Machine's Learning

Can a model trained only on birdsong classify elephant calls — without any fine-tuning at all? A new paper from Geldenhuys and Niesler runs frozen-embedding transfer from bird-trained and speech-trained foundation models to African and Asian elephant calls, and gets within 2.2 percent of an end-to-end supervised baseline. Even more striking: the second layer of the network outperforms the final la...

When you ask a modern language model to "think step by step," it writes out intermediate reasoning before answering and tends to do better on hard problems. The field has been treating those written steps as the reasoning itself. A new position paper from Wenshuo Wang argues that the evidence currently favors a different picture: the real reasoning happens in the hidden internal states moving thro...

The dominant recipe for building capable language models is depth — more layers, more parameters, more distinct transformer blocks. A new mechanistic interpretability paper from Nam, Gromov, Yaida and colleagues looks at what happens when you take a much smaller model and run the same block over and over again in a loop. Inside, the internal state moves through cyclic trajectories that settle into...

A puzzling result from a recent paper on multi-agent AI: more capable, reasoning-enabled language models cooperate LESS in social dilemmas than older, weaker ones. CoopEval takes the puzzle seriously and tests four classic mechanisms for restoring cooperation — repeated play, reputation, mediation, and binding contracts — across six modern LLMs. Without any mechanism, welfare collapses to 7% of op...

When AI agents play poker against each other, do they start modeling each other's minds the way humans do? A recent paper ran three Claude agents through a hundred hands of Texas Hold'em with a clean factorial design — memory present or absent, poker skill present or absent — and found a perfectly categorical result: agents with memory climbed a five-level ladder of theory-of-mind sophistication....

AI isn't just chatbots and agents. There are microphones in forests right now using machine learning to count chimpanzees, elephants, and rare birds — and that count is increasingly the basis for real money decisions about habitat conservation. Today we look at DeepForestSound, a region-specific acoustic detector for African tropical forests, and what it means when the detector layer becomes the m...

Last episode we talked about keeping AI agents from being attacked. Today we look at the failure mode that emerges when no one is attacking the agent at all — when the agent's own memory drifts over time through accumulated biased input. Memory misevolution as a path-dependent phenomenon, with institutional drift and the Overton window as the cross-domain parallel.

AI agents that can use tools — browse the web, read files, call APIs — face a serious vulnerability called indirect prompt injection. Today we look at ClawGuard, a runtime security framework that takes a different approach than training models to refuse: it sits between the AI and its tools, checking each action against rules the user has pre-approved. The shift from behavioral to architectural se...