986aignan
RLHF has a way of squashing variance. If the different models are trained on each other or on the same human feedback, then the training would reduce the variety of answers while the similarity of feedback would lead the common answer to be similar.
The different models may prefer the same controversial takes for the same reason that they tend to prefer the same jokes and names. Maybe the authors of the article would have had better luck if they'd used verbalized sampling prompts or something like it.[1]
Maybe you could use a trusted timestamping service to stamp keyframes or stills from your video while you're recording? If the time of the recording is not in question, it would bound the kind of video editing people could plausibly claim that you could've done.
It's not foolproof - if you have enough face data of the person you want to frame and you've been to the room where the conversation takes place, you could have made a fake video beforehand and spliced it into the conversation at the time of the recording. But that would at least take a lot of preparation and quite a bit of luck and timing to find a place in the conversation that's just long enough to copy your edited version in without arousing suspicion. For dashcam videos it would be difficult to merge your edit with existing traffic, etc...
And something like attributions can be determined even if the LLM is paraphrasing, if I'm understanding Cosma Shalizi correctly[1]. He gives Grosse et al.[2] as a reference for doing so in a scalable manner.
It turns out the server is using referer checking and I've got referer disabled. Compare the following curl invocations:
curl 'https://fabiensanglard.net/jurrasic_park_computers/00_11_15_Apple_Powerbook_100.webp' \
-H 'Referer: https://fabiensanglard.net/'
curl 'https://fabiensanglard.net/jurrasic_park_computers/00_11_15_Apple_Powerbook_100.webp'
At the end of the thread, David Given reports that some PC floppy drives drop pulses when trying to read 800k Mac disks: https://github.com/davidgiven/fluxengine/issues/75#issuecomm...
So the Cambridge page might be saying "use an Applesauce controller" as a sort of shorthand for "use a drive from an actual Apple computer".
As you mention, there's a problem where if you want to know yourself fully - including observability of your mind - you need to be aware of the thing that is aware, and aware of the thing that is aware of the thing that is aware, etc, down to infinite regress.
But that doesn't necessarily mean that it's impossible. It would suggest that such a being works by moving along a set of fixed point. Suppose x is the state of the physical mind (whether that's a brain, a set of chips, or something else). Suppose f(x) returns the updated state after this mind takes into account what it knows (i.e. f(first-order aware) = first- and second-order aware). Then if the mind-state always changes so that x is a fixed point of f, then the mind already knows all there is to know about itself.
So the question, on the mind side, is whether there exists a sufficiently generally intelligent type of mind for which moving along fixed points is computationally easy. I don't know. It could be possible, it could be impossible. Finding fixed points is hard in general, though.
There's also the problem of sensors, which may seem simpler but hides a similar difficulty. What happens if whatever thing that measures blood sugar fails? Or if the immune system goes auto-immune due to an immune cell defect that makes it report benign cells as foreign? Then you need an integrity sensor for the sensors - then you need an integrity sensor for the integrity sensor... At least this fixed-point problem for sensors should be easier than the one for general minds.
The consciousness:complexity ratio would probably express itself in that a mind of a certain complexity with the fixed-point-of-f constraint would be less intelligent on average than a mind of the same complexity without such a constraint.
If the mind is supported by or comes from the physical world, then the hard question is "why is there something it is like to be me"?
If the physical world is supported by or comes from the mind, then the hard question is "why is the product of my thoughts sometimes incredibly malleable and other times not at all?"
From a pragmatic perspective, there are certain events that behave the same whether the mind came first and is somehow restricted in certain capacities, or if the natural world came first and is imposing itself on the mind (through whatever supports it).
For instance, falling down stairs is going to hurt in either case. If the physical world exists independently, that happens because you either are or have a body which is also subject to its laws. If there's a mental monism, that happens because you can't shape all your thoughts, and those thoughts you can't shape act on some other part of you in a way that injures what you think of as your body.
But you can view consciousness as a natural phenomenon without being reductionist. In a Hempel's Dilemma-like turn, you could say something like: "consciousness, like mass, is a property of arrangements of matter and exists wherever matter is arranged in a particular way. Disrupt the arrangement, as with anesthetics, and the consciousness goes away."
You end up with something like integrated information theory: https://iep.utm.edu/integrated-information-theory-of-conscio...
From such a perspective, the article's byline, "Consciousness is not separate from the physical world — our “soul” is of the same nature as our body and any other phenomenon of the world", is true. Like mass or charge, consciousness is merely another property or feature of stuff of combinations of matter that exist in the physical universe.
But there's still a "hard problem of consciousness" with such a theory. The distinguishing feature of qualia-like consciousness remains: it can only be properly verified from the inside. Researchers may devise theories that say "if property X holds, then the lump of matter is conscious" (like Tononi is doing with IIT). And the theory they develop may be quite tight - for all actions where it predicts temporary loss of consciousness, people exposed to the experiment say "I wasn't conscious at that time". But until they can solve the hard problem - being able to detect the what-its-like from the outside, the hard problem remains.
