Alan

This account is to link to other sites and communicate about blog related things. Also, I like audiobooks.

[Alan @AlanRominger]

FWIW, I asked ChatGPT 5.6 Sol (High) if quantum gravity could plausibly be solved by AI in the next year. It thinks "yes" (there is a real chance), but won't be recognized as such within that year.

[Zohar Komargodski @ZoharKo]

The result about S^6 having a cmplx structure is striking. The best mathematicians were working on it (I believe it was an obsession of Jim Simons). It is a little surprising that it is in the convex hull of what was known before. (Assuming the argument holds up -- and experts tell me it looks very plausibly correct.)

Even (very few) physicists encountered this question in the study of rigid supersymmetry in 6 dimensions. That is how I learned about it around 10 years ago when I was looking into rigid supersymmetry in lower dimensions, where we encountered transverse holomorphic foliations.

I have to check it more carefully but I believe with this new result, we now know a topological S^6 does admit a Hermitian metric (albeit a highly non-standard, squashed one). From a physics perspective, this implies the existence of rigidly supersymmetric backgrounds on a topological 6-sphere. Maybe there is some connection to localization of the partition function.

[Alan @AlanRominger]

I have planned a 2nd post on ridiculously giant fusion power plant designs - inertial fusion without compression. Here is a meme for it.

[Alan @AlanRominger]

Marking for the prediction of quantum gravity being solved in the next year.

Progress in math demonstrates ability to contribute at the tip-top human level. But quantum gravity is civilizationally meaningful, and unobtainable otherwise.

[Michael McNair @michaeljmcnair]

People still do not appreciate what is about to happen in science.
I am by no means a professional theoretical physicist, but it is my passion. And what I have been able to accomplish in the past month or so working with the newest AI models is nothing short of unbelievable. I don’t want to get ahead of independent peer review, and I’m fully aware of how easily people can fool themselves with previous models. But by forcing the outputs through rigorous code execution and mathematical oracles, some of what has come out of this work has the potential to be genuinely, fundamentally groundbreaking.

The models have finally crossed a threshold (and to be clear that is in math and coding) where the amount of unresolved math and physics questions are about to fall at a pace that will be mind-blowing.

Phase Transition
Over the past month, we’ve seen a slew of groundbreaking, Nobel-tier math discoveries. My theory has been that when genuine AI assisted (or fully derived) breakthroughs in math and science finally appear, progress won't be linear…it will go exponential.

People have wondered why years of dramatic improvements in AI benchmarks didn't trigger an explosion of actual mathematical discoveries. The reason is that if an AI is substantially below the human frontier, improving it from, say, the level of a good undergraduate to a strong graduate student to an ordinary professional mathematician can produce huge benchmark gains but surprisingly little genuinely novel math. A mathematician at that level is working inside territory humanity has already saturated. They're just discovering things we already know.

But when AI matches the upper tail of the humans who actually create the frontier, the situation completely changes.

Open problems aren't randomly distributed across a smooth difficulty spectrum, because humanity has already solved everything accessible below the edge. Now an incremental capability bump doesn't just solve harder known problems, it pushes the model straight into a massive, untouched inventory of unresolved questions living where only the best human minds operate. That alone will produce a phase transition.

And the math frontier has historically been bottlenecked by bandwidth, because there are so few humans that are capable of actually doing it. But when an AI reaches Terry Tao-level capability, the equation isn't one AI vs Terry Tao. It’s millions of instances operating at that tier, continuously, across thousands of problems, backed by perfect recall of the literature, code execution, numerical experimentation, ect. That level of parallelism alone will create exponentials.

Frontier math hasn't just been constrained by how thin it is, but also by how specialized those people have to be. To operate at the frontier of a modern math, you spend decades focused on learning an enormous amount of domain specific information. Which creates a tradeoff where the deeper your expertise is, the harder it is to also posses deep cross domain intuition. AI changes that.

To be clear, I’m not arguing that humans become irrelevant once models reach frontier capability. A human brings the things the model might not have…a strange analogy, an intuition about a pattern they noticed in another field, a thought experiment, or the sense that two apparently unrelated phenomena are actually manifestations of the same underlying structure. The model then provides the rigorous mathematical capability that historically would have prevented people from actually pursuing those ideas.

The successful combination is a human that thinks like Einstein- someone who has intuition and creativity that steers the search (Einstein famously used thought experiments to guide his research). The AI is like Paul Dirac, who famously on let the math lead him and followed it whereever the formal structure pointed, as opposed to starting primarily from physical intuition.

