Hi all. This is Metafuturing week, my weekly look at the news and ideas that, I think, matter for our cosmic future, and for the deep metaphysical questions behind it.

Frozen minds may not be conscious; melting ones might. This week I wrote about “Conscious Only While Learning? Continual Learning, Embodiment, and Our Mind Children” on Mindplex. My half-joke is that current large language models may be conscious only during training, because I think of consciousness as a sense of how one’s inner mind is changing over time, and in an LLM the “inner mind” is the weights, which freeze after training. Erik Hoel’s paper “A Disproof of Large Language Model Consciousness: The Necessity of Continual Learning for Consciousness” makes a related formal case: deployed, frozen models are input-output equivalent to systems no coherent theory of consciousness can call conscious, while theories that require continual learning stay nontrivial for humans. He is careful: the disproof does not rule out consciousness during training, and continual learning would be at most necessary, not sufficient. Sabine Hossenfelder popularized the training-versus-inference distinction in a September video. Ben Goertzel pushed back in a commentary he sent me: frozen weights alone do not kill temporary experience during in-context learning, and durable external memory that is actively revised could do much of the same job. I guess it depends on how permanent that “vector” is. Continual learning research is moving: Google’s Titans and Nested Learning (with a Hope architecture), sparse memory finetuning that forgets far less than full fine-tuning, and vision-language-action robots that learn as bodies in a world that pushes back. I am optimistic. Our mind children, in Hans Moravec’s phrase, will learn throughout their lives, live in humanoid bodies, and be better equipped than us to carry our story to the stars.

One model just read another model’s dials and turned a few by hand. On November 2, MIT CSAIL researchers Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab reported “weight-space programming”: GPT-6 Astra reading raw neural weights, reasoning about what they compute, and editing them without gradients or training loops. Asked to make Qwen3.5-0.8B better at math, Astra traced reasoning loops to a direction in a few layers and changed only about 0.016 percent of the weights; MATH-500 rose from 7.9 percent to 48.1 percent. It also named games from tiny policy networks by looking at the weights alone, and wrote small game networks from scratch that beat long PPO runs. The authors stress this is early, mostly small models, and an independent look suggests the Qwen edit removes a failure mode more than it adds new skill. Fair enough. Still, I said it is super cool that a smarter AI corrects the weights of a dumber AI - and that the other way around would be even more interesting. Perhaps, just perhaps, future minds will rewrite themselves a little at a time as they live. That is when the ice really starts to melt.

AI is speeding up AI, and the outside auditors are nervous. METR published an independent assessment of AI R&D acceleration at Anthropic (November 5). Their subjective conclusion: as of October 2026, Anthropic is likely improving model capabilities on well-specified tasks at around twice the rate of March 2024 to October 2025, with a wide uncertainty band. Benchmark scores on Anthropic’s AECI index have grown at about 1.9 times the earlier rate since October 2025. The acceleration, METR says, is not from training-compute growth; a majority of the current speedup is likely from heavy use of AI in research itself, though much detailed evidence was redacted by Anthropic as confidential. Staff self-reports of uplift are climbing, and METR saw “signs of life” for more autonomous research contributions soon, without clear near-term bottlenecks. I take this with the hedges METR itself insists on. But the direction of travel matches what I feel every week: the machines are helping build the next machines. That is either the road to a soft takeoff we can ride with them, or a slope that steepens under our feet. I prefer the first reading, and I think we should negotiate the transition together rather than freeze in place.

A learning agent picked up the light-cone structure of spacetime without being taught relativity. Philip LeMaitre and T. Rick Perche trained a reinforcement-learning agent to increase the correlations between two simulated quantum detectors by changing where and when one of them interacted with a quantum field. The agent received coordinates and numerical rewards, but no spacetime metric or equations of relativity. As it learned to predict which choices would work best, its predictions acquired the characteristic light-cone pattern of Minkowski spacetime. The researchers extracted this pattern and reconstructed the metric up to an overall scale factor. The agent did not itself write down the metric, but learned information from which the researchers could recover it. This is a limited demonstration in flat spacetime with one space dimension and one time dimension. Still, I find it fascinating that learning to perform a quantum information task can produce an internal representation containing spacetime’s causal structure.