Though, as you're saying, if you just want something that predicts observable outcomes, then consciousness theories that say "this anesthetic-like thing produces what, to the outside observer, is indistinguishable from loss of consciousness", might be good enough.
I think SV was asking what onion services, which can't really use recaptcha, do to prevent the DDoS storm.
And I would imagine the answer is obscurity, since the dark web isn't nearly as well-mapped as the public web. That and some Anubis or other PoW would probably go far.
> The glutathione hepatic values in mice obtained by intraperitoneal injection of the ester are superimposable on controls and the oral LD50 was found to be greater than 2000 mg kg^-1 and the intraperitoneal LD50 was 1900 mg kg^-1 ...
That's for pyroglutamic and glutamic acid esters of paracetamol: https://pubmed.ncbi.nlm.nih.gov/8799871/
and more general analogs apparently can also be designed to not produce NAPQI:
> Thus, in 2020, N-sulpharyl-APAP prodrugs 39–40 (Fig. 11) were developed. [...] They are not hepatotoxic because they do not generate toxic metabolite NAPQI, even in concentrations equal to a toxic dose of APAP (600 mg kg^−1 in mice).
https://pubs.rsc.org/zh-tw/content/articlepdf/2024/ra/d4ra00... p. 9702.
These would probably require trials, though.
Say a particular finetune prioritizes profits right now and makes recommendations like "cut down on maintenance, you can make up for it later with your increased profits and their interest". It produces more profits, and wins the AB test. Later the chickens come home to roost.
You can reduce the problem by using long-term indicators, but then each AB test is very slow.
Biasing actual buying advice would be feasible, but it would have to be handled very carefully to not be too obvious.
[1] https://tmedweb.tulane.edu/pharmwiki/doku.php/acetaminophen
Virtual memory probably isn't what you meant, but take something like user privilege separation. It's usually considered a good idea to not run software as root. To interpret the statement generously, privilege separation does restrict immediate freedom: you have to escalate whenever you want to do system-level changes. But I think josephg's statement:
> Sandboxing gives users more control. Not less. Even if they use that control to turn off sandboxing, they still have more freedom because they get to decide if sandboxing is enabled or disabled.
can be directly transposed to user privilege separation. While it's true that escalating to root is more of a hassle than just running everything as root, in another sense it does provide more control because the user can run arbitrary code without being afraid that it will nuke their OS; and more freedom because you could always just run everything as root anyway.
Maybe josephg's sense of freedom and control is what you're saying there is a trade-off between. But the case of privilege separation shows that some trade-offs are such that they provide a lot of security for only a little bit of inconvenience, and that's a trade-off most people are willing to make.
Sometimes the trade-off may seem unacceptable because OS or software support isn't there yet. Like Vista's constant UAC annoyances in the case of privilege separation/escalation. But that doesn't mean that the fundamental idea of privilege levels is bad or that it must necessarily trade off too much convenience for control.
I think that's also what josephg is suggesting about sandboxing. He says that the clipboard problem could probably be fixed; then you say, "but there are other examples". What remains to be shown is whether the examples are inherent to sandboxing and must degrade a capabilities/sandbox approach to a level where the trade-off is unacceptable to most.
The service then links the token to your account and uses ordinary detection measures to see if you're spamming, flooding, phishing, whatever. If you do, the token gets blacklisted and you can no longer sign on to that service.
This isn't foolproof - you could still bribe random people on the street to be men/mules in the middle and do your flooding through them - but it's much harder than just spinning up ten thousand bots on a residential proxy.
[1] https://github.com/aw31/openai-imo-2025-proofs
[2] https://arxiv.org/pdf/2507.15855 Appendix A
Domain names are a good example, because as skissane said, you could just make another DNS root. The trouble is convincing people (browsers) to use it. The problem in attempting to overturn Facebook isn't mainly the coding, either, but having a critical mass care. Those barriers don't seem like absolutes the way land is; they're just very high, high enough for those who control them to extract economic rent.
In city planning and building design, the problem is even more severe. The planner doesn't know what people are going to settle where, what their desired needs are (or are going to be), and so on. That doesn't mean that there's no such thing as an awful solution, nor that you can't say anything at all. (A house probably needs windows, and you probably shouldn't stick a polluting industrial zone right next to a bunch of them.) It just means that trying to "micromanage" a city or complex building fails - for the same reason that micromanaging an organization fails.
(This is a requisite variety or "seeing like a state" argument.)
It seems to me that any "paper trail" scheme of the sort you describe would have to solve the problems of DRM to work: making the elements that report on the real world (in this case, the CCD) tamper-proof, making the encryption key impossible to extract, designing robust watermarks to avoid analog holes, etc.
Here's a paper by Markus Schulze proposing such a method: https://aso.icann.org/wp-content/uploads/2019/02/schulze4.pd... He uses some very large districts, but it should work for smaller districts too.