The combination of a creative, cross-domain knowledgeable human, interacting with the AI will be most powerful. The human can choose unusual conceptual directions and the AI can explore the consequence-space at insane depth.

But cross-domain creativity usually suffers from the fact that most analogies are wrong. So AI’s usefulness isn’t just that it can generate more speculative connections, but that it can subject those connections to rigorous mathematical pressure. The model creates a filter between creativity and scientific validity. Without that filtering mechanism, scaling human intuition would mostly scale pseudoscience.

And the superpower isn’t just math or coding alone, it’s the combination of the two that creates something like a closed discovery loop. The coding dramatically reduces the cost of being wrong, and because research is an iterative process, lowering the time and cost of each iteration compounds into enormous productivity gains.

Where the Breakthroughs will Arrive First
Math will be the field that see’s the most immediate advances for a couple of reasons. Math (and code) have an uniquely strong oracle - a fast, automatic way to check if an attempt was good. Because of that, the entire research loop can occur entirely in compute.

There are a lot of unresolved physics problems that can be solved this way as well, but there are two kinds of problems. In the first kind, we already possess essentially all of the relevant empirical and theoretical information, but we haven’t found the correct synthesis yet. Those will be quickly solved (and there’s a lot of those). The second class consists of data-constrained problems, that hit a physical data bottleneck, where you actually need to run real world experiments to test hypotheses and see if they correspond to reality before you can progress your theory. Those will fall much slower.

I will make a truly bold prediction: quantum gravity will be solved in the next year.

Markets
The relevant point for those in markets is to ask yourself how these stocks are going to react when suddenly a deluge of groundbreaking discoveries start getting announced. Because that's what's about to happen.

This next year is going to get weird.

[Alan @AlanRominger]

@jwt0625 @attack_orange You can use fossil cold from the Oort cloud, yes.

[Alan @AlanRominger]

@jwt0625 @attack_orange I used to think this would be different in space, and breathlessly did calculations and wrote blog posts on it.

But no, simple T^4 radiator law sets optimal operating temperature notably higher.

We will compute hot until fusion fuel availability matters which is never-ish.

[Alan @AlanRominger]

@is_OwenLewis What's the exact line between a free floating habitat and a ship? What about cyclers?

[Alan @AlanRominger]

I have daydreamed about heaver-than-air loitering aircraft. Not good ideas, that's why I never shared. Just one example - why do helicopters not have a skirt? C'mon, Delta P right? Think of the eddies...

One look at this, I don't need to draw the force diagram, pure genius.

[Hunter Weiss @Hunter_Weiss]

I laughed when I watched this one. Amazing that someone made this, and it flies

[Alan @AlanRominger]

I just published Design of 1 TWe Fusion Power Plant https://t.co/UWrc9PN6YE

[Alan @AlanRominger]

@PhilaSzn @scaling01 Completely valid. It is a real problem, and creates a new form of overfitting which is like AI group think.

Domains with cheap data synthesis with see 99% of the gains. So, verifiable math stuff. Letting the AI use Linux.

We will make an AI Linux homunculus

[Alan @AlanRominger]

If true, this scenario is good for humans in the general. Kurzweil gets proven right, we have a singularity of cyborgs. LLMs are the next layer of the neo cortex. The thalamus directs 1,000 internal brains which direct a million agents.

[Adam Hunt @RealAdamHunt]

Recently I've flipped from being bullish to being bearish about AI.

I think I'm updating my bearishness to be more solidly bearish. Early thoughts (which I hope to be disproven in the next year or so, I would prefer progress) and my reasoning:

The whole 'it turns out if you keep training and scaling the models more they develop broad new capabilities in lots of domains' thesis is wrong (sorry Demis). The recent batch of models haven't got more general, they've got less general. This is most obvious in the fact that their language outputs have got much worse in comparison to e.g. o3. If they were gaining generalist capacities we would expect them to be describing their work in ever more graceful and comprehensive prose!

The image that was being shared as the AGI thesis (November 2025, Tomas Pueyo) was the spiky bubble that has a current spike or two out past human capabilities (e.g. on coding or math) but below human on other capabilities on the other spikes - the future prediction was that as the models scale/advance, every spike would grow bit by bit until the whole center encompasses the human capabilities, with super-superhuman on some spikes. I think it seems like what's actually happened in the last few models has been that the coding/math spike has grown, but leaving behind or even at the cost of the other spikes. The models are no better at some simple logic, language (and sometimes worse!).

