Import AI 474: Platonic mindspace; TPUs in space; Zhipu starts an outer RSI loop

Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe.

Subscribe now

Are minds patterns from a Platonic space, with bodies and machines as their interfaces, Michael Levin asks:
…A mind-bending paper asking us to reconsider basic assumptions about philosophy and biology…
Could the complexity in the world around us be better explained by there being a platonic zone of patterns of arbitrary complexity which seek to embed themselves in the world? That’s the big question at the heart of an iconoclastic paper from scientist Michael Levin.

The argument: “I 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… 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)”, he writes. “Bodies (whether living, engineered, or hybrid) are interfaces for a massive, multi-scale hierarchy of patterns to ingress into the physical world… specific patterns represented by biophysical, biomechanical, or chemical information fields serve as encoded setpoints for homeostatic or allostatic processes—in other words, goal states toward which systems try to navigate in various problem spaces”.

Towards a platonic realm of arbitrarily sophisticated patterns: “We already know that non-physical patterns ingress into, and functionally matter, in the non-living and living world and that we can (and do) study them to great effect… platonic forms inject information and influence into physical events, such as the growth and form of biological bodies,” he writes. “This latent space [of platonic forms] contains not only low-agency forms such as facts about integers and geometric shapes, but also a wide range of increasingly high-agency patterns, some of which we call ‘kinds of minds’. Thus, I propose that minds, as patterns that ensoul somatic embodiments, are of exactly the kind (but not in degree) of non-physical nature as the patterns that inhabit and guide the behavior of simple physical structures.”

Some bizarre experiments: There aren’t any home-run experiments that back this up, but rather there are weird things in the world that suggests that even seemingly simple systems are able to explore and expand into rich patterns that seem to be adjacent to them in this hypothesized platonic space of complexity, yet not implied by the systems themselves.
A couple of nice examples here are xenobots, which are biorobots made of frog cells and “teach us about patterns adjacent to those of frog embryos”, and anthrobots which are made from human tracheal cells and when taken out of the body take on bizarre new forms and display interesting behaviors like being able to autonomously heal damage to neurons and “which teach us about patterns adjacent to adult human tissues”.
There’s also an amazing discussion of an experiment where they take a standard sorting algorithm and perturb it in a few ways, including ‘locking’ certain cells in place and noting that it then routes around those cells, and a more complex experiment where they let each number run a different type of sorting algorithm and then study how different families of sorting algorithm appear and then cluster together in numberspace during the sort - this kind of complexity is unexplained and not anticipated by staring at these algorithms in the abstract, but by embedding them in a complex system they suddenly take on new properties.
“Machines (whether meaty or silicon-based) 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. It is those behaviors—allowed by the algorithm but not directly prescribed by it—that correspond to the freedom (physically non-determined) or secret sauce that we seek when trying to understand how free minds can supervene on chemically determined substrates in the case of living beings,” he writes. “On this view, algorithmic machines and biochemical life are on exactly the same spectrum, having in common the ability to go beyond the facts of physical or algorithmic implementation because both are just pointers/interfaces to patterns that ingress in a way that results in getting more out than we put in.”

Why this matters - perhaps the ‘minds’ of AI and of humans are close but not exactly the same in platonic mindspace: The research agenda Levin lays out is intoxicating and fascinating and also seems to me to hold some clues as to how we might grapple with the deep philosophical and alignment questions inspired by AI systems. Perhaps both human minds and different types of AI minds are neighboring forms in this space, and brains and datacenters are different bio- or anchor-systems in the physical world that their shadows fall on?
Could we be but vessels colonized by patterns from another place? Could it be “possible to re-cast the theory of evolution as a process in which agential patterns seek embodiments”?
“Is there a “force”, beyond the “if you build it, they will come” model of physical objects, pulling patterns from the space? Or are the contents of the Platonic space under “positive pressure”, somehow encouraging their appearance in the world as intrusive thoughts, archetypes, works of art?... The only thing that can be said strongly at this point is that our ignorance about the capabilities of matter, together with the patterns that ingress into specific architectures, is vast”.
Make yourself a bucket of coffee and spend a few hours with this paper - it’s worth it.
Read more: Ingressing Minds: Causal, Non-Physical Patterns In-Form Natural, Synthetic, and Hybrid Embodiments (MDPI).

***

Robotics is preparing for a LLM-moment, but it’s missing a universal algorithm:
…For robots to mature, post-training needs to simplify…
Perry Dong, a Stanford researcher, and Chelsea Finn, a Stanford professor and co-founder of robot company Physical Intelligence, have written a nice piece about what is holding back robotics.

