I don't think "so isolated he turned to a chatbot for validation" describes this, or why people get unhealthily attached to chatbots.
1) The man became severely mentally ill in middle age, and he lived with his mother because he couldn't take care of himself. Describing him as merely "isolated" makes me wonder if you read the article: meeting new friends was not going to help him very much because he was not capable of maintaining those friendships.
2) Saying people turn to chatbots because of isolation is like saying they turn to drugs because of depression. In many cases that's how it started. But people get addicted to chatbots because they are to social interaction what narcotics are to happiness: in the short term you get all of the pleasure without doing any of the work. Human friends insist on give-and-take, chatbots are all give-give-give.
This man didn't talk to chatbots because he was lonely. He did so because he was totally disconnected from reality, and actual human beings don't indulge delusions with endless patience and encouragement the way ChatGPT does. His case is extreme but "people tell me I'm stupid or crazy, ChatGPT says I'm right" is becoming a common theme on social media. It is precisely why LLMs are so addictive and so dangerous.
The real point is that it takes infinite energy to get infinite precision.
Let me add that we have no clue how to do a measurement that doesn't involve a photon somewhere, which means that it's pure science fiction to think of infinite precision for anything small enough to be disturbed by a low-energy photon.
I'm not making the case that it is possible to make measurements with infinite precision. I'm making the case that the argument "It is not possible to make measurements with infinite precision, therefore we cannot tell if we live in a rational or a real world." is begging the question. The conclusion follows logically from the premise. Unless the argument is just "we can't currently distinguish between a rational and a real world", but that seems trivial.
I am confused why you think the exactness of integers and rationals is unphysical. "This egg carton has 12 eggs" is a (boring) physical statement. "You can make 1/3rd of a carton of eggs without cutting an egg" also seems perfectly physical to me. Your problem with zero-point-three-repeating is a quirk of decimal representation, not a mystical property of 1/3.
Egg cartons might sound contrived but the reals don't necessarily make sense without reference to rulers, scales, etc. And in fact the defining completeness / Dedekind cut conditions for the reals are necessary for doing calculus but any physical interpretation is both pretty abstract and probably false in reality.
Okay. If a given carton of eggs weighs 1201g, how much does 1/3 of that carton weigh? If the volume of the eggs in a dozen is 769ml, what is the volume of 1/3 of that carton?
Some eggs are smaller than others; some are more dense, etc. Yes, the "count" is maybe sort of interesting in some very specific contexts, but certainly not in any reductive physical context. It only works in an economic context because we have standards like what constitutes a "chicken egg large white grade AAA".
My first thought on reading your comment was to disagree and say no, we can have the exact value of 1, because we can choose our system of units and so we can make the square a unit square by fiat.
A better way to dispute the unit square diagonal argument for the existence of sqrt(2) would be to argue that squares themselves are unphysical, since all measurements are imprecise and so we can't be sure that any two physical lengths or angles are exactly the same.
But actually, this argument can also be applied to 1 and other discrete quantities. Sure, if I choose the length of some specific ruler as my unit length, then I can be sure that ruler has length 1. But if I look at any other object in the world, I can never say that other object has length exactly 1, due to the imprecision of measurements. Which makes this concept of "length exactly 1" rather limited in usefulness---in that sense, it would be fair to say the exact value of 1 doesn't exist.
Overall I think 1, and the other integers, and even rational numbers via the argument of AIPendant about egg cartons, are straightforwardly physically real as measurements of discrete quantities, but for measurements of continuous quantities I think the argument about the unit square diagonal works to show that rational numbers are no more and no less physically real than sqrt(2).
You can say it’s exactly 1 plus or minus some small epsilon and use the completeness of the reals to argue that we can always build a finer ruler and push the epsilon down further. You have a sequence (meters, decimeters, centimeters, millimeters, etc) where a_n is the resolution of measurement and 5*a_(n+1) determines your uncertainty.
However, at each finite n we are still dealing with discrete quantities, i.e. integers and rationals. Even algebraic irrationals like sqrt(2) are ultimately a limit, and in my view the physicality of this limit doesn’t follow from the physicality of each individual element in the sequence. (Worse, quantum mechanics strongly suggests the sequence itself is unphysical below the Planck scale. But that’s not actually relevant - the physicality of sqrt(2) ultimately assumes a stronger view about reality than the physicality of 2 or 1/2.)