This makes sense from a simple RL perspective; you can't RL something endlessly on one domain of tasks and expect it to improve on the other tasks. The fact that early LLMs did seem to improve generally was a byproduct of the written language corpus covering everything - that corpus is general, so training it on that gave the appearance of something generally intelligent and becoming more generally intelligent as it got better at replicating that corpus. But the actual logic and underlying ground truths behind the language aren't captured efficiently enough and weren't effectively RLd in - they top out at some point (I guess this happened around the time that there was the 'has scaling hit a wall' discussion in late 2024). Chain of thought was then a genuine breakthrough, along with web search, which plugged into that general LLM global-corpus intelligence to lead to post 2024 gains.

The AI companies have since worked out that coding works (and pays) really well (basically this is because the entire job is nearly perfectly recorded and exists as training data, and you can set up clear benchmarks and rewards). The recent models (and benchmarks) have been maxxing that and we've seen degradation on normal English use for that reason. This could still be transformative, leading to extremely powerful (and potentially dangerous, particularly in cyber security) models but it's not a pathway to AGI.

I'm probably at about 40% confidence about this. It fits my current observations of AI progress and has a basic explanatory model. It doesn't account for potential breakthroughs, which is a major reason for discounting.
To make some predictions, I guess if I'm right this will become broadly apparent and more widely acknowledged in the next year or two, as we see how the spikiness of models that keep getting released develops.
Maybe there will be efforts to concentrate on specific spikes e.g. health or law which require going back to earlier models and RLing on a different data set/with different rewards/benchmarks. Maybe those separate models can be linked together to give a more apparently general model. How capital intensive that is/the potential profitability will be a defining question. But I just don't see general abilities emerging atm, and I don't think we will any time soon. Good news - a whole industry of tackling important specific problems/sectors can open up!

[Alan @AlanRominger]

@without__excuse @AkalStation @DrChrisCombs Coat the existing tiles with the ablative layer, and have a temporary sealant between the tiles you pull out like a stencil.

[Alan @AlanRominger]

@AkalStation @DrChrisCombs Ablative dippin' dots

[Alan @AlanRominger]

@HopDavid41 The current solutions are deorbit and disposal orbit.

We need a 3rd solution, which is collect and bag. Orbital landfill. We will reuse them later. First as mass, like you say, later, materials.

Need several, based on orbit parameters, transfer requirement.

[Alan @AlanRominger]

Marek M. Kamiński, Games Prisoners Play: The Tragicomic Worlds of Polish Prison

Its key points about game-theory in prison are morbid and fascinating. People are weird, but sometimes it's really information asymmetry.

[Alan @AlanRominger]

Reading about Novatron fusion. Seems pretty institutional / legit. Surprising how this is very new, and recently funded. Comparatively open on technology details, at least, for now.

[Alan @AlanRominger]

I just published Mining Geography of the Solar System https://t.co/XZejXWf5Ba

[Alan @AlanRominger]

@is_OwenLewis @AstraPublica @SpaceX They are closer to the data, so that counts for a lot. If nothing else, there will be a market for data warehouses as that data is sitting in the queue to go to orbit.

At least until humanity moves to space habits.

[Alan @AlanRominger]

I just published L1 Diamond Stabilizer https://t.co/82hUZFbz5z

[Alan @AlanRominger]

Strange experience, was having an AI look for AI text. It could competently identify this pattern in the text. But it always _praised_ the writing style. Suggests it does it... because it thinks it's good.

[Eric Levitz @EricLevitz]

The most puzzling AI-ism to me is probably the "Not x. Not y. But z."

Not the em-dashes (an essential piece of punctuation). Not "That isn't x, it's y" (a useful if inelegant way to clarify an argument). But consecutive examples of what your subject isn't -- conveyed in fragmentary, staccato sentences -- before a declaration of what it is.

Feel like this is an inherently irritating rhetorical device. And I don't recall regularly coming across it in pre-AI writing. So, I don't understand why LLMs are so in love with the template

[Alan @AlanRominger]

I started thinking about how to make a petawatt radiator in space. It's about the size of Mars, but something felt wrong about it. So I thought about an exawatt radiator. Legitimately hard in a dimensional analysis sense. But I've got something now.