Glonce

Import AI 474: Levin's Platonic mind paper, Google TPUs in space, Zhipu's Infra Agent

researchThe Glonce editorsImport AI🔥63Glonce rating63Impact50Novelty55Relevance85Surprise60Players55Credibility75
Import AI 474: Levin's Platonic mind paper, Google TPUs in space, Zhipu's Infra Agent

Import AI issue 474 is a newsletter digest of four items. The first is a paper by the scientist Michael Levin, titled 'Ingressing Minds: Causal, Non-Physical Patterns In-Form Natural, Synthetic, and Hybrid Embodiments' (MDPI). Levin argues that the emerging sciences of synthetic morphology and diverse intelligence suggest non-physicalist models of mind. His central claim is that the relationship between mind and brain is the same as the relationship between mathematical patterns and the morphogenetic outcomes they guide; in his words, 'mind:body is as math:physics'. Bodies, whether living, engineered or hybrid, are interfaces through which a multi-scale hierarchy of patterns ingresses into the physical world. He describes a latent Platonic space that holds low-agency forms such as facts about integers and geometric shapes, and also higher-agency patterns, some of which we call 'kinds of minds'.

The newsletter says there are no home-run experiments backing this up, only odd phenomena. Xenobots are biorobots made of frog cells and are said to teach us about patterns adjacent to those of frog embryos. Anthrobots are made from human tracheal cells; outside the body they take on new forms and can autonomously heal damage to neurons. The paper also discusses a standard sorting algorithm that was perturbed in several ways. When certain cells were locked in place, the algorithm routed around them. In a second experiment each number ran a different type of sorting algorithm, and different families of algorithm clustered together in numberspace during the sort. Levin's reading is that machines, like living things, do things that are allowed by the algorithm but not prescribed by it, and that algorithmic machines and biochemical life sit on the same spectrum as interfaces to patterns. The newsletter author calls the research agenda fascinating, suggests it may offer clues for alignment questions, and wonders whether human and AI minds are neighbouring forms in this space.

The second item is a post by Perry Dong, a Stanford researcher, and Chelsea Finn, a Stanford professor and co-founder of the robot company Physical Intelligence. They note that language modelling converged on a shared post-training recipe with four steps: start from a strong pretrained model, define environments and reward (for example preference models), run RL optimisation against the reference model, and watch for pathologies such as reward hacking. They write that robotics 'is sitting almost exactly where language modeling was: The pretraining has scaled beautifully'. What is missing is the model learning from its own experience. They call for an algorithm built for fine-tuning frontier robotics models that stays stable on models with billions of parameters and learns from little enough experience to be practical on real hardware, plus standard defaults for defining success, resetting the scene between attempts, and giving human feedback. They describe EXPO(-FT), which they are developing: it learns to repeatedly improve actions from the frontier model using reinforcement learning with small edits from a lightweight policy, then absorbs that into the frontier model. The newsletter says it is early in development and not widely used.

The third item is an update on Google's Project Suncatcher, the initiative announced last year to put computers in space and eventually train AI systems there. Google is preparing, with its partner Planet, to send some of its chips to space on SpaceX's Transporter-18 rideshare mission. Google has stress-tested its TPUs against the g-forces of launch. It also tested radiation: 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'. Google is now working on cooling in space, which the newsletter says will likely be a challenge, since chips generate a lot of heat and radiating it away in a vacuum is very hard. The newsletter author predicts it is very likely that humanity moves a very large amount of computation into orbit very quickly.