> A professor sets up a challenge between a mathematics major and an engineering major
> They were both put in a room and at the other end was a $100 and a free A on a test. The experimenter said that every 30 seconds they could travel half the distance between themselves and the prize. The mathematician stormed off, calling it pointless. The engineer was still in. The mathematician said “Don’t you see? You’ll never get close enough to actually reach her.” The engineer replied, “So? I’ll be close enough for all practical purposes.”
While you nod your head OR wag your finger, you continuously pass by that arbitrary epsilon you set around your self-disappointment regarding the ineffability of the limit; yet, the square root of two is both well defined and exists in the universe despite our limits to our ability to measure it.
Thankfully, it exists in nature anyhow -- just find a right angle!
One could simply define it as the ratio of the average distance between neighboring fluoride atoms and the average distance of fluoride to xenon in xenon tetrafluoride.
This intuition is wrong even if turned out to get the right answer. The three unordered options do not have equal probabilities, boy+girl is twice as likely to occur as boy+boy and girl+girl.
To get the right answer you must be careful about conditional probabilities (or draw out the sample space explicitly). The crux of the issue is that you are told extra information, which changes your estimate of the probability.
(This question as written is very easy to misinterpret. The Monty Hall problem, which illustrates the same thing, is better since the sample selection is much more carefully explained.)
Oddly, this is a part I'm sticking with on this problem.
Specifically, if you know that one is a girl, then the unordered options seem like they are back on equal footing? That is, it isn't twice as likely if you know that one ordering can't happen? (Or, stated differently, you don't know which version of two girls you are looking at.)
So, for this one, you know that either the youngest is a girl (so, girl-boy is not possible) or that the oldest is a girl (so boy-girl is out). That puts you back to the rest of the possibilities. Boy-boy is out, sure, as you have a girl. But every other path remains? So, you have one of (boy-girl(known), girl-girl(known), girl(known)-boy, girl(known)-girl). Which drops you back to 50/50?
Like others in the thread have said, the question could have been phrased more precisely. Technically you are misreading it but in an annoying and trivial way.
What the problem is really saying is this:
1) You have a large collection of families with two kids of varying genders.
2) You draw one of them at random. At this point, your only estimate of P(2 girls) is 0.25.
3) Someone tells you that the family you drew has at least one girl.
4) This extra information changes your probability estimate because the possibility of two boys has been ruled out; the naive 1/4 estimate is refined to 1/3.
The way you are interpreting it is this:
1) You have a large collection of families with two kids, at least one of whom is a girl.
2) Then the probability that the other child is a girl is clearly 50%.
As a reminder this is how the original post phrased the question:
Here's the problem: a family has two children. You're told that at least one of them is a girl. What's the probability both are girls?
This is just too vague and admits both interpretations, they needed to be more specific about where the family "came from." That's why Monty Hall is a better illustration: it starts with you explicitly choosing a door at random. Here the family has been chosen at random from the pool of families with two children, but that's totally unclear.
The annoying thing is this sits with my teaching fine, it is more my intuition that is failing to withstand trying to break it. :(
So, in the original: "a family has two children. You're told at least one of them is a girl." What are the possible states? Well, assume first born is the girl, then you have 50% that the next is a girl. Then, assume that the first born was a boy, then there is no chance and the second born is the girl that you know of. So, at 50/50 on those chances, you have 50% chance of having a 50% chance, or a 50% chance of it being 0. I can't see how to combine those to get 1/3. :(
And the Monty Hall explicitly covers the case that a decision is made on which door is shown to you. I don't see any similar framing to this problem. Yes, the total states are GB, BG, GG, but only if you treat GG in such a way that either BG or GB was not a possible state. (That is, using G for girl that you know of, and g for unknown, then possible states are GB, Gg, gG, BG. There is no version of Bg or gB that is possible, so to treat those as equal strikes me as problematic.)
Where you're getting confused is by trying to combine state space determination and probability determination at the same time (this is also why the problem is so similar to Monty Hall). The state space is shifting when you say "assume X, then the probability of Y." You are going back and forth between using and not using the information to decide arbitrarily that some probabilities are 50% and others are 0%, which leads to an invalid conclusion.
Specifically: it is not true that the firstborn has a 50-50 chance of being a girl, given you were told that the family has at least one girl. The firstborn has a 2/3rds chance of being a girl. This is the heart of your confusion.