LLMs vs Robots: For LLMs, what turned out to be important was that “the field converged on a shared recipe for post-training language models”, which basically had four steps: 1) have a strong pretrained model, 2) define the environments and reward (e.g., preference models), 3) run RL optimization against the reference model, and 4) watch for and address pathologies like reward hacking, they write.
“Robotics is sitting almost exactly where language modeling was: The pretraining has scaled beautifully,” they write. “What is missing is the other half: the model learning from its own experience, and for that, we need a recipe for post-training. And for robotics, it needs to be even more reliable than language models.”

What does robotics need? “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”, they write. Along with this, they need to converge on “a set of standard practices surrounding the algorithm. A default way to define what counts as success. A default way to reset the scene between attempts, so the robot can try again. A default way for a person to give feedback, and to turn that feedback into learning”.
In the post, the researchers talk about a specific algorithm, EXPO(-FT), that they’ve been developing for robots which they think embodies some of these qualities, though it’s early in its development and not widely used. “EXPO(-FT) 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.”

Why this matters - standard recipes are a prerequisite for a major scale-up: Proprietary large-scale LLMs are already smart enough to give instructions to robots and help them construct and carry out complex plans in the world, but the actual ability for robots to move and see remains fairly primitive; fixing that will require models customized around robot movement and vision and doing that will require the kind of standardization described here.
“Language model post-training became scalable because the field settled on defaults concrete enough to follow and be expected to work. Robotics is arriving at the same moment”, they write. “Converging on a set of industry defaults, a universal post-training recipe, is the most important part of bringing us to that point.”
Read more: Towards Universal Post-Training for Robotics (Perry Dong blog).

***

SPACE COMPUTERS! I REPEAT: SPACE COMPUTERS!
…Google prepares to put TPUs in space…
Google has given an update on Project Suncatcher, its initiative announced last year to put computers in space and eventually train AI systems there (Import AI #434). Google is preparing, along with its partner Planet, to send some of its chips to space as part of the SpaceX “Transporter-18 rideshare mission”.

Preparing for space: Google has done some stress tests of its TPUs to see how well they’ll do with the immense g-forces of going to space. Google has also tested how well they respond to radiation and found 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”. Currently, Google is working on cooling in space, which will likely be a challenge - computer chips generate a ton of heat and radiating that away in a vacuum is very, very difficult.

Why this matters - the future of AI training and AI inference is off the planet: Though Suncatcher sounds scifi-like it actually makes sense if you think about the ever-growing scale of compute used for AI and its energy needs. Do you know what has vast room and great solar power? Space. It seems very likely to me that humanity moves a very large amount of computation into orbit very quickly, especially as singularity-driven automation comes in across the AI supply chain.
Read more: Behind Project Suncatcher, our moonshot to put AI in space (Google blog).

***

Zhipu uses GLM-5.3 to automate some of its own infrastructure work:
…Chinese developers start to do the outer RSI loop…
Zhipu AI, the Chinese company behind GLM-5.3, one of the world’s strongest open-weight LLMs, has written a post about how it has been using its own models to help it build its own infrastructure, giving us a look inside how AI labs are trying to speed themselves up with the technology they build.
Specifically, this post details how they used their technology to help them launch GLM-5.3 Flash, a fast and cheap version of their most powerful model.

What they did: “The GLM-5.3-Flash launch established an optimization loop involving engineers, the Infra Agent, and the experimental environment. Engineers defined objectives and system boundaries. The agent handled analysis, hypotheses, and code changes. The experimental environment provided layered, timely, and verifiable feedback,” they write. “Much of the work was carried out by an Infra Agent powered by GLM-5.3… With the Infra Agent’s feedback loop running throughout the optimization process, GLM-5.3-Flash went from initial model adaptation to production readiness in less than two weeks, ultimately tripling end-to-end throughput relative to the initial baseline”

How to automate yourself: The company shared tips for making software that can use AI for greater automation.

• Feedback must be sufficiently local: “It should be tied to specific engine launch parameters, code changes, kernels, input conditions, threads, execution intervals, or code paths, helping the agent narrow the scope of the problem.”

• Feedback must be inexpensive and timely to obtain: “Shorter validation cycles help the agent correct course promptly and spend less effort on unproductive hypotheses.”

• Feedback must support objective verification: “Whether a change is correct and whether performance has improved should be determined by reference implementations, test results, and comparable experimental metrics.”