Two angles. What is happening in AI, and the neuroscience behind why it works. Subscribe for the daily digest.
Anthropic released Claude 4 with the ability to chain extended thinking and tool use in a single turn. This collapses what previously required multi-step orchestration into one API call.
ReadThe brain constantly generates predictions about incoming sensory data and only processes the difference between prediction and reality. This is the computational basis for attention, surprise, and efficient learning.
ReadA new DeepMind paper establishes scaling laws specifically for agentic systems, showing that tool diversity matters more than parameter count once models exceed 70B. Implications for how we architect multi-agent pipelines.
ReadDonald Hebb's 1949 principle remains the foundation of how biological neural networks strengthen connections through repeated co-activation. Modern deep learning optimizers are, at their core, approximations of this rule.
ReadDopamine neurons encode reward prediction errors, not rewards themselves. This distinction, discovered by Wolfram Schultz in the 1990s, directly inspired the temporal difference learning algorithms behind modern RL systems.
ReadApple opened its Foundation Models API to third-party developers, enabling on-device inference with zero cloud dependency. First major platform to offer free, private LLM access at the OS level.
ReadHumans hold roughly 4 chunks in working memory at once (Miller's 7 was revised downward by Cowan in 2001). AI context windows are functionally unlimited but suffer analogous degradation: attention dilutes with length.
ReadDuring sleep, the hippocampus replays sharp-wave ripples (SWR) of daily experiences at 10x speed, transferring them to long-term cortical storage. This biological mechanism directly informs how we design memory consolidation in Aldric.
ReadMirror neurons fire both when performing an action and observing it in others. Discovered by Rizzolatti in macaques, they offer a neural substrate for empathy, imitation learning, and theory of mind.
ReadEbbinghaus demonstrated in 1885 that distributed repetition produces stronger memory traces than massed practice. Spaced repetition systems exploit this effect, and it has direct parallels to how retrieval-augmented generation schedules re-indexing.
ReadFrontier research directly informs how Aldric is built.
Learn about Aldric