feral

I don't think I'd go this far.

In the EMH people do still trade, it's just that only those with new information trade profitably.

In the joke scenario, as a parallel, you just need the economist to model that there's a chance they are the first person to see the bill.

Then you have a game theory game, where your chance of being the first person to see the bill depends on what everyone else thinks their chance was to be the first person, and hence did or didn't pick up the bill.

Then there is some optimal mixed strategy where people try pick up the bill with a certain probability, and it all works out.

The Wikipedia page on three body problem shows many periodic patterns.

There appear to be stable configurations for the 4 body problem too, even if the 4 masses are similar.

Its not fair to say there's no periodic pattern possible with 4 suns, then, even though most initial configurations might lead to chaotic behavior.

I think "reach" here is used in the context of: "the extent or range of something's application, effect, or influence"

So this article is talking about "to what extent did Irish culture/language/poetry influence early English poetry.

Why might we think there could be an influence there? I'm Irish, but no expert of any sort on history, so going to speculate here, but there's a common story in Ireland that Christianity came to Ireland late, just before the 'dark ages' in Europe, and that Christian culture and monasteries thrived in Ireland, and then Irish monks later carried this learning back into Europe. You can think of it a bit like the Foundation scifi stories!

So maybe that's a factor here.

Zafka seems to be recommending "The Design of Everyday Things" here, from that link which appears mangled to me. (looks like there are many books with similar names!)

This is what I do, too.

I'm not a big believer in making detailed notes during a 1:1

When necessary, I capture key points and action items in my phone while walking.

Not entirely the same, but since the pandemic and WFH, I've started doing a lot of my 1:1s while walking, voice only. I'd be really interested to know if anyone else has started doing this?

Typically, both people walk, using good Bluetooth headphones.

I was finding the combination of staring at the screen while coding, and then also during meetings, to be very straining.

Something about having to focus intently on the screen during a 1:1 was particularly tiring. Maybe because it's socially rude to look away while the other person is talking, so you end up really staring at the screen?

Now, instead, during a walking meeting, I exercise and get to look into the distance.

The most surprising thing for me was the ability to have technical conversations like this. I expected we'd really suffer without a whiteboard. Sometimes this is true, but I find we can have a lot of very technical conversations with voice only, and actually it's often easier than with video chat - because with voice only, I can focus more on the technical issues being discussed.

Finally, there are some days where I have a couple of hours walking meetings scattered throughout the day, and this is a lot of extra movement, which has to be good.

I wish there were more employers experimenting with different communication modalities like this; I feel the pandemic for all it's costs, has shown there are some great opportunities for other ways of working which have been underexplored.

artificial neural networks aren't like actual neural networks or brains

Just to zoom right in on neural networks:

People often say this, and I never see a solid argument.

I know very little about biological neural networks.

Clearly they are very different in some respects, for example, meat vs silicon.

But I never see a good argument that there's no perspective from which the computational structure is similar.

Yes, the low level structure, and the optimization is different, but so? You can run quicksort on a computer made of water and wood, or vaccum tubes, or transistors, and it's still quicksort.

Are we sure there aren't similarities in terms of how the various neural networks process information? I would be interested in argument for this claim.

After all, the artificial neural networks are achieving useful high level functionality, like recognizing shapes.

Yes, however we are reasoning probabilistically here.

We're not making absolute statements about either new or old things.

We're not saying any one new thing definitely won't last another 10 years, the same way we're not saying any one old thing will definitely last another 10 years.

The whole discussion is about what things are likely to do.

There's a place for formal logic in discussions of probabilistic models, but it can also confuse people.

Believing that any rocker that is new won't make it is wrong, in fact you can be certain that some will.

If you were made to bet money repeatedly on whether a new rocker or an old rocker would still be a rocker in 5 years, you would end up with more money if you always picked the old rocker. This means that older rockers last relatively longer and new rockers last relatively shorter ("on average" / "by expectation" / "probabilistically"); that's literally the same statement.

That the underlying mechanism might be gradual removal doesn't change this, and doesn't matter.

I salute you, Thaumasiotes, great explanation.

Does no rain implies streets are dry?

It would, if the only thing that could make the streets wet is rain; ie if we were in a limited model where that was true.

We're talking in the context of a limited model in this thread. My point was that the statements are equivalent in this model.

Alternatively:

Imagine we were talking about the size of a pizza and how big a dinner it will lead to.

A blog says "wider diameter pizzas generally lead to bigger meals. Narrower diameter pizzas generally lead to smaller meals".

The first comment here says "just because wider pizzas implies bigger means doesn't mean narrower pizzas implies smaller meals, the blog has made a logical error!".