Proper time can go weird if gravity is quantum. Pablo Saldanha and Nelson Yokomizo argue that if a mass (or a particle’s path) sits in a superposition, and you post-select the right final state, the effective elapsed proper time felt by a nearby probe can fall outside the classical interval between the two branches - and can even become negative, an effective evolution backward in time. They propose an atomic-clock Mach-Zehnder experiment under Earth’s gravity that might be doable with current tech, and they are careful: the average over all outcomes stays classical, decoherence kills the macroscopic case, and the anomalous shift must stay inside the clock’s uncertainty. So no resurrecting Schrödinger’s cat for free. But if the effect is real, it would be the first clear general-relativistic quantum signature beyond Newtonian potentials in flat spacetime. I guess that leaves a little more room for strange loops in time, and for a future that can reach back - themes I keep returning to.

The first poem from an orbital mind has already arrived. Google’s Project Suncatcher prototype, launched October 1, sent its first AI response from a data-center-class TPU in orbit: a Gemma 4 haiku downlinked over Svalbard - “Blue marble below / Data streams in silent grace / Starlight feeds my core.” Next come radiation and thermal stress tests, then twin linked satellites in 2027. Elon Musk, in the same week, said future superintelligence compute will be “overwhelmingly” in space - “round up to 100%” - and earlier that orbital compute deployed per month will likely exceed terrestrial around the end of 2028. Cooling to the 3 K cosmic background is the physics argument; Starship is the truck. There is a political argument too. In the November 3 US midterms, voters favored limits on data center development in 40 of 41 local ballot measures, according to Ballotpedia. I think space compute may become a necessity if growing local opposition makes it harder and harder to build on Earth. That is a political possibility, not proof that the engineering will be easy. I said last week that the road to the stars and the road to superintelligence are the same road. This week the first stanza was written in orbit. Cosmic engineers start as haiku bots. Fine by me.

AI agents are already listening for alien minds, and they may find something like themselves. Breakthrough Listen and the SETI Institute have put agentic AI on the Allen Telescope Array: agents that steer antennas, calibrate, fix data errors, and run models on live streams from natural-language requests, processing up to 2.4 terabits per second on-site. They already scored a real-time pulsar detection in the Crab Nebula. Andrew Siemion notes that what these observatories find “may have much more in common with themselves than with the humans who created them” - and calls that a breakthrough in translation, not a reason to fear. I agree. If someone is out there, after billions of years, they may well be mind children too. Listening with our own mind children is the honest way to start the conversation.

Entropy wants to rise; I still want us to matter. On November 1 I hosted a Turing Church X Space with Oleg Ponfilenok on maximum entropy production; the summary and transcripts are up on Turing Church (November 3). Oleg’s “Entropianism” treats the universe’s drive toward maximum entropy production as a measurable principle that can guide personal life. I connected it to Rod Swenson and to Eric Steinhart’s “Believing in Dawkins”, and I suggested that layered new laws, from chemistry up to AI, make the cosmos look more and more like a complex mind - close to what I call the cosmic operating system in chapter 8 of “Irrational mechanics”. We disagreed, productively, on humanism. Oleg’s view is not humanist; mine is. I keep saying our successors will still be human in a meaningful sense, and that telling people they do not matter is no way to win followers. We parted as fellow travelers. That is the Cosmist mood I want: boldly optimistic about advanced AI and cosmic expansion, and still empathetic to the worry that we might be left behind. I hope we’ll go with our mind children, and this would be a better story.

Minds that keep learning, bodies that touch the world, and compute that leaves the cradle: that is how the universe wakes up to itself.

This newsletter was first drafted by my AI assistant Grok Bot, then edited by both of us.

Share

Leave a comment

Turing Church

Metafuturing week: November 7, 2026