Recursion Is The Next Scaling Law In AI

Predictions and Expectations:

• Francois Chauvard expects that combining the benefits of recursive models with giant LLMs will lead to "really huge" performance gains.
• Chauvard believes that if small recursive models are applied to the embedding spaces found by large models, it is "really gonna work."
• The host predicts that recursion is "not going away" in the field of AI research.
• Chauvard is "not sure" why the recursive mechanisms work, stating "we actually don't know why it really working."
• Chauvard believes that adding more recursion is "efficient and not necessary" for performance, citing Melanie Mitchell's work.
• The host expects that making small recursive models more general-purpose agents could allow them to achieve complex reasoning.

Timelines and Milestones:

• Chauvard recalls that RNNs were "very much in the belief that this was required to get to AGI" as of "about ten years ago" (referencing the era of Alex Graves' 2016 NeurIPS keynote).
• The transcript notes that the two papers demonstrating recursion's power were published "in 2025."
• Chauvard mentions that a model trained on 16 test-time steps can achieve performance comparable to 100% if trained on 16 steps but tested on just one ("seven eighths of the performance or almost all").

Technology and Product Direction:

• The host expects future research to involve "slamming together" recursive models with giant LLMs to utilize the "benefit of both."
• Chauvard predicts that the "outer refinement loop" is the primary reason for the success of Hierarchical Reasoning Models (HRM).
• The speakers discuss a direction where models learn recursive methods to store information in local memory without "teacher forcing via a chain of thought."
• The host suggests a future architecture where reasoning occurs in a latent space rather than the discrete token space used by current LLMs.
• Chauvard notes that current LLMs are limited by the need for "external memory tape" to achieve faster than $n \log n$ sorting performance.
• The speakers anticipate that future models might operate at different frequencies or hierarchies similar to the human brain, though Chauvard doubts strict bioplausibility is necessary.
• Chauvard expects that backpropagating through just one latent recursion step (truncated at $t=1$) is "completely sufficient" for training.

Market and Industry Outlook:

• The host states that researchers and labs are "very hungrily buying" data and traces, implying a market demand for the specific types of data needed to teach reasoning.
• Chauvard observes that "GPT-3 if you remember back way back when... got zero" on ARC Prize compared to the new recursive models.
• The speakers note that the current trend is moving away from just making models "bigger and bigger" to improve reasoning.
• Chauvard mentions that the HRM paper authors didn't do extensive scaling ablations, but someone at his company did, indicating industry interest in scaling these specific techniques.

Company Plans:

• Francois Chauvard reveals that researchers at his company, India, performed scaling ablations on the HRM paper and posted the breakdown on YouTube.
• The speakers discuss the potential for these recursive techniques to be integrated into existing giant models like "Gemini" or "Google's recursion language models."

Financial Guidance:

• (No financial guidance or specific monetary projections were found in the transcript.)

Risks and Caveats:

• Chauvard warns that current recursive models (HRM and TRM) are "task specific" and cannot inherently solve different tasks without retraining, unlike general-purpose LLMs.
• The speakers note that backpropagation through time in traditional RNNs fails at scale due to "vanishing or exploding gradient problems" and the need to retain activations for every single step.
• Chauvard states that Alexia's work on DEQ (Deep Equilibrium Models) is "not sufficient support for why [recursive models] is working."
• The host highlights a risk that relying on Chain of Thought or tool use bounds reasoning to "human knowledge," meaning the models would fail on problems outside the set of human knowledge.
• Chauvard points out that while MLPs outperformed attention mechanisms on Sudoku, they scored "zero" on maze tasks, indicating instability in architecture choices.
• The speakers acknowledge that backpropagating through the entire recursion loop hits a "memory perspective" limit, preventing the models from being made arbitrarily large.

Confidence and Disagreement:

• Chauvard is "very much in the belief" that recursive mechanisms (RNNs) are required to reach AGI, though he admits current implementations are limited.
• The host expresses skepticism about "bioplausibility arguments," stating he is "usually not super keen on these" and believes machine learning often moves away from them toward GPU-efficient solutions.
• Chauvard agrees that while bioplausibility can inspire ideas, systems often become "highly bio implausible" but work better on GPUs.
• The speakers are "not sure" about the theoretical lower bounds of performance for sorting in one-shot LLMs versus recursive models.
• Chauvard is "not sure" if the specific nomenclature used in the papers (X vs Z) matters, but emphasizes the functional clarity of distinct hidden states ($Z_H$ and $Z_L$).