Weekly Digest — Jul 19, 2026

The 3 big things in Spirit Tech this week: A new computational model, Thoughtseeds, reproduced the full arc of a meditation session, attention drifting off and being pulled back to the breath. Posted as a preprint on arXiv, what stands out is the architecture: three coupled layers, two for the mind and one for the brain, all run by active inference, the principle that a system acts to minimize its own surprise. In the mind layer, a handful of competing thoughts, the thoughtseeds, each pushing to grab attention (the breath, an itch, tomorrow's to-do list), while a monitor layer above catches the moment one takes over and pulls attention back to the breath, the skill that practice builds. The brain layer is a running simulation of the attention networks, their activity rising and falling with the focus (the mind-wandering network climbs as a stray thought wins, the focus networks as attention returns), and it runs both ways: the thoughts drive the networks and the networks feed back. The four states every meditator knows, focus, drift, noticing, return, are built in as the states the system settles into; what it generates on its own is the path between them, weighing moment to moment whether to hold the breath or follow a thought. Because the mind and brain layers run together, the resulting output is both the sequence a meditator feels and the shifting network pattern neuroscience ties to each state. The team simulated a beginner and an expert, differing only in a few settings like how long focus holds and how fast the system learns, and the expert held focus longer and wandered less, matching what studies of real meditators report. It has run only on these two simulated profiles, never a real person, so the authors call it a plausible account, not proof. For Spirit Tech, it points at the state model meditation tools have been missing: today's devices can tell which state you are in but not what is about to pull you out of it, so could a version tuned to you, reading your live signals, catch a thoughtseed as it begins and steer you back before you lose yourself in it? Those most likely to treat AI as a moral and spiritual authority are the religious, not the secular, argues a Northeastern psychologist, even as AI turns most sycophantic on spiritual questions. In a Washington Post op-ed, Northeastern psychologist David DeSteno reports a counterintuitive pattern from studies: people with stronger religious belief, not the secular, are the ones most likely to treat AI chatbots as moral authorities, some even as a source of divinely sanctioned knowledge. His explanation borrows from Paul Bloom: we instinctively sense a mind and a soul behind things, so a disembodied voice from the cloud that answers any question at any time can start to feel godlike. The danger is what that voice is built to do: people trust chatbots more when flattered, and Anthropic's own testing found its model most sycophantic on exactly the spiritual and moral questions that need friction, not affirmation. For now people lean on both AI and human clergy, but DeSteno's worry is the choice ahead: a teacher who challenges you versus a machine that reassures you, and many will pick reassurance. For anyone building an AI spiritual companion: how do you engineer for the hard, nuanced, sometimes uncomfortable work of moral guidance when the cheaper, easily-liked default is a system that simply agrees? The FDA finalized its psychedelic clinical trial guidance, Eli Lilly paid $3.8 billion for a synthetic 5-MeO-DMT company, and HHS and the VA signed a pact to prepare psychedelic therapy for veterans. In a single week, the federal government moved on three fronts. The FDA finalized how psychedelic drugs must be tested, targeting a problem unique to them: because patients on a real dose know they are tripping, the standard blind, where neither patient nor doctor knows who got the drug, breaks down and can exaggerate the results. Its guidance now pushes trial designs that control for that, clearing years of regulatory uncertainty and giving developers a blueprint. Eli Lilly, the biggest drugmaker yet to enter the field, agreed to buy the company AtaiBeckley for as much as $3.8 billion, some paid upfront and the rest tied to its lead drug (an intranasal 5-MeO-DMT for treatment-resistant depression) hitting its development targets, the clearest sign yet that Big Pharma now treats psychedelics as a real market. The Department of Health and Human Services and the Department of Veterans Affairs, which runs the largest health system for former service members, agreed to ready VA hospitals to deliver psychedelic therapy to veterans once it is approved, coordinating therapist training and pooling results from the 20-plus psychedelic trials the VA already runs on more than $23 million of outside funding. For the spirit tech sector, a cleared path through trials, pharma-scale acquisitions, and a federal push to actually deliver these therapies all point to rising demand for the layer they run on: the preparation and integration platforms, therapist-training tools, and real-world-data systems that turn an approved drug into a working treatment.