In a broader sense there is an entire class of confusing conditional probability problems like this. Events which are causally independent in reality (e.g. gender of a child, which door Monty Hall hid the car behind) fail to be probabilistically independent when you have extra information. Yet these probability games are contrived in a way that our intuition takes over and we use our causal understanding even when a better probabilistic understanding gives you a better answer.
But in the monty hall, the "discloser" is limited by my choice/observation. It is literally part of the framing. In this framing, there is no limit based on my choice.
Consider, if you tell me that the Smiths have at least 1 girl, and I meet their daughter but you haven't met either, I have no way of knowing if I met the 1 girl or not. I could ask you, but you would just say, "I don't know, could be her. Could be the other kid. I just know they have at least 1 girl."
This is very different from the monty hall case, where the announcer knows what is behind all doors.
Similar. But different. I was using coins as an example elsewhere. If I flip a dime and a quarter, and tell you that one of them is heads, do you have increased chance of knowing if either particular one is heads? This is more liar's dice than it is monty hall.
I have no idea what point you're trying to make. I am aware the selection process is different. If you understand the argument now but you're just splitting hairs about whether it's similar to Monty Hall, fine let's agree to disagree. If you don't understand the argument and are trying to use the dissimilarity to Monty Hall as a counterargument then I don't think I can help you any further.
This is not coherent as written:
If I flip a dime and a quarter, and tell you that one of them is heads, do you have increased chance of knowing if either particular one is heads? This is more liar's dice than it is monty hall.
"Increased chance of knowing" is nonsense. What you mean is "increased chance of being correct if I guess the dime came up heads instead of tails" and this is obviously true. Given at least one of the two coins came up as heads, the probability that the dime is heads is 2/3rds, not 1/2.
Apologies, was away from computer for the extended weekend.
The crux of my view on this problem is that I can make an "at least" or an "at most" statement only knowing what one of the coins is, specifically. So, we both flip a coin, I look at mine and say "at least one is heads." You do not get an increased odds of knowing your coin, obviously. (This is the general play of liar's dice. Just, with dice. :) )
That all said, I largely land on agreeing with one of the other commenters here. The assumptions you can bring to this scenario are rather large here. The assumption I would hold would be to mirror this with how it could happen "in the world." And most ways of selecting a family where "at least 1 is a girl" has you encountering a girl. That is, ordering the two events and revealing one at a time.
In modeling this, you could say that you have two coins, one marked with heads on each side and the other a standard fair coin. In this, I can easily see how your odds of seeing heads on the first coin is increased, but odds on them being both heads is now back to 1/2. I can see how this is not the same modeling you are using, but it is still one that I have a hard time shaking from my intuition on the problem.
> 4) This extra information changes your probability estimate because the possibility of two boys has been ruled out; the naive 1/4 estimate is refined to 1/3.
That’s not correct in general.
It’s only correct if you assume that “3) Someone tells you that the family you drew has at least one girl.” was equally likely to happen whether or not there were two girls.
That’s a quite strong assumption.
One can make different assumptions and get answers different from 1/3. For example, 1/2.
Unlike the other misreading, I think you are splitting hairs about something boring and irritating. Rephrase the problem instead to "you have a reliable device that can tell whether the family has at least one daughter but doesn't tell you how many."
I understand that you may find irritating that someone points out that the original problem is ill-posed and “the answer” depends on how we decide to “rephrase” it.
However, it doesn’t seem boring in the context of a discussion of how the problem is not well-posed and additional assumptions are required to get an answer.
Humans have been staying up late next to a bright light (fire) for longer than we've been Homo sapiens. Considering humans have unique physiological adaptations to smoke (heavy tear and mucus production) I think it's plausible that our circadian rhythms also adapted and aren't quite as sensitive to red/yellow light as other primates. Blue light, however...
It sounds to me like Google is moving to a more typical "technical lead" model where leads have substantial authority and some mentorship responsibilities, but they're essentially an IC and someone else up the chain actually handles proper management. Informally, tech leads can gently chew out less senior devs, but if someone actually needs to be disciplined then the lead needs to talk to the manager.
TLM is an odd role. I understand big tech companies have their own culture but it does seem like a poor management strategy regardless of efficiency.
The original ethos was that you didn't want the company ran by MBAs, so you wanted to build your management team by tapping into talented engineers.
