cubefox

Karma: 2,369

cubefox 17 Jun 2026 9:19 UTC
2 points
0
on: Two Classical Answers to “What do Two Variables Share?”
Can you make the relevance of this research more tangible by giving an informal example of Alice and Bob “observing” two different correlated variables, as well as a shared “latent variable”? I have difficulty imagining what those variables would model in practice.

cubefox 15 Jun 2026 13:49 UTC
2 points
0
in reply to: Richard_Kennaway’s comment on: A simple argument for trying less hard
This reminds me of this passage from Richard Jeffrey’s The Logic of Decision:

Socrates argues (Symposium, 200) that Love is not a god because to desire something is to be in want of it: you cannot desire what you already have. The point is also made in contemporary dictionaries. Thus, The Concise Oxford Dictionary defines a desire as an unsatisfied appetite, longing, wish, or craving. This account of the matter seems roughly right, although it needs some refinement. Thus, to vary the example of the Symposium, since people do not always know when they are loved, it is entirely possible to desire someone’s love when you already have it. Therefore, it seems better to say that you cannot desire what you think you have. Taking propositions as the objects of desire, the doctrine becomes: one who believes that a proposition is true cannot desire that it be true.

Which does suggest that both believing and wanting is related to reducing prediction error, because wanting something is wanting it to be true, and wanting something to be true is wanting to increase your degree of belief that it is true. So they are indeed related but they don’t seem hard to distinguish.

cubefox 14 Jun 2026 9:17 UTC
2 points
−2
in reply to: Seth Herd’s comment on: Philipreal’s Shortform
Some form of ID verification, or (less precisely) IP geo-blocking. This is already done on many websites for other purposes (copyright, local laws, sanctions etc).

cubefox 14 Jun 2026 9:03 UTC
2 points
0
in reply to: Morpheus’s comment on: leogao’s Shortform
I think an antonym is a polar opposite, not a negation. (hot --> cold rather than hot --> not hot)

cubefox 14 Jun 2026 6:28 UTC
4 points
0
in reply to: Mo Putera’s comment on: Mo Putera’s Shortform

This makes me wonder how the tribes of math will change as a result, as well as the relative status and prestige of subfields.

David Bessis has a long essay on the future of mathematics: The fall of the theorem economy. He argues convincingly that obtaining theorems was never a fundamental goal of mathematics. The real goal was always increasing human understanding of mathematical concepts. He argues that theorems are only valued so highly because they have been the most important (but imperfect) proxy for measuring contributions to mathematical understanding. You had to have deep understanding to come up with a proof. Bessis calls this monomaniacal focus on theorems “Hardy’s curse”, and the dynamic of theorems being the hard social currency of mathematics (citations, academic positions), the “theorem economy”. With the automation of proving theorems, they will lose most of their social prestige due to goodharting, and the theorem economy will break down, while leaving mathematics itself mostly untouched. The essay also touches on a lot of other interesting points.

cubefox 14 Jun 2026 5:47 UTC
2 points
0
in reply to: leogao’s comment on: leogao’s Shortform
Because of how it sounds? There is already “überall”/everywhere.

cubefox 13 Jun 2026 17:42 UTC
5 points
−6
on: Reward Hacking at the 1937 World’s Fair
The Apollo program was probably the clearest case of US American “reward hacking”. (Or the moon race between the US and USSR more generally.)

cubefox 13 Jun 2026 15:50 UTC
4 points
−2
in reply to: Seth Herd’s comment on: Philipreal’s Shortform
In the short term yes, in the medium term frontier AI companies, forced by the US government, can restrict access to frontier models to US Americans. Which would accelerate the US economy over the rest of the world, while the justification will be “national security”.

cubefox 13 Jun 2026 9:58 UTC
6 points
−4
in reply to: emanuelr’s comment on: Philipreal’s Shortform

I think that it’s good that a government body just demonstrated it’s willing to pull a frontier model offline

The government didn’t pull a frontier model offline, it cut off access for non-Americans.

cubefox 12 Jun 2026 17:47 UTC
3 points
1
on: The Skeptic, the Bayesian, Empiricism and Claims to Know:
Some form of “prior uncertainty” can be described with the Beta distribution. If we have a coin (or any other binary event) there is a difference between having observed 0 outcomes (0 heads and 0 tails), having observed one heads and one tails, and having observed 50 heads and 50 tails. All of them can be quantified as evidence for “50 probability for heads”, but the degree of reliability of this estimate varies widely between these cases. The Beta distribution is a subjective probability distribution over possible but unknown objective probability values of an event. If you are very sure the objective probability of your event is around 50%, the Beta distribution has a narrower peak than if you are more uncertain about your probability.

