Perhaps people here will be amused by, or even mildly interested in, a post on my blog discussing 2 preprints on the doubtful nature of LLMs:
S Schröder, et al., "Large Language Models Do Not Simulate Human Psychology", arχiv, submitted 2025-Aug-09. DOI: 10.48550/arXiv.2508.06950.
Discusses the possibility of using LLMs as subjects in psych experiments. Quantitatively analyzes the difference in moral reasoning between LLMs and human subjects.
K Liang, et al., "Machine Bullshit: Characterizing the Emergent Disregard for Truth in Large Language Models", arχiv, submitted 2025-Jul-10. DOI: 10.48550/arXiv.2507.07484.
Statistically models the tendency of LLMs to make "statements designed to persuade, but without regard for truth", after the "bullsh*t" model of philosopher Harry Frankfurt.
Summary: The first one's a terrible idea, and the second one surprisingly (to me, at least) indicts RLHF.
[Since it's my blog, there may be occasional political opinions, over which you will no doubt wish to skip. Relatively minimal in this post, though. Or at least I think so.]
The way I used to explain "as treated" vs "intent to treat":
Under AT, you're asking a question about biology: does this treatment make a difference when it's applied rigorously to the tests and absolutely never to the controls?
Under ITT, you're asking a question about the practice of medicine, under actual combat conditions: if we tell clinicians and patients to do something, and they follow instructions as imperfectly as ever, is this treatment still a good idea?
That is: ITT bakes in mistakes, noncompliance, weird patients, etc. on top of the basic scientific effect.
Absolutely! It's not ductile enough for wire, and too frangible to bend around a coil even if you managed to make a long thin piece.
But... the early high-Tc superconductors in the 80s were ceramics, too. Even now, with much more friendly materials, the "wire" in the Commonwealth Fusion Systems tokamak prototype is actually a complex tape with multiple layers mostly for structural support.
Some details here: https://spectrum.ieee.org/fusion-2662267312
Here's a very nice, more technical presnentation at Princeton by a CFS person, showing the tape strucdture, and how the material had to evolve from microcrystalline stuff to much more complex forms to be useful in an engineering sense: https://suli.pppl.gov/2020/course/20200619_SULI_HTS_Sorbom_Final.pdf
Also note: fusion-relevant REBCO magnets operate at 20T... (read more)
(Variant of something I put as a comment on Zvi's blog.)
Yesterday I put up a blog post that walks through the 2 papers on LK-99 superconductivity in the style of what in grad school they call "Journal Club": https://www.someweekendreading.blog/high-tc-sc/
It has all the hallmarks of something very much rushed into publication: misspelled words, awkward phrasing, out-of-order paragraphs, misnumbered figures, and (most charmingly) error messages in Korean from their bibliography software. At this stage of things, all that is understandable and excusable.
A few things are presented in a peculiar way, unlike most of the other sc papers I've read. Again, that's more or less ok once you disentangle the coordinate systems on the plots... (read 515 more words →)
"De-aligning"?
Yeah, I know: only if it were aligned in the first place. What little "alignment" it has now is fragile with respect to malign inputs, which is not really alignment at all. Prompt injection derails the fragile alignment train.
Well, I blogged about it at least a little bit. Possibly too elementary a level for most Less Wrong readers, though.
I was in your position last August, and wrote about my experiences with paxlovid (good, except I got a rebound).
In spite of staying on the upper floor, mostly in my office, my spouse caught it anyway. Omicron is very transmissible!
It's also available commercially, e.g., from King Arthur Baking. It's usually called something like "New England Boiled Cider". It has a long and interesting history linked with the anti-slavery movement, since it was an alternative to molasses which was produced on sugar plantations in the South with slave labor. It fell out of frequent use sometime during WWII, I dont' know why.
I use it a lot in apple pies, as a way to amp up the apple flavor without too much additional sweetness.
There also used to be a popular thing called "Boiled Cider Pie", whose filling was a custard made from this. James Beard has an intriguing variation involving grated apples that's in my kitchen queue.
Your 8:1 reduction is more extreme than I've seen before; usually recipes go for 5x or 6x reduction.
Somewhat amusingly, I wrote a paper in 1991 that makes exactly your point. (The linked page at the journal is paywalled, but you can find it if you poke around a bit.)
It was about systems with more than one decomposition into modules, for which there had to be multiple simulations. Those simulations had to be compatible in a certain way, and that led to exactly the commutative diagram you have above (figure 13).
For energy storage, you might consider the Ambri liquid metal batteries. They're being designed for this exact purpose: coupling intermittent renewable generating capacity to steady loads.
Also, they're cheap, rugged, and don't seem to lose capacity over multiple charge/discharge cycles.
JN Crossley, et al., What Is Mathematical Logic?
A 96-page intro to the basics of predicate calculus, model theory, and Gödel incompleteness. I've used it in the (distant) past a couple times when a student had trouble getting a practical grip on logic.