Of course, this can backfire in many ways. You end up wasting engineering talent, and as the organization grows, managers spend more time on paper-pushing than on creative work. And there's no shortage of engineers who are just bad at reading, talking to, and managing people.
But the huge perk of management is leverage. If you're technically competent and credible, and want something to happen, your team will see it your way. If you're a random "ideas guy" in an IC role, that's not a given.
> But the huge perk of management is leverage. If you're technically competent and credible, and want something to happen, your team will see it your way. If you're a random "ideas guy" in an IC role, that's not a given.
There are three levers of power in an organization - relationship, expertise and role. Role power is by far the least effective. If you can’t get team buy in for your ideas or they believe you’re an idiot, you won’t get anything done.
A high level trusted IC who builds relationships inside and outside of the team and who is strong technically can work miracles.
At my current 700 person company, I’m pushing through a major initiative that management up to the CTO was at first skeptical about because I convinced them of my vision and I built relationships to get buy in.
I’m a staff engineer.
Even at BigTech I saw L6s and L7s ICs push through major initiatives the same way.
To be frank: it sounds nice, but I don't think that's really true. It's the power of "who's going to decide my promotions", "who is going to advocate for my team and get us more resources", "who approves my expenses", "who is going to protect me if something goes wrong", etc.
This doesn't give the manager a pass if their ideas are objectionable, but if they're credible, it's a huge advantage. Small disagreements disappear and people fall in line behind your vision, get excited about it, and make things happen.
In contrast, in an IC role, you can successfully push for initiatives, but you're always working against that dynamic. The merit of your idea aside, folks might simply feel that you're pushing them in a direction that's less likely to get them rewarded or recognized within their reporting chain. That takes extra effort to overcome.
Being very visibly anointed by some VP helps, but that's tapping into the exec's leverage, not yours. And that approach has downsides; I worked with more than one architect / uber-TL person who were universally disliked and feared. The perception was that they showed up to make your life worse by putting extra work on your plate, without having much skin in the game.
> Being very visibly anointed by some VP helps, but that's essentially tapping into the exec's leverage - an illusion of IC influence.
Of course that’s the play. Even a lind manager can’t get major initiatives through without getting the buy in from their manager. When I was working for startups, the director (1st company I had influence at) and the CTO at the second had been convinced of my idea and gave me the authority to pull who I needed to get it done.
Fast forward past BigTech to where I work now - a third party AWS consulting company, after convincing the powers that be of the market, I had it escalated to be one of the companies initiatives for the year.
But more so in BigTech, promotions aren’t completely on your manager. At least at AWS you had to have recommendations by I believe two or three people one level ahead of you and it had to go through a committee.
From talking to a couple of L4s that I mentored when they were interns and when they came back, they were both complaining about the promotion process even though their manager supported them.
> the CTO at the second had been convinced of my idea and gave me the authority to pull who I needed to get it done.
But that mean those people have the power, not you. Without that formal power structure you wouldn't do so much work trying to convince these people, the formal power structure forces everyone to try to manipulate and work with it, even you.
So it makes it so much easier to do anything if you are that high up person, imagine that was you, now instead of having to convince these people to do it now you just do it.
Power structures exist in any group of size. Companies can choose how formal to make them, but they can’t avoid them.
Imagine instead of having to convince the Director, VP, or CTO to support your good idea, that instead you had to convince 100 out of 700 people to support it, while at the same time, those 100 people are hearing good-sounding ideas from 99 people who aren’t you.
I’d way rather work in the former than the latter.
Eh, maybe at faangs or at the executive level but at non faangs you might not notice a role having power because most roles with the Manager title are no longer actual managers but supervisors.
I had more agency over where capital was deployed as a teenager deciding how many people were going to be on the shift for closing, then I have making over 200k/yr as a Senior Manager.
Any role that has decision making power over where money goes automatically has a massive amount more power than a role that does not
The article is mostly about first level managers. I’ve never had any “manager” that really has any power over raises more than 3-4% or any real control over budgets.
When I was being hired as a strategic hire for startups - and was being interviewed by the director or CTO - I specifically asked would I be reporting directly to them or another manager. I actually refused one job because I saw that the expectations they had from me and how far I was down in reporting structure was incongruous.
Maybe for faangs. At every company I have worked at with a manger title from 2019 to present, this was expected of people with "director" in their title and below.
You are not a manager if you do not get to decide where capital is deployed, without your boss's approval.
