usgroup

Yeah totally agree. That the size of the tattoo or the number of them not increasing risk makes no sense. Somewhat like claiming whether you smoke a cigarette or 20 a day, the risk is the same. If the latter was true it would more likely indicate that there is some other commonality in that group increasing the risk.

Also the slicing and dicing, “11 more than the index year” and so on, is multiple hypothesis testing on the face of it; I wonder if they adjust for that.

Thank you. Your helpfulness and balanced evaluations are refreshing.

Does that imply that every possible branch is optimised separately and then convoluted thereafter?

If so, does it scale for very branchy programs?

Do you have any comparisons to a Gibbs based approach for any of the use case examples?

Noteworthy the difference an exact fitting makes over an approximation for a model like this. Fortran is amazing for this kind of exact work.

Not really. The world of Bayesian modelling has much fancier tools: Hamiltonian MC. See MC Stan. There’s also been Gibbs samplers and other techniques which support discrete decisions for donkeys years.

You can write down just about anything as a BUGS model for example, but “identifying the model” —- finding the uniquely best parameters, even though it’s a globally optimisation —- is often very difficult.

Gradient descent is significantly more limiting than that. Worth understanding MC. The old school is a high bar to jump.

I used to replace strict Boolean conditionals with sigmoids in order to achieve continuous transfer for Bayesian change-point models.

Does this do something similar or is it fancier?

I’m sorry to read about so much pain experienced by so many.

Do the people doing the traumatising typically know they are doing it? Do they typically deny it if confronted?

You should then be able to show that objective reality exists without recourse to subjective experience.

I’m all ears :-)

Yes, this section of Wikipedia is spot on in my opinion:

https://en.wikipedia.org/wiki/Occam's_razor#Controversial_as...

But we don't have an "outside-in" view? You have an inside + "guess the outside" view. Just because we can imagine an "outside-in" view, it doesn't mean we have it.

If there is an "objective" meaning to our lives which is without recourse to an inside-out view, then we have no access to it.

Occam's Razor, fewest entities/assumptions, etc is so problematic, because it is not invariant to being re-parameterised. I.e. under one description X is more complex than Y, and under another Y is more complex than X.

Here is Roger Penrose talking about the fine tuning of the Universe:

https://www.youtube.com/watch?v=yDqny7UzyR4

Many use these observations to argue for design. Say it became a scientific consensus that design was the strongest hypothesis, because it just kept making good predictions. Design implies purpose. Would life still be meaningless?

Is this all just a regress to "nothing is just as good as something therefore all life is meaningless"? I think that's where you have to end up to defend Nihilism.

I think that "life is objectively meaningless" is an unfounded statement. You should start by proving that to be true. I also think it preferentially loads words like "objective" and "existence". It also doesn't correspond to any common sense meaning of the word "meaning", since a servile dog's life is meaningful, a doctor's life is meaningful, and so on, in normal speech.

If you mean to say something like "all existence is objectively meaningless, and it is not objectively 'better' than nothingness", then it becomes another statement you need to usefully prove in order to go on with your argument.

Nihilism strikes me as a very specific commitment. There is a difference between "there is no meaning", and "we don't know the meaning", or a myriad other recall/precision related questions such as "can we know the meaning", "is one of the following candidates potentially the meaning", and so on.

I can see why for practical reasons some may lean into it, but I don't see it being epistemologically well founded.

> First, the contradiction thing. It's just.. An error/panic, why? Anyway, fine.

I suspect due to the details of the proof. This condition looks likely to me for very small set cardinality making the algorithm inappropriate for all-weather. See page 3 of the paper where Algorithm 2 is introduced. They show that in the failure condition, the likelihood of the algorithm returning a value outside of the epsilon bounds is higher.

> Then, there's the whole premise of 1..m: I'm still not sure if the size needs to be known upfront or not.

m sizes the threshold, if it is too small, the error bounds guaranteed by the algorithm will be larger than expected, and vice versa if m is too large.

> And btw, I didn't know about the Chernoff bounds and delta/epsilon are not explained at all in the paper, which added to the confusion a lot.

Papers typically do not spend words on basics and those are standard concepts.

> Perhaps there should be a method for estimating the confidence/error rate.

You can't resize the stream easily because p implicitly depends on m via the thresh cardinality condition and if you were to change m then your p updates would not correspond. As a result you may not be able to rely on the error bounds. Note though that stream doesn't mean infinite or very large: take it to mean one item at a time. The point of this algorithm is to bound space complexity to something small. Have a look at thresh and note that log2(1e50) is just 166: if you did have a very large stream of indeterminate length you could also just pick a very large m.