I reply "in the context of roughly round pizzas the two statements are the same".

You then give the example about the rain and streets. Hopefully it's clear that your general logical point isn't relevant to the pizza discussion, where there is a relationship between width and area.

Similarly, in the Lindy model, which we are actually discussing, there is a relationship between age and expected lifetime. That relationship is more complex than the pizza one, which confuses things, but hopefully the pizza example makes it clear why there isn't a general logical error here.

go "wow, I find this good and I think it will keep being around and good for a while".

Sure, you can have a much richer model of the world than the Lindy model provides.

fact that it is young doesn't mean it is doomed, or even that it is not good. The Lindy effect says nothing in that direction.

No; it says that young things are less likely to last than old things. If the average expectancy is small, sure it doesn't guarantee any one thing is doomed (some aren't) but it absolutely does tell you that most of them aren't going to last long.

Imagine a friend tells you that they know someone who is planning to become a professional rock star when they leave school.

They are probably not going to make it, because most people who try don't. You can't be sure, because some people do become rockstars. But it's not an error to be sceptical of their chances.

If you hear they are still gigging after 3 years, even if they haven't made it yet, you are probably a tiny bit less sceptical.

That all makes sense, right?

I mean, the 'initial statement' we're talking about here was my hypothetical example, right?; just pretend I decided not to leave out "assuming constant speed" after I first wrote it, and we're good? Unless you are referring to a different statement in which case we're all mixed up and have exceeded the carrying capacity of HN threads :)

saying Pareto distributions follow linear relationships

No; not sure where you got that from.

For a product at T = 0, why would it be optimal to assume its lifespan would be T = 0 as well, rather than the average age of all products?

I'm not saying it would be. (The original blog might, but that's irrelevant to my posts here.)

If you knew the average lifespan of all products, that would be your best estimate of the lifespan of a new product. (You can't use average age naively without thinking about right censoring.)

If the average lifespan of a product was short, then the average lifespan of a new product would be short. As the product aged, it's expected lifespan would increase.

I.e. the longer something has been around the longer it will be around, or, equivalently, the shorter it has been around the shorter it will be around (both obviously taking about expected times.) Make sense now?

But it does, if you assume constant speed. (I almost wrote that in my comment but decided it was too obviously implied to mention.)

So if we assume that's implicit in my example, how does the Wikipedia page on the contrapositive help now? (It doesn't, as my argument isn't relying on any property of it.)

Absolutely there are real situations where the new stuff is going to last longer than the old stuff, even the old stuff that's been through a selection process. E.g. modern manufacturing techniques have improved overall longevity of all 2022 models.

But I think you are outside the Lindy effect model at that point.

To put this in the example of the original post: The author says visual studio code is expected to last less time than VIM.

You could counter by saying: "hey, maybe, uh, the rise of Product Management as a discipline has meant that modern software overall will have longer lifetimes, and hence it's not fair to guess that VIM will outlive VSCode".

And that'd be a fine position. But imo the right way to frame that isn't "the author did an incorrect logical inversion of the Lindy effect model"; rather it would be "I don't think the Lindy effect model applies to this domain".

(No one is saying the Lindy model is universal.)

I guess you could say you want to apply it only within a given year of software; so, we're happy to look backwards and apply the Lindy model to software written in 2011, but we've no idea how to think about the lifespan of software written in 2022, and aren't allowed make any inferences from software written before 2022.

That's fine, but that's an additional constraint we've added, is outside the Lindy model, and, really, we're in "all models are wrong, some are useful" territory here, where I'd ask "is it really useful to throw away all that previous data? Wouldn't it be a better starting point to use the lifetimes of previous years as at least a prior?" And if you grant that, then I think theres no logical error here.

Ok, so let's say discuss your formulation instead:

Why can't we 'invert' it, just because it's statistical effect? Yes, in your formulation, some of the new things in the pool will go on to live a long time, while others will be selected out.

But so what? We are talking about the expected lifetime of an item in the pool, conditioned only on it's age. There's no fundamental problem making a statement that this expected lifetime is short for new things, even if some fraction of those new things will last a long time, right?

After all, we don't know any one individual item that's been around a long time will last a lot longer. We only know we expect it to. Because even long lived items have finite lifetime, hence they'll eventually die (and when they do it'll be really surprising, because they've been around so long; but it will happen eventually.)

And so the statement is always talking about expected lifetime, whether for items that have already lasted a long or short time.

(Hence I still don't think there's really any logical 'inversion' here.)

Disagree: even in your formulation, no 'quick jump' is needed.

Your statement A:"The longer something has been around, the longer it will be around" is not the opposite of B:"the shorter something has been around, the shorter it will be around".

