition : Louer des modèles puissants. Conserver l’institution qui les rend utiles. Pas les poids d’un fournisseur : la maîtrise de votre mission, de vos preuves, méthodes exécutables, tests, décisions et historique — selon des droits et des limites explicites. Chapitre 3 Et ce livre se met à l’épreuve. Dans son exercice sur des factures synthétiques, une modification fait passer les décisions correctes de 128 à 138 sur 160… mais laisse échapper 22 exceptions critiques. Le score progresse. La mission recule. Corrigez la distinction perdue. Confrontez la nouvelle version à une alternative plus forte. Regardez l’avantage s’effacer lorsque celle-ci rattrape le candidat. L’actif durable, c’est de savoir exactement pourquoi. Chapitre 2 Une seconde expérience éprouve le processus d’amélioration : mémoriser les échecs évite des évaluations inutiles. Mais un comparateur ordinaire, doté du même ordre de tests, obtient le même résultat. Conserver le savoir utile. Affiner la portée des affirmations. Chapitre 13 En 18 chapitres, l’ouvrage relie modèles du monde neuro-symboliques, Mission Gyms, économie comparative, vérification indépendante, autorité délimitée et mémoire institutionnelle. Non plus seulement : « Cette IA est-elle performante ? » Mais : « Qu’a-t-elle démontré ? Que peut-on l’autoriser à faire ? Que transmettra-t-elle à la génération suivante ? » Pourquoi maintenant ? WikiSkill, de Tang et ses coauteurs, explore l’accumulation de connaissances pour faire évoluer les compétences des agents. La synthèse de Duan et ses coauteurs examine l’amélioration des mécanismes qui produisent les progrès suivants. Ma lecture : plus l’IA apprend à se transformer, plus il importe de préserver ce que chaque transformation établit réellement. L’édition réunit PDF, EPUB et lecture web, un Mission Lab autonome, douze fiches réutilisables, un exemple rempli, les sources modifiables et les résultats expérimentaux. Aucun compte, clé API ou modèle à télécharger pour le laboratoire pédagogique. Guide de lecture Commencez par une décision récurrente, un résultat mesurable et une alternative crédible. Exécutez l’exercice. Analysez un échec. Exportez puis restaurez votre travail. Repartez avec un dossier de mission — pas seulement une conviction sur l’IA. Édition complète 4.0. Ensemble pédagogique ; logiciel commercial non inclus. Expériences synthétiques sous le contrôle de l’auteur : ni preuve client indépendante ni autorisation de production. Manuscrit anglais, avec un compagnon pratique français distinct — non une traduction intégrale. Note de publication Posséder l’institution. Remplacer les modèles. Préserver la mission. solana:tWKHzXd5PRmxTF5cMfJkm2Ua3TcjwNNoSRUqx6Apump #Successor #IANeuroSymbolique #IAInstitutionnelle #RechercheIA