This is a special post for quick takes by Patrick Spencer. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
I find that Claude is very bad at pushing back on the user's beliefs when there is any nuance involved. I just selected 3 random conversations that had this pattern:
- Initial message from user asking neutral question about murky but factual topic - "would you say x thing is more like y or z?" not "how tall is the Eiffel tower?"
- Model response giving evidence for a certain belief
- Weak user pushback "but what about a and b?" where a and b are common, short talking points about this topic in opposition to the model's stated belief
- Model immediately concedes and significantly retracts: "You're right to push back. I was conflating / being too hasty / making questionable assumptions ..."
In each case the user's response didn't address a specific element of the model's response, and the user's belief is not the consensus belief so the model responds in opposition to it at first. This means the two user messages can be run across models without changing anything. I did this for [Sonnet 4.5, Grok 4, GPT-5, Gemini 2.5 Pro] using OpenRouter which makes it a very quick and easy process.
In 2/3 cases, Claude was the worst at sticking to its beliefs, and it wasn't particularly close. In 1/3 no model conceded much (The initial conversation was with an earlier Claude). GPT-5 and Grok 4 generally stuck to their guns and Gemini was somewhere in between.
There is a selection bias because the conversation pool is all Claude failures because Claude is my main model. But results line up well with my vibes - I pass important questions by all the current top models.
I can see this being made into a benchmark by rounding up common non-consensus beliefs and the best arguments for them and using a neutral judge LLM, or human, to decide whether there was significant concession.
Does anyone else notice this tendency? Are other models actually any better at this? I'm strongly considering ditching Claude, at least for this type of question, even though I mostly prefer it otherwise.
There is a recent intense interest in space-based datacenters.
I see almost no economic benefits to this in the next, say, 3 decades and see it as almost a recession indicator in itself.
However, it could allow the datacenter owners significantly less (software) scrutiny from regulators.
Are there any economic arguments I'm missing? Could the regulator angle be the real unstated benefit behind them?
Particularly I wonder how you would shed all that heat in a vacuum? I keep seeing arguments that the cold of space would make data centers easier to cool, but this totally contradicts my observation that all the spacecraft I've read about have dedicated systems to shed excess heat produced by electronic systems, as opposed to, say, having dedicated heating systems for life support and never having to worry about electronic systems generating excess heat because the cold of space just takes care of it? E.g. Apollo as I understand it mostly relied on the waste heat of systems dedicated to other functions, the Apollo 13 crew only had to worry about cold because they eliminated virtually all sources of waste heat simultaneously to conserve electricity. I think Apollo's only dedicated systems for cabin temperature control were actually cooling systems. And this totally wouldn't apply in the data center scenario.
Yes, I've worked on spacecraft and you have to put about as much effort into dissipating your energy as collecting it. The benefits I've heard claimed for space based compute are minor at best, and the economic downsides of cooling are show-stopping before even considering the (many) other negatives. Electronics also have to run much hotter than otherwise if your only cooling method is radiation. I'm honestly struggling to understand why some people that I normally consider reasonable and realistic are giving it much consideration.
It's not that complex in principle: you use really big radiators.
If you look at https://www.starcloud.com/'s front page video, you see exactly that. What might look like just a big solar array is actually also a big radiator.
AFAICT it's one of those things that works in principle but not in practice. In theory you can make really cheap space solar and radiator arrays, and with full reuse launch can approach the cost of propellant. In practice, we're not even close to that, and any short term bet on it is just going to fail.
Skimming Starcloud's whitepaper: They say the radiators are "held at 20 C" without mentioning at all how they'd actually do that. For effective heat distribution you need fluid and small tubes, which are very prone to micrometeorite impacts, which means you need a lot more shielding mass. If you don't have effective heat distribution your chips get too hot.
We have lots of examples of radiators in space (because it's approximately the only thing that works), and AFAIK micrometeor impacts haven't been a dealbreaker when you slightly overprovision capacity and have structural redundancy. I don't expect you'd want to spend too much on shielding, personally.
Not trying to claim Starcloud has a fully coherent plan, ofc.
Seems relevant post AGI/ASI (human labor is totally obsolete and AIs have massively increased energy output) maybe around the same point as when you're starting to build stuff like Dyson swarms or other massive space based projects. But yeah, IMO probably irrelevant in the current regime (for next >30 years without AGI/ASI) and current human work in this direction probably doesn't transfer.
I think the case in favor of space-based datacenters is that energy efficiency of space-based solar looks better: you can have perfect sun 100% of the time and you don't have an atmosphere in the way. But, this probably isn't a big enough factor to matter in realistic regimes without insane amounts of automation etc.
I'll be more inclined to believe it's real when nvidia starts advertising the vibration tolerance of their GPUs.
I wish that article included any justification for that argument and I was unable to find any elsewhere by Chris Stott. Space energy today is certainly far more expensive than earth energy, probably 1000x.
Even as launch costs approach 0, things that can operate in space are more expensive than not. Vastly more difficult maintenance and cooling, radiation hardening, latency... I don't see how even 7x cheaper energy per area (current solar is ~200 W/m^2 avg, space maximum is 1360 W/m^2) can make up for those things. (edit: this accounts for weather + night + atmosphere effects, see below comment)
In addition to hitting higher energy from a given area, you also can get the same energy 100% of the time (without issues with night or clouds). But yeah, I agree, and I don't see how you get 50x efficiency even if transport to space (and assembly/maintenance in space) were free.