The fourth item covers Zhipu AI, the Chinese company behind GLM-5.3, which the newsletter calls one of the world's strongest open-weight LLMs. Zhipu wrote about using its own models to build infrastructure, specifically to help launch GLM-5.3 Flash, a fast and cheap version of its most powerful model. Zhipu describes an optimisation loop of three parts: engineers defined objectives and system boundaries, an Infra Agent handled analysis, hypotheses and code changes, and the experimental environment gave layered, timely and verifiable feedback. 'Much of the work was carried out by an Infra Agent powered by GLM-5.3'. With that loop running throughout, 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. The newsletter headlines this as Chinese developers starting the 'outer RSI loop'. Zhipu also shared tips for building software that AI can automate; the first is that feedback must be sufficiently local, tied to specific engine launch parameters, code changes, kernels, input conditions, threads, execution intervals or code paths, so the agent can narrow the scope of the problem.

Key facts

  • Michael Levin's paper proposes that minds are patterns from a Platonic space that bodies and machines act as interfaces for, summed up as 'mind:body is as math:physics'; the newsletter says no home-run experiments back it yet.
  • Perry Dong and Chelsea Finn argue robotics needs a universal post-training recipe like the four-step one LLMs converged on, and offer their early-stage EXPO(-FT) algorithm as one candidate.
  • Google is preparing, with Planet, to send some chips to space on SpaceX's Transporter-18 rideshare mission; its Trillium TPUs can survive a radiation dose greater than a five-year space mission would deliver, and cooling is the open problem.
  • Zhipu AI says an Infra Agent powered by GLM-5.3 did much of the work to bring GLM-5.3-Flash from initial model adaptation to production readiness in less than two weeks, tripling end-to-end throughput versus the initial baseline.

Why it matters

The issue ties together several shifts at once. Levin's paper pushes a philosophical frame in which human and AI minds could be neighbouring patterns in one space, which the newsletter author thinks may bear on alignment questions. The robotics post argues that standard recipes are a prerequisite for a major scale-up, as they were for language models. Google's Suncatcher update moves orbital compute from an idea toward a launch plan. Zhipu's post is a rare look at how a lab uses its own model to speed up its own infrastructure work, which the newsletter frames as the 'outer RSI loop'.

Who it affects

Researchers in biology, philosophy of mind and AI alignment are the audience for Levin's paper. Robotics teams working on fine-tuning frontier models are the target of the Dong and Finn argument. Anyone planning AI compute and its energy supply may watch Suncatcher. Engineers building inference infrastructure, and teams considering agents for optimisation work, can look at Zhipu's account of the Infra Agent loop.

How to use it

Each item links to a primary source: the Levin paper (MDPI), the Dong blog post 'Towards Universal Post-Training for Robotics', and Google's blog post 'Behind Project Suncatcher, our moonshot to put AI in space'. For practical takeaways, Zhipu's first tip is to make agent feedback local: tie it to specific launch parameters, code changes, kernels, input conditions, threads, execution intervals or code paths, so the agent can narrow the problem. Dong and Finn's wish list for robotics is also a checklist: a default definition of success, a default way to reset the scene, and a default way for a person to give feedback and turn it into learning.

How solid is it

This is a secondary-source digest, so each claim is the newsletter's account of someone else's work. The quotes from Levin, Dong and Finn, Google and Zhipu are given as the newsletter quotes them. The Trillium radiation result and the Transporter-18 plan come from Google's update as relayed here. Zhipu's less-than-two-weeks and tripling figures are its own claims, with no absolute throughput numbers. The available text of the issue breaks off partway through the Zhipu item, so its remaining tips are not covered.

Risks and caveats

The newsletter itself says there are no home-run experiments behind Levin's hypothesis, only unusual phenomena such as xenobots, anthrobots and perturbed sorting algorithms. Dong and Finn present EXPO(-FT) as early in development and not widely used. Google is still working on cooling in space, which the newsletter expects to be a challenge. The 'outer RSI loop' label is the newsletter's framing: Zhipu's own wording describes engineers, an Infra Agent and an experimental environment working together, with engineers setting objectives and system boundaries. The prediction that computation moves into orbit very quickly is the newsletter author's opinion.

“mind:body is as math:physics”

— Michael Levin, in his paper 'Ingressing Minds', as quoted in Import AI 474