January 2026

Imitation learning (IL) provides a data-driven framework for approximating policies for large-scale combinatorial optimisation problems formulated as sequential decision problems (SDPs), where exact solution methods are computationally intractable. A central but underexplored aspect of IL in this context is the role of the emph{expert} that generates training demonstrations. Existing studies employ a wide range of expert constructions, yet lack a unifying framework to characterise their modelling assumptions, computational properties, and impact on learning performance. This paper introduces a […]

Ver mais

LLMs compress large amounts of meaning/context/latent assumptions into finite internal representations. When the semantic load is close to those limits, small surface changes can push the model into a different internal pathway even though the meaning hasn’t changed. The output stays fluent but coherence across prompts breaks This is compression-aware intelligence and its a way of explicitly reasoning about what happens when meaning exceeds representational capacity. Helps explain why LLMs contradict themselves on semantically equivalent prompts submitted by […]

Ver mais

🧪 I was able to finally solve DeepMind’s Alchemy Meta RL benchmark using a new theoretical framework: Epiplexity For many years, I’ve been working on DeepMind’s Alchemy meta-reinforcement learning benchmark as a side project – a notoriously difficult task that requires agents to discover hidden “chemistry rules” that get shuffled each episode. The breakthrough: Instead of only selecting models by reward, I select by epiplexity – a measure of structural information extraction from the recent paper “From Entropy […]

Ver mais

Federated learning (FL) has emerged as a transformative distributed learning paradigm, enabling multiple clients to collaboratively train a global model under the coordination of a central server without sharing their raw training data. While FL offers notable advantages, it faces critical challenges in ensuring fairness across diverse demographic groups. To address these fairness concerns, various fairness-aware debiasing methods have been proposed. However, many of these approaches either require modifications to clients’ training protocols or lack flexibility in their […]

Ver mais

Last year, I closed a seven-year chapter of my life with my startup (an AI writing assistant): $1.35M in funding, 300,000 users, and an acqui-hire exit to a large company. But behind the scenes of that deal, I was weathering a two-year “perfect storm”: plummeting revenue, the loss of a second business, and a painful divorce. In this article, I’ll share how I managed to navigate this chaos using nothing but a single Google Doc instead of complex […]

Ver mais