Physicist and dabbler in writing fantasy/science fiction.
Ben
Karma: 2,488
A breif thought. For an LLM, presumably documents with dates on them all get fed in in some non-chronological order. The current date is preaumably given to the LLM in some kind of system prompt. (Is it?). But if it were not given the currnet date in this way it would not know if it was being run 3000 years in the future or if it was being run in 1965. (Less certain than 3000 years in future).
A human, even before you tell them todays date, has some kind of anchoring in time. The LLM is not like that, so the questions you are asking are intrinsicaly a little harder. Especially as some documents in the training data might not be dated. For example, a document that says ‘Obama is President’ with no date provided will influence the LLMs training and ‘world view’.
Not fully satisfting as an answer. If i were playing a game where you gave me newspaper cuttings for a fictional world with fictional events. Gave me the fictional worlds 14 month calaender, then asked me who was King of Threposia in nontis, and the current date, I would make many mistakes but i would not make that mistake—i would understand you want last nontis.
I think the phrase ‘Proof by lack of imagination’ is sometimes used to describe this (or a close cousin).
I am not understanding the mechanism here, could someone explain it please?
A consumer buys an air ticket. The airline makes a loss on this (or not much profit). That same consumer now has air miles on a frequent flyer card/account that they can use for perks. How does the airline make money subsequently? Does it require that the consumer used a credit card (instead of a debit card) to buy the ticket? Or does it require that the consumer uses the air miles in a specific way?
These are not clear to me. Perhaps from some combination of me (1) living in UK not USA, (2) using a debit card, not a credit card and (3) having only ever ‘used’ frequent flyer cards as a tiebreak where the airline has overbooked and needs to give someone a free upgrade, and as the cardholder I am first in that queue.
Thank you very much for sharing that paper! Its a really nicely written paper, I like their figures a lot.
I think you have slightly misunderstood the paper (either that or I am missing something). In the paper, I think they are abusing the word “spin”. Every single place the paper says “spin”, they don’t actually mean spin (as in, the intrinsic spin angular momentum of light), they actually mean direction. So, when reading the paper try and read it through a mental translator where “left handed spin” translates to “left propagating”.
The spin angular momentum of light is (for a plane wave in vacuum) controlled entirely by its polarization, either left handed circular polarization or right handed. Importantly, this polarization depends on the fact that their are 2 spatial dimensions that are orthogonal to the propegation direction, so that for example the electric field could be expressed as: E = (1, i, 0) in an (x, y, z) basis and z the propegation direction. (Similarly (1, -i, 0) for the other polarization with the opposite spin).
In this paper they define what they call the “left handed” and “right handed” operators in the unnumbered equation immediately under equation (10). However, these operators are NOT left hand polarized and right hand polarized light waves. The operators differ, not by the relative phase of orthogonal electric field components, but by the relative phase of the electric and magnetic fields. This means they are “left travelling” and “right travelling” (IE propagating left or right) light waves. They have confusingly chosen to call these terms “spin”, I think this is because the equation they have derived looks like a Dirac equation, and in the Dirac equation those terms are called spin. But they are not the actual spin angular momentum of the light, they are completely unrelated.
In short, they don’t actually consider real spin at all, they just rename “direction” to “spin”.
They say theyr are in full agreement with Stephen Barnet (option number (1) in my post), that Minkowski’s momentum is the canonical one (to be used in Heisenberg uncertainty type situations) and Abraham’s is the kinetic one (to be used in Newtonian recoil calculations).
I previously thought “Atomic Weapons Establishment” was like “Medical Establishment”. But no, the “Atomic Weapons Establishment” is a real organization with buildings that bear that name and employees and a logo and everything somewhere in London.
I was so surprised to hear that that I immediately googled to see if there was a building in Washington DC somewhere that was literally called “The Military Industrial Complex” (there is not).
If I told someone ‘I bet stockfish could beat you at Chess’ i think it is very unlikely they would demand that I provide the exact sequence of moves it would play.
I think the key differences are that (1) the adversarial nature of chess is a given (a company merger could or should be collabroative). (2) People know it is possible to ‘win’ chess. Forcing a stalemate is not easy. In naughts and crosses, getting a draw is pretty easy, doesn’t matter how smart the computer is, I can at least tie. For all I (or most people) know company mergers that become adversarial might look more like naughts and crosses than chess.
So, I think what people actually want, is not so much a sketch of how they will loose. But more a sketch of the facts that (1) it is adversarial situation and (2) it is very likely someone will loose. (Not a tie). At that point you are already pretty worried (50% loss chance) even if you think your enemy is no stronger than you.
I was once in a situation much like yours considering the same question.
Arguments I considered at the time (when considering if I should “dress up” more in search of a romantic partner):
Do I actually know how to dress better?
Would dressing up differently than normal make me feel like I was impersonating/deceiving, thereby feeling more insecure and countering any gains in dress with losses in confidence?
