Jack Clark's Import AI 474, published September 28, 2026, gathers three threads worth reading together: a philosophical paper from scientist Michael Levin proposing that minds are non-physical patterns that inhabit bodies, a Stanford argument that robotics needs a universal post-training recipe, and an update on Google's plan to fly TPUs to orbit on a SpaceX rideshare.
Levin's paper, published in MDPI, argues that synthetic morphology and diverse intelligence research point toward a non-physicalist model of mind. He writes that he wants to "argue that the emerging sciences of synthetic morphology and diverse intelligence suggest non-physicalist models of mind and show how they can be empirically investigated." His core analogy is that the mind-brain relationship mirrors the math-physics relationship.
“I propose that the relationship between mind and brain is the same as the relationship between mathematical patterns and the morphogenetic outcomes they guide (more broadly, mind:body is as math:physics)”— Michael Levin, Scientist and author of the paper
Bodies, in Levin's framing, are interfaces. He describes them as "interfaces for a massive, multi-scale hierarchy of patterns to ingress into the physical world," with biophysical and chemical information fields acting as encoded setpoints toward which systems navigate. The Platonic latent space he posits contains not just facts about integers and shapes but also high-agency patterns he is willing to call "kinds of minds."
Key facts
- 01Michael Levin's MDPI paper argues minds are non-physical Platonic patterns that use biological or engineered bodies as interfaces into the physical world.
- 02Stanford's Perry Dong and Physical Intelligence co-founder Chelsea Finn say robotics lacks a universal post-training recipe analogous to the one that unlocked LLMs.
- 03Dong and Finn's proposed EXPO(-FT) algorithm fine-tunes frontier robotics models using RL edits from a lightweight policy absorbed back into the frontier model.
- 04Google's Trillium TPUs survived radiation testing equivalent to a five-year space mission and will fly on SpaceX's Transporter-18 rideshare with partner Planet.
The empirical hooks Levin cites are unusual: xenobots built from frog cells and anthrobots built from human tracheal cells, both of which take on behaviors and forms not implied by their source tissue. He also points to a perturbed sorting algorithm that routes around locked cells, arguing that machines "also do other things that are not in the algorithm, as do we, and these things are not just unpredictable complexity, it is intelligence and other components of minds."
The second thread is a Stanford blog post from Perry Dong and Chelsea Finn, the latter a co-founder of robot company Physical Intelligence. Their claim is that large language models took off because "the field converged on a shared recipe for post-training language models" built around a strong pretrained model, defined environments and rewards, RL optimization against a reference model, and mitigations for pathologies like reward hacking.
Robotics has the pretraining but not the post-training recipe, they argue. What is needed is "an algorithm built specifically for fine-tuning frontier robotics models, one that stays stable when applied to models with billions of parameters, and that learns from a small enough amount of experience to be practical on real hardware," alongside standard practices for defining success, resetting scenes, and turning human feedback into learning.
Their candidate is EXPO(-FT), which "works by learning to repeatedly improve actions from the frontier model using reinforcement learning with small edits from a lightweight policy, and then absorbing that into the frontier model itself." Dong and Finn concede it is early and not widely used, but frame the broader point as a call for industry defaults: "Converging on a set of industry defaults, a universal post-training recipe, is the most important part of bringing us to that point."
The third thread is hardware in orbit. Google is preparing to send chips to space with partner Planet as part of SpaceX's Transporter-18 rideshare mission, an extension of Project Suncatcher, the initiative announced last year to eventually train AI systems in space. Stress tests on g-forces and radiation are complete; Google says its Trillium TPUs "hold up remarkably well, and can survive a radiation total ionizing dose greater than what they would receive during a five-year space mission."
The open problem Google flags is cooling. Radiating heat away in a vacuum is difficult, and the newsletter notes the company is still working on it. Between Levin's Platonic mindspace, a missing robotics recipe, and TPUs bound for orbit, Import AI 474 sketches a field pushing simultaneously on philosophy, method, and physical substrate.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