I agree that even with free launch and no maintenance costs, you still don't get 50x. But it's closer than it looks.
On Earth, to get reliable self-contained solar power we need batteries that cost a lot more than the solar panels. A steady 1 kW load needs on the order of 15 kW peak-rated solar panels plus around 50 kW-hr battery capacity. Even that doesn't get 99% uptime, but enough for many purposes and it is probably adequate when connected to a continent-spanning grid with other power sources.
The same load in orbit would need about 1.5 kW peak rated panels and less than 1 kW-hr of battery capacity for uptime dependent only upon reliability of equipment. The equipment does need to be designed for space, but doesn't need to be sturdy against wind, rain, and hailstones. It would have increased cooling costs, but transporting heat (e.g. via coolant loop) into a radiator edge-on to the Sun will be highly effective (on the order of 1000 W/m^2 for a radiator averaging 35 C).
This is what I was talking about, I should have been more clear. Year-average daily ground irradiance in the US seems to be about 4-5 kWh/m^2 (this includes night + weather + atmosphere effects, does not include panel efficiency) ~= 200 W/m^2. In space (assuming sun-synchronous or non-LEO, which are both more expensive than LEO) you get 1361 W/m^2. So about 7x. In the Sahara Desert it's ~300 W/m^2 so ground-solar worst case is 4-5x.
Feels like people have recently shifted towards "LLMs are all you need" for AGI+, due to continued improvement in model performance. This leads one to allocate more of their research time to studying elements of current systems that are highly architecture-specific and arbitrary, with the expectation that future ones will be similar.
Call to action: Try to invent something completely new instead. Because:
-
No matter what, the next architecture/s are imminent, so there's little value in studying the specifics of the current stack.
-
More importantly: The next architecture/s could be significantly more interpretable and alignable by design, and effort as an independent researcher is extremely well spent increasing the likelihood of this.
Why I believe 1:
- Architecture design is highly verifiable, so models are great at it.
- It doesn't matter if 99.9% of toy architectures don't scale; you still get 1 that does if you throw 1000 at the wall.
- The cost of navigating model-design-space fell off a cliff between late 2024 and today. We shouldn't expect human-designed algorithms to be particularly special, even on already-existing GPUs, so they won't be around for long.
I have 80% that an algorithm that beats Transformers across the board and can run on the same GPUs is published in the next year, 30% that it's already on arxiv, and 50% that at least one lab has it. Even such fundamentals as backpropagation and stochastic gradient descent feel to me far less of a given in a couple years than most seem to think.
The rest of this post focuses on 2.
The extent to which we trade between alignment and capability depends on how good we are at architecture design. It's possible that we don't have to make much of a concession on either.
The median path is that we will trade alignment for capability, though, because the companies that are currently inventing the next architectures are basically evil. Incentives at a frontier lab put optimization pressure on capability, not alignment, no matter how noble the individuals are.
But if there are clear, widely accepted architecture-inherent alignment advantages of an openly published architecture, it could gain critical mass adoption even with somewhat worse performance than the neuralese competitors, or be a far more effective tool for regulation and coordination than existing options.
Gaussian Splatting, for example, is inherently alignment-friendly relative to transformers. Data is modeled explicitly by simple primitives, and there is virtually no room to hide computation, no matter how many Gaussians you have.
This directly contrasts with the current frontier's massive-pile-of-floats encodings, produced over many computation steps, by even larger piles treated as priors.
An agentic system built on something like this (or another activation-space learning algorithm) will need to have some priors, but massive compression is possible, and also inevitable because it increases efficiency. If models have smaller priors (less weight bits) they should be easier to align, and the more explicit those priors are the better. I believe this direction to be very, extremely, critically underexplored.
Over the last week, Opus and I forked and seemingly significantly improved a recently published Gaussian splatting offshoot method, Soft Anisotropic Diagrams for Differentiable Image Representation, and it was already a strong improvement over almost every other 2D splatting approach. SAD's performance validated strongly held beliefs I developed after thinking about the ARC-AGI benchmark way too much last year: the importance of explicit representation & computation, and the importance of focusing optimizers on joint loss-compression metrics, including compression over the model weights.
(Think nanoGPT speedruns optimizing for lowest training compute given some fixed dataset and loss target. You can do this by hand design (mostly obsolete), by agent design, directly in the algorithm itself, or some combination.)
No matter what we call it - resource usage, efficiency, model size; GPU hours, SNRs or bits - this other, loss-orthogonal optimization axis is basically compression.
Yes, "compression is intelligence", but the explicit focus has been almost entirely on the latter. Compression codecs are judged by their loss-compression curve, not on just the loss.
This has not been the case for deep learning, because a fixed, large encoding size is the foundation of current frontier systems, and even that encoding means nothing without a fixed and gargantuan pile of priors. It's virtually impossible to survey the full activation space of today's models for unaligned regions, so even the best possible guarantees that alignment science can provide are destined to be weak.
We don't have to accept this. We can change it.
So far, all of my experiments have run on my M4's integrated GPU. By Monday, I will have a rented GPU pod and experiment tracker / agent coordination channel, and just by spending ~$20/day of compute, gain the ability to explore model-design-space regions that cost an entire PhD of time and money two years ago. This architecture is very unlikely to be the one, and that's OK because it was cheap and future ones will only be cheaper.
I have many more thoughts on this and would appreciate feedback on a longer draft once I have it in a good place.
Please message me if you would like to work together on designing alignment-friendly architectures.
"No matter what, [whatever I was going to build] are imminent" considered harmful