Would this contribute to acquiring the wrong type of romantic partner? In the sense that, my interests, priorities etc might be strongly correlated with my poor dress sense, and that therefore changing the dress alone might disproportionately help with partners that are a bad match.
My very small data sample is that, I didn’t change anything. Then, at a fancy dress party (where everyone was weirdly dressed and my costume had not been picked by me, but was part of a matching set with friends) I met someone and things went great. I don’t know what to take from this, maybe fancy dress parties (or other settings with “non normal clothes”, like a wedding) are good for people in your situation. At the very least, if your clothing choice is proving to be a barrier then events like this provide you with a good opportunity, to either solve the problem, or possibly help diagnose if you could benefit from different everyday clothes.
Of course, there is a strong chance the fancy dress aspect was coincidence.
This was a fun post. I liked the way the “how many layers deep” idea was foreshadowed and built up to.
I see you are mostly on substack now, so you probably won’t see this.
I was trying to think of a clean example of a many-layer deep interaction, and I think I have identified it in the way that my parents and their friends pay bills at a restaurant. (Obviously you are socially obligated to offer to pay, so you do. But they know that was a “forced move”, which means that they can’t take your offer to pay as a strong sign that you are genuinely happy to pay, so they don’t accept the offer. But, you know that they know all that, so you can see that them rejecting your offer is also a somewhat forced move on their side, so you don’t accept their rejection of your offer … ).
Perhaps I am mudding the waters too much. I agree with your logic, and with your conclusions. I agree you are better off taking the option where you serve one less person to increase the total payoff.
What I was trying to say in the original post is that for most things there is the thing itself, and the measurement of the thing. For example maybe noisy thermometers are off from the actual temperature by some random variance. Things feel slightly more suspicious for utility, because the measure of the thing kind of is the thing itself, the split between the actual value and the measured value feels less defensible.
I don’t think that is right.
If we ask 100 people which of two ice cream flavors they prefer, and get a 51⁄49 split, that does not at all imply that 99 people couldn’t tell the difference and picked randomly, with the 100th person uniquely able to tell the difference. What we have is only similar to that on a population level. Its not that their is one person who can tell the difference perfectly and many who can’t tell the difference at all, but instead many people who can all unreliably tell the difference a little bit.
You take a drug that reduces your chances of heart disease by 1%. You don’t get heart disease. You will almost certainly never know if you would have got it without that drug.
That is interesting. My guess would have been that you would learn fastest in jobs that are just a little above your current skill set. (Learn fastest does not equal ‘most happy’).
Although, your claim does seem to fit better with my lived experience.
While the argument itself is nonsense, I think it makes a lot of sense for people to say it.
Lets say they gave their real logic: “I can’t imagine the LLM has any self awareness, so I don’t see any reason to treat it kindly, especially when that inconveniences me”. This is a reasonable position given the state of LLMs, but if the other person says “Wouldn’t it be good to be kind just in case? A small inconvenience vs potentially causing suffering?” and suddenly the first person look like the bad guy.
They don’t want to look like the bad guy, but they still think the policy is dumb, so they lay a “minefield”. They bring up animal suffering or whatever so that there is a threat. “I think this policy is dumb, and if you accuse me of being evil as a result then I will accuse you of being evil back. Mutually assured destruction of status”.
This dynamic seems like the kind of thing that becomes stronger the less well you know someone. So, like, random person on Twitter whose real name you don’t know would bring this up, a close friend, family member or similar wouldn’t do this.
In a vacuum, that logic seems good.
But, I know of several chain cooperatives that are very large, with lots of shops, etc. I live in the UK if that changes things at all compared to the USA. The ones I know are John Lewis, Waitrose, and Co-Op. (This is sort of double counting as John Lewis owns Waitrose).
So, all but one of the Co-Ops I know about are huge companies, although that comes with an obvious selection bias. (The only small one I know about is a random coffee shop I went to in Bristol that had a sign informing customers that they were helping to support a “radical, collectivist movement” (or something like that), they had 5 types of milk, none of them from animals. This sort of thing is very Bristol.)
I don’t know how the US system works at all, and have only a shallow understanding of the UK one (mostly from watching Yes Minister), but I think in the US a lot of posts that in the UK would be civil service are instead political posts. For example, I think US politicians directly pick which ambassadors to send, which is not the UK system.
They are probably very different systems.
How many hour of fully focused time from a knowledge worker do we need for them to pay for themselves? Until we have answered this question, we cannot possibly answer the question of whether employers are getting enough hours.
The linked post seems to implicitly accept that the number of deep-focus hours that employers need from their employees is comparable to the number of hours they are contracted to work. (So, like, most of a 9am to 5pm day for example.)