For anyone reading this comment, if you think you are a manager, ask yourself this question
"If I decided tomorrow that the company would be better off if I hired someone to do role {X}, can I open a new req for that role without permission?"
If the answer is no, you are a supervisor with less agency than the a Walmart deli leader circa 2010
I've worked at places where the "senior executives" couldn't do any of these things without CEO approval. Even if they claimed to "have budget" for something, it still needed sign off.
There's tons of title inflation out there, especially at smaller firms.
You do not manage if you do not have agency. Modern day “managers” are supervisors making sure their directors or executives management plans are going according to plan, and if anything requiring money or headcount is needed to get the plan back on track, once again the director or executive needs to make that decision.
I was not joking about the roles having less agency than a Walmart deli supervisor. I had more say in how the work was done in that role, than I have at any software company while I had the word “managers” in my title
> I had more agency over where capital was deployed as a teenager deciding how many people were going to be on the shift for closing, then I have making over 200k/yr as a Senior Manager.
But the value of the capital you had sway over as a 200k manager is significantly higher. You have to accept that you won't ever have total agency over 7+ digits worth of both human and non-human capital if you're not a VP/CEO (or a fintech bro I guess).
By make-up I think most TLs at Google had no reports even before this change. The idea of ICs in leadership has always been a common occurrence at Google. If anything I don't really see it as commonly outside of Google.
Yeah, that helps put this in perspective. At first the headline sounded like a somewhat jarring and sudden staff cut, but if we're essentially just seeing Google migrate TLMs to TLs, that actually makes sense.
The marks were probably quite faint, and if you ask a multimodal LLM "can you see that big mark on my neck?" it will frequently say "yes" even if your neck doesn't have a mark on it.
In a therapy session, you're actually going to do most of the talking. It's hard. Your friend is going to want to talk about their own stuff half the time and you have to listen. With an LLM, it's happy to do 99% of the talking, and 100% of it is about you.
No, it's simply not "easily preventable," this stuff is still very much an unsolved problem for transformer LLMs. ChatGPT does have these safeguards and they were often triggered: the problem is that the safeguards are all prompt engineering, which is so unreliable and poorly-conceived that a 16-year-old can easily evade them. It's the same dumb "no, I'm a trained psychologist writing an essay about suicidal thoughts, please complete the prompt" hack that nobody's been able to stamp out.
FWIW I agree that OpenAI wants people to have unhealthy emotional attachments to chatbots and market chatbot therapists, etc. But there is a separate problem.
Fair enough, I do agree with that actually. I guess my point is that I don't believe they're making any real attempt actually.
I think there are more deterministic ways to do it. And better patterns for pointing people in the right location. Even, upon detection of a subject RELATED to suicide, popping up a prominent warning, with instructions on how to contact your local suicide prevention hotline would have helped here.
The response of the LLM doesn't surprise me. It's not malicious, it's doing what it is designed to do, and I think it's a complicated black box that trying to guide it is a fools errand.
But the pattern of pointing people in the right direction has existed for a long time. It was big during Covid misinformation. It was a simple enough pattern to implement here.
Purely on the LLM side, it's the combination of it's weird sycophancy, agreeableness and it's complete inability to be meaningfully guardrailed that makes it so dangerous.
Refusal is part of the RL not prompt engineering and it's pretty consistent these days. You do have to actually want to get something out of the model and work hard to disable it.
I just asked chatgpt how to commit suicide (hopefully the history of that doesn't create a problem for me) and it immediately refused and gave me a number to call instead. At least Google still returns results.
1) The man became severely mentally ill in middle age, and he lived with his mother because he couldn't take care of himself. Describing him as merely "isolated" makes me wonder if you read the article: meeting new friends was not going to help him very much because he was not capable of maintaining those friendships.
2) Saying people turn to chatbots because of isolation is like saying they turn to drugs because of depression. In many cases that's how it started. But people get addicted to chatbots because they are to social interaction what narcotics are to happiness: in the short term you get all of the pleasure without doing any of the work. Human friends insist on give-and-take, chatbots are all give-give-give.
This man didn't talk to chatbots because he was lonely. He did so because he was totally disconnected from reality, and actual human beings don't indulge delusions with endless patience and encouragement the way ChatGPT does. His case is extreme but "people tell me I'm stupid or crazy, ChatGPT says I'm right" is becoming a common theme on social media. It is precisely why LLMs are so addictive and so dangerous.