Of course, this doesn’t cover the more general types of prior uncertainty (e.g. rational vs irrational) you are discussing here, but at least it is something.

cubefox 12 Jun 2026 16:15 UTC
2 points
0
in reply to: RobertM’s comment on: Yoav Ravid’s Shortform
I’m a different person but I have a similar experience, I don’t use the “Your Feed > For You” section because it shows me subjectively random posts, though perhaps I could tune it to work better. Previously there was a “Top Comments” section on the front page which I personally found more useful.

cubefox 12 Jun 2026 13:47 UTC
4 points
0
in reply to: avturchin’s comment on: PSA: Almost nobody is directly working on superintelligent alignment
Another one: Nathan Young, Dear AGI,

cubefox 12 Jun 2026 10:59 UTC
3 points
0
in reply to: Shankar Sivarajan’s comment on: Shankar Sivarajan’s Shortform

Mine is that it’s indeed simply saying the government doesn’t want to work their products, even indirectly.

The first is very different from the second. “Indirectly” is an extremely wide concept, and very few companies are designated supply chain risks. Usually when deciding to buy or not buy some product, you don’t care at all which tools the company uses, or the company from which the first company buys tools, etc.

cubefox 11 Jun 2026 20:13 UTC
2 points
0
in reply to: tailcalled’s comment on: tailcalled’s Shortform
I assume there is a reason you don’t want New Foundations?

cubefox 11 Jun 2026 18:11 UTC
2 points
0
in reply to: tailcalled’s comment on: tailcalled’s Shortform
I thought usually ∈ is assumed to be acyclic, then one could identify a set with its ∈ DAG. But if you have a universal set it probably must be an element of itself, so we don’t even have a DAG. (I recently learned that the classical definition of “acyclic” only rules out finite cycles, so even a DAG allows for infinite cycles? Strange.)

So identify a set with its directed ∈ graph? Not sure whether this makes sense.

More to the point, this SEP article might be interesting.

cubefox 10 Jun 2026 13:26 UTC
2 points
0
on: Logits as a new monitor for evaluation awareness
This is cool research. But I want to emphasize that the usage of AUROC for the evaluation of any binary classifier is generally questionable because a high AUROC value only indicates 1) a large true positive rate and 2) a large true negative rate (= low false positive rate).

But a high value of an appropriate statistic for a binary predictor should maximize all four of these values:
1. P(actually true | predicted true) (= true positive rate, recall, sensitivity)
2. P(actually false | predicted false) (= true negative rate, specificity, converse of false positive rate)
3. P(predicted true | actually true) (= positive predictive value, precision)
4. P(predicted false | actually false) (= negative predictive value)
A high value of the AUROC only means 1 and 2 are large, while 3 and 4 could be very small. In which case we would clearly have a bad classifier despite a high AUROC value. For details see this paper.

There is another statistic which can be used instead of the AUROC while avoiding this problem: the phi coefficient / “MCC”. This is simply the binary version of the standard Pearson correlation.

A high value of the MCC (close to +1) indicates that all four of the above probabilities are large, and a high negative value (close to −1) that all four are small. If the predictor and the measured variable are independent (the classifier guesses randomly) the value of the MCC is 0.

The linked paper above goes so far as to say

In this short study, we explain why the Matthews correlation coefficient should replace the ROC AUC as standard statistic in all the scientific studies involving a binary classification, in all scientific fields.

cubefox 10 Jun 2026 8:00 UTC
2 points
0
in reply to: Mateusz Bagiński’s comment on: niplav’s Shortform

I consider it coherent to have qualia and be unaware of the fact, insofar as we’re taking qualia to be a good concept.

This would only make sense if they misunderstand the word “qualia”. Which seems quite unlikely for Carl Zimmerman who said he studied philosophy at MIT.

cubefox 9 Jun 2026 11:34 UTC
2 points
0
in reply to: Mo Putera’s comment on: Mo Putera’s Shortform
There is a difference between solving intent alignment for instruction following, and full value alignment. The latter would plausiblybbe guaranteed to be “safe”, or even more optimal than merely safe. (Utopia etc.)

cubefox 8 Jun 2026 21:57 UTC
4 points
0
in reply to: niplav’s comment on: niplav’s Shortform
Related: The Zombie Preacher of Somerset

cubefox 7 Jun 2026 5:20 UTC
7 points
0
in reply to: papetoast’s comment on: papetoast’s low quality shortforms
So AI papers are currently good enough that they can’t be trivially distinguished from human papers, making Pangram necessary, but not yet good enough to produce AI research that is at least on a human level. From the outside this looks like a sign that RSI fairly close now.

Tangentially, it’s somewhat interesting that Pangram is a twist on Turing’s original test: In the original, it was a human who had to distinguish between a human and an AI based on text, now it is an AI that distinguishes between both, since AIs are apparently better now than humans in distinguishing between humans and AIs. So Pangram is a CAPTCHA, but conventional captchas weren’t better than humans at distinguishing between AIs and humans.