But I think this is kind of backwards. It is more likely that the number of hours “spent at work” has already been chosen with some knowledge of human behavior, and that the employers are perfectly aware that the number of hours of deep concentration they are buying is a considerably lower number.
It would maybe create a bit of a free-rider problem, if you had a situation where valuable data could be acquired at a cost (by doing a survey or experiment, or having fast messengers to run to London from Waterloo), then you kind of get exploited by anyone who just copies you.
On the other hand, in the immediate (non-equilibrium) mode the “copy the best trader” strategy is helping the market make good predictions. If their really is someone out there who is making perfect predictions, then having a lot of money follow their lead is going to result in better predictions.
I am curious about why you felt the discussion about minorities was so derailing. Couldn’t you have just said “Yes, that is a problem as well. However, the thing I am working on is...”
If it is any consolation, I have never seen that specific discussion, but in situations I have seen that feel analogous most of the people in the audience are actually more sympathetic to your side than they might appear. Its not like anyone is going to interject with “Well, that was a pointless question.”
I teach lifesaving, and it is true that children can drown in even very shallow water. At the most basic mechanical level, you just need to be able to find a head position where they are lying down with both nose and mouth submerged for it to be possible. I once collected drowning data from a coroner’s office, and its really sad how many children drown when their mum gets a phone call during bath time.
I suspect that most people assume water depth is a good metric for safety, and that its actually quite a bad way of measuring it. If you have someone who can swim, then it doesn’t matter if the pond is ten-thousand leagues deep. If they become unable to swim (eg. they have an epileptic seizure, or hit their head and fall unconscious, or try and hold their breath underwater and fall unconscious, get into a playfight with a sibling and get pushed under the water, become entangled in a net, become confused, or are so surprised to fall in that they freeze, or accidentally breath in water and start panicking) then it doesn’t make that much difference if its only 30cm deep, that is still deep enough to immerse the face.
I would guess that the difference between being supervised, and not, is much much bigger than the difference between 10cm and 10,000km.
“Shallow enough to stand up”, while presumably somewhat important, is not an all-important break point, because most things that would disrupt your ability to swim (confusion, unconsciousness, injury, panic, entanglement, cold water shock) might also disrupt your ability to stand.
In this case it sounds like the child, after landing in the water, was so surprised/shocked by it that she froze, and didn’t put her legs down. (I would guess the claim afterwards that she couldn’t was rationalization, and that the real reason was she was stunned with surprise, just based on the sense that I can’t imagine the dress was really that restrictive).
I thought the same thing. But looking at it, its still mostly wrong, but it is slightly less crazy than it first sounds.
I compared the watts per square meter coming down from sunlight (about 1000 at sea level according to the top google hit) and compared it to the watts of an air con system, 3000 acordong to some google hit (in the long run it will only heat the outside by its power consumption, although in the short term the heat from your house will add more), then we see the ac is like another 3 square meters of sun light.
So if you live somewhere where the density of dwellings is low, say a detached house with garden, then 3 extra square meters is nothing compared the square meter-age you already cover. But if you live in a 20 story appartment building in a city centre surroudned by similar buildings, and everyone runs ac, then maybe the ‘dwellings per square meter’ will be high enough that the ac will be adding energy that is non-negligable compared to the solar energy. (If we took +15% as our ‘non negligable’ threshold then the critical density is about 0.05 dwellings per square meter. Meaning in 100 square meters we have 5 dwellings adding 15 effective sunlight meters.) So maybe in Singapore this actually matters a little.
It still seems weird to single out ac though. The heat dissipated by driving a car through the city is surely much larger.
One aspect of this that i think is potentialy significant is processes which ‘stick’ in a non-working state.
I have on several occasions gone on the quest to find paper or ink, only for the printer to spit out 100 pages that other people had queued up when that resource was depleted. Those other people all, presumably, knew their document had not printed but decided not to get ink or paper. (Maybe their print wasnt that important to them. Maybe the instruction ‘check paper in drawer B2’ on an unfamilar machine was intimidating).
As i brought the paper/ink to the printer and reflected on the time it had taken, i consoled myself that it was now fixed, not just for me but for those who followed. I had unstuck the process for everyone. However, the queued up documents of people who had already given up, peehaps days ago, plus my print put us back into low-ink territory.
Compare this to stairs. Imagine i put an annoying to move barrier in front of the stairs somewhere. The first person would arrive, and remove it, and then it would inconvenience zero additional people.
I am not sure if printers jam, break oe run out of ink more often than most machines. I am sure that when they do they stay down longer.
A related contributing factor. Every organisation i have been in discourages printing when it ‘isnt nesseaary’. This means that some of the people leaving the printer in the stuck position are doing so because they worry that if they start asking where the ink is kept someone might badger them about whether the print they are doing is nessesary. If people have a low level sense of guilt (perhaps enviromentalism related) when they are dojng somethjng they are less likely to ask for help in ways that draw attention.