Rather they mean the same thing. 'longer' and 'shorter' here are just English language ways of referring to the same time t that an object has been around.

If someone tells you "the longer a distance is, the more time it takes to walk it", that is exactly the same as "the shorter a distance is, the less time it takes to walk it"; there's no logical leap there.

I could conceive a rule that says "archeological artefacts are likely to be around for a long time" and it'd be a mistake to conclude that this means that non-archelogical artefacts will only be around for a short time.

But that doesn't seem to be how the Lindy effect is formulated, either on Wikipedia or on your post, so there doesn't seem to be an error in applying it to new things.

The original comment was: "So Artificial Intelligence is a superset of Machine Learning. What are some AI algorithms that are still in use, that is not Machine Learning"

It seems we agree with Wikipedia that ML contains RL?

It's true RL is also studied in other fields.

I struggle to see how that means that RL is a good answer to "What are some AI algorithms that are still in use, that is not Machine Learning".

Hmm, I don't see that.

In the spirit of the cutting edge, any chance you could give me a chain-of-reasoning on that inference?

RL is generally considered as a type of ML.

Eg Wikipedia: "Reinforcement learning (RL) is an area of machine learning " https://en.m.wikipedia.org/wiki/Reinforcement_learning

I think this is the idea you are looking for: https://en.m.wikipedia.org/wiki/Kolmogorov_complexity

I think there's lots of ways to formulate this to see how it's percolation. I'm not a percolation expert though so please call me out if you are, and I'm wrong:

To sketch here:

Consider R the reproductive number of a disease (or R_e, the effective reproductive number, to be specific). This depends on both the inherent contagiousness of the disease, but also on the behavior of the population. If people choose to have fewer contacts (or bars are closed) then R decreases.

Let's say covid is spreading. We ask people to limit their contacts, and see if this stops spread.

We can think of this as trying to remove edges from the contact graph that the disease spreads on.

The contact graph becoming connected or unconnected as we remove edges is clearly percolation.

(Subject to some modeling assumptions about edges being removed at random, but these assumptions are common to both Erdos renyi graph models and SIR style compartmental models).

Make sense?

What I mean is that R varies by disease (or for a given disease, according to interventions, etc).

The size of the total epidemic varies sharply as R is changed in this way.

And, to give an example we've probably all become familiar with, this this is equivalent to how a disease becomes an epidemic as the R value crosses 1.

This is also broadly true in Ireland, (and I presume the UK). Among non-canoeists, 'canoe' can refer to a canoe or kayak, and people refer to "Canadian canoes" when they want to specify an open legged canoe.

(Although sit on top kayaks are usually called as such, confusingly).

where a specific employee would boost profits by $100k

This isn't a good model. Boosting your company's profit by 100k doesn't mean they can boost any company's profit by 100k. If they could, they'd capture a lot more of the 100k.

May be missing some subtlety, but at a quick read I would respond to this as:

- The author has not deeply understood the subjective interpretation of probability, as widely used by bayesians, and applicable to this sort of reasoning

- the author, and the sources the author quotes based on a quick read, seem unfamiliar with quanifying uncertainties not just as point estimates, but instead by using a probability distribution, which you can use to easily tell the difference between a mean probability of .5 based on say 5 heads and 5 tails, vs one based on 500 heads and 500 tails.

The point that it's often difficult to tell whether a particular probability should be 0.0001 or 0.00001 is a fair one.

But overall I don't take much reassurance from this article.

no possible blind study,

Not sure I agree with this.

I am definitely not saying this should be done, but as a thought experiment:

Imagine you worked for Fitbit, and could choose to undercount steps slightly.

You have a cohort of users who will walk each day until they reach 10k steps. You undercount steps for a random subset so that they end up walking 10.5k.

You repeat this experiment so you make the folks with a 5k goal actually walk 5.5k, etc

This is a pretty good blinded study. The difference is slight enough that the users don't notice the extra increase in distance consciously.

(Obviously you have to take into account the risk of churning people from the service somehow.)

If you were to check all cause mortality somehow (look up public death records? Some other way?) That would seem like a pretty good study design.

Again, perhaps unethical to do this without consent; but not impossible; you could ask a cohort of users to opt in to this sort of study.

If the claim is that it's distant dependent, this would likely take you a lot closer to causality.

Firstly, that was 60 years ago.

Second, I think thats misleading. The Irish military did scenario analysis on this in case they were asked, and concluded it was hopeless, if that's what you mean.

do we mean that no-one is going to bomb or shoot people in the UK or Ireland to attempt to achieve a political objective

No, that's not what we mean.

War between the UK and the RoI, the actual states involved, is in fact pretty unthinkable.