From the abstract: "All the current state-of-the-art algorithms are, however, beyond the reach of an undergraduate textbook owing to their reliance on the usage of notions such as pairwise independence and universal hash functions. We present a simple, intuitive, sampling-based space-efficient algorithm whose description and the proof are accessible to undergraduates with the knowledge of basic probability theory."

From the abstract: "All the current state-of-the-art algorithms are, however, beyond the reach of an undergraduate textbook owing to their reliance on the usage of notions such as pairwise independence and universal hash functions. We present a simple, intuitive, sampling-based space-efficient algorithm whose description and the proof are accessible to undergraduates with the knowledge of basic probability theory."

From the abstract: "... All the current state-of-the-art algorithms are, however, beyond the reach of an undergraduate textbook owing to their reliance on the usage of notions such as pairwise independence and universal hash functions. We present a simple, intuitive, sampling-based space-efficient algorithm whose description and the proof are accessible to undergraduates with the knowledge of basic probability theory ...."

I found the paper took about as long to read as the blog post and is more informative:

https://arxiv.org/pdf/2301.10191

It is about estimating the cardinality of a set of elements derived from a stream. The algorithm is so simple, you can code it and play with it whilst you read the paper.

The authors are explicit about the target audience and purpose for the algorithm: undergraduates and textbooks.

Yes, just like the algorithm that decides whether any given algorithm terminates could be "one algorithm" (if it was possible) :-)

In your statement, I'm not sure by which definition physically grounded things are a sufficient condition for being computable, but I think Penrose depends on the Turing notion of computability and the hardness of the halting problem.

We can of course move goal posts, redefine computability however we like, to get whatever conclusion we care for, but I think that Penrose effort is intellectually honest.

I think -- roughly speaking -- Penrose argues that human capabilities such as transductive reasoning are clearly not computable, therefore falsifying the idea that the mind reduces to an algorithm. He then goes on to propose how nonetheless what the mind does might be physically grounded, even if not in purely computational machinery.

Perhaps its where I live, or the people I know, but at my kids pre-school, I suspect that few 4 year olds could play naughts and crosses to a draw. I think that sort of awareness started around 5.5-6, where it became more normal.

Gobblet Gobblers -- on a cursory look -- seems to me like a complication on top of naughts and crosses. Namely, adding the ability to mask opponent pieces, and replace existing pieces.

As a side note it seems to me that one could replicate Gobblet Gobblers by using coloured coins of 3 sizes, with the smaller coins trumping the bigger ones thereby implying stacks.

Doubtless there are some four year olds to whom some people could teach chess. However, I'd say that learning chess, beyond how the pieces move, was an extraordinary level of competence for a 4 year old.

I focus on just using stuff with my kids. Things like piloting a remote control car is hard for a 4 year old, but they still want to do it. Playing Tetris on a slow speed, naughts and crosses, simplified chess, making objects out of paper, painting acorns, building train tracks, Lego, and so on.

You need engagement first, in order to cause learning, and I guess any process that causes both learning and engagement makes sense, but in my experience at young ages, that's more likely to be on the doing/using rather than creating side of the spectrum.

Truly. Immediately awash with nostalgia. Although I do feel its missing a spinning GIF, a visit counter and a visitors book :-)

Worth highlighting that make defines a DAG, and that you can run make tasks in parallel which will automatically bottleneck (as desired) on common dependencies, and fan out otherwise.

The sort of massive C++ build that make can handle are typically much more complicated than your average ETL pipeline. So, there is plenty of room to grow into make.

Shallow insight, but separating the algorithmic/representational and sqlite operational parts into own C projects, is possibly a good idea.

I'd expect the rate of evolution to be significantly different between the two parts, and if you are using an algorithm library adopted by others, you may get progress "for free".

A consequence of being an auto regressive model is not being able to plan token output. I think the author’s example is one of the many corollaries.

You could prompt the LLM differently , for example to write a Python program that does the random part, and then act on its output.

Yes precisely. If I had a pound for each time a dev told me they wish they were a doctor, and another pound for each time a doctor told me they wish they were something else!

I've many regrets similar in vein to that of your friend. Over time I've decided that it's rather reductionist to think that things connect together so simplistically. The "how my life could have been better if X, ceteris paribus" can be a really painful analysis, but -- in this case favourably -- there is no "ceteris paribus" in real lives.

In some sense life is obviously path dependent, since everything you've done is at least dependent on what didn't happen to you.

In another sense, there are many paths which lead to similar things, and path dependency can be Markovian, wherein the whole path doesn't matter in the first place, just the last bit.