Import AI 474: Platonic… | Jack Clark from Import AI Summary

1. Key Themes

Robotics is at an "LLM moment" but lacks the standardized post-training recipe that unlocked language models

The article argues robotics pretraining has scaled well, but the field is missing a shared, reliable methodology for fine-tuning on real hardware.

"Robotics is sitting almost exactly where language modeling was: The pretraining has scaled beautifully. 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."

The needed elements are concrete and standardizable — success criteria, reset procedures, and feedback mechanisms.

"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."

Recursive self-improvement (RSI) is quietly moving from theory to practice inside AI labs

Zhipu AI used its own model (GLM-5.3) to build its own inference infrastructure, compressing a multi-week engineering task and dramatically boosting performance.

"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."

This signals labs are operationalizing self-improvement loops with defined engineering discipline (local, cheap, and verifiable feedback), not just using AI as a coding assistant.

"Feedback must be sufficiently local... Feedback must be inexpensive and timely to obtain... Feedback must support objective verification."

Compute infrastructure is expanding beyond Earth

Google's Project Suncatcher is a serious, tested effort to move AI compute into orbit, motivated by space and solar constraints on Earth.

"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.'"

Jack Clark frames this as an inevitable consequence of compute and energy scaling trends.

"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."

AI is being embedded directly into physical scientific discovery loops

Periodic Labs trained a model (Periodic Neon) using reinforcement learning tied to real lab equipment, not just simulated or textual data, to accelerate materials science.

"Periodic Neon got a 55.3% success rate on their internal evaluation set ('FrontierXRD'), a 20X improvement over the 2.7% success rate of Kimi 2.6."

This is part of a broader wave of "AI scientist" startups automating discovery across multiple domains.

"Startups like Periodic Labs are part of a new wave of companies which are seeking to turn AI systems towards automating science, ranging from materials science (Periodic), to AI R&D itself (Recursive, Mirendil, Discovery Loop), to science broadly, to chip design (Ricursive)."

2. Contrarian Perspectives

Minds may not be purely physical computational products — they could be "interfaces" to a Platonic space of patterns. Michael Levin's paper challenges standard physicalist assumptions in AI/neuroscience by proposing that brains, bodies, and even algorithms are merely vessels that patterns from an abstract mathematical-like space "ingress" into, rather than minds being generated purely by their substrate.

"Bodies (whether living, engineered, or hybrid) are interfaces for a massive, multi-scale hierarchy of patterns to ingress into the physical world."

This reframes AI alignment debates: rather than asking whether an LLM's mind is "real," the more useful frame may be whether human and AI minds are neighboring — not identical — points in a shared abstract space.

"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?"

Lab workers are ambivalent, not celebratory, about AI automating their own jobs. Rather than pure enthusiasm about RSI progress, Zhipu's own engineers express a bittersweet resignation about being replaced by their own creation.

"Today, GLM-5.3 has become an indispensable daily coding partner for everyone on the team, and it is moving steadily toward replacing us."

They still hold onto a belief that humans retain a durable edge in high-level design — but even that is framed as fragile and temporary.

"We believe humans should continue to hold that line for a long time to come... Progress at this boundary will not slow down simply because we want it to."

3. Companies Identified

Physical Intelligence — Robotics company co-founded by Chelsea Finn. Mentioned as a source of thinking on why robotics lacks a universal post-training recipe akin to LLM RLHF pipelines.

"Chelsea Finn, a Stanford professor and co-founder of robot company Physical Intelligence."

Google (Project Suncatcher) — Initiative to launch TPUs into space for AI training/inference, partnering with Planet and SpaceX. Mentioned as evidence that compute infrastructure is moving off-planet.

"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.'"

Zhipu AI (Z.ai) — Chinese AI lab behind GLM-5.3, used to automate its own infrastructure development. Mentioned as a real-world case of recursive self-improvement (RSI) in action.

"Much of the work was carried out by an Infra Agent powered by GLM-5.3."

Periodic Labs — AI startup building models trained via RL on real physical lab experiments (Periodic Neon) for materials science discovery (superconductors, magnets). Mentioned as a leading example of "AI scientist" automation.

"A 1-trillion parameter model 'that outperforms GPT-6 Astra on a critical scientific analysis task using relatively little compute.'"

Old Models Foundation — Organization/site preserving legacy generative AI models (DRAW, DeepDream, DCGAN, StyleGAN, etc.) for public use. Mentioned as a novel effort to document AI history and aesthetics.

"Today, the Old Models Foundation lets you access ancient models... through to the comparatively modern ones."

Companies referenced as part of the "AI scientist" wave: Recursive, Mirendil, Discovery Loop, Ricursive — noted collectively as competitors/peers automating AI R&D and chip design.

"ranging from materials science (Periodic), to AI R&D itself (Recursive, Mirendil, Discovery Loop), to science broadly, to chip design (Ricursive - not a typo!)."

4. People Identified

Michael Levin — Scientist proposing a non-physicalist theory of mind based on synthetic morphology and diverse intelligence. Mentioned as author of an "iconoclastic" paper reframing mind-body relationships.

"I propose that the relationship between mind and brain is the same as the relationship between mathematical patterns and the morphogenetic outcomes they guide."

Perry Dong — Stanford researcher and co-author (with Chelsea Finn) of a piece diagnosing what's missing in robotics post-training. Mentioned as originator of the EXPO(-FT) algorithm concept.

"Perry Dong, a Stanford researcher, and Chelsea Finn... have written a nice piece about what is holding back robotics."

Chelsea Finn — Stanford professor and Physical Intelligence co-founder. Mentioned as co-author diagnosing the lack of a universal robotics post-training recipe.

(See above quote; role established alongside Perry Dong.)

Brian Eno — Music artist quoted regarding the aesthetics of new mediums, used to frame the appeal of legacy AI image models.

"Whatever you now find weird, ugly, uncomfortable and nasty about a new medium will surely become its signature."

5. Operating Insights

For AI/robotics builders: prioritize standardizing feedback loops before scaling models. Both the robotics piece and Zhipu's RSI writeup converge on the same operating principle — feedback must be local, cheap, fast, and objectively verifiable to make automated optimization loops actually work.

"Feedback must be sufficiently local... Feedback must be inexpensive and timely to obtain... Feedback must support objective verification."

Use small, targeted RL fine-tuning on top of large pretrained models rather than brute-force scaling. Periodic Labs achieved outsized gains (20x improvement) with a small compute footprint (1,300 H200s) by doing targeted RL on domain-specific lab data atop an existing pretrained model (Kimi 2.6), rather than pretraining from scratch.

"Neon was created by midtraining on top of Kimi 2.6 and then doing reinforcement learning on data from its labs, which is the key differentiator for Periodic."

Internal automation tooling can compress multi-week engineering cycles into days. Zhipu's Infra Agent example shows that dogfooding your own frontier model on infrastructure work yields measurable, fast returns — a replicable playbook for other AI companies.

"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."

6. Overlooked Insights

Google's space-compute cooling challenge may be the actual bottleneck, not launch or radiation survivability. While the headline is about TPUs surviving launch g-forces and radiation, the harder unsolved engineering problem — thermal dissipation in vacuum — is mentioned almost in passing but could determine feasibility.

"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."

The framing of "comparative advantage" for humans in AI-driven engineering is narrowing to high-level design decisions only — and even that is treated as temporary. This is a subtle admission from an AI lab itself (not an outside critic) that the human role in technical work is shrinking faster than expected.

"We believe humans should continue to hold that line for a long time to come... Progress at this boundary will not slow down simply because we want it to."