usgroup

It was very influential and had all kinds of interesting stories (e.g. https://en.wikipedia.org/wiki/Fifth_Generation_Computer_Syst...).

I'm not sure what qualifies as dead. Prolog is still around although as a small and specific community, perhaps comparable in size to the APL community at least within an order of magnitude.

I was almost sure that Prolog would be on the list, but apparently not.

I'm assuming:

Vendor LLM APIs + Software engineer = AI Engineer

Prolog seems cursed to be forgotten and re-discovered in a never ending cycle.

Theory is our best device for cultivating good judgement. My advise is to deeply invest in understand computer science and mathematics. Those are the foundations which will make it most likely to understand new application landscapes based on them.

Probability theory. It is closely aligned to my intuition now, but when I first learned it, it was difficult to accept beyond manipulating formulae.

Linearity aka most of linear algebra. Again, beyond manipulating formulae, many concepts eventually become intuitive with enough application, but its a hard won intuition to acquire.

x ~ Binomial(N,p) and you wish to estimate p.

Here are a whole collection methods for how to estimate p and calculate a confidence interval for it: https://en.wikipedia.org/wiki/Binomial_distribution#Confiden...

One of the methods is Bayesian; the rest are not.

Not mentioned in the list, but you can also use likelihood ratio intervals calculated from a likelihood profile: another Frequentist method.

None of the methods -- including the Bayesian, requires an informative prior.

>frequentism largely pretends that it doesn't exist. The ongoing replication crisis shows why this is not merely pedantry, but the single most urgent issue in science.

If you mean that Frequentist methods have no way of dealing with parameter uncertainty then your statement is false.

If you mean that some people who use Frequentist methods don't deal with parameter uncertainty then it may sometimes be the case.

It depends what the null hypothesis is here, but by construction, under a reasonable null, the p-value for an appropriate test would not be acceptable under a Frequentist framework.

I think Bayesian methods have made ground in sciences such as Sociology, Psychology and Ecology, which are mostly observational, but still attempt to make models with intepretable parameters.

With observational studies, representing confounders and uncertainty is a primary concern, because they are the most important source of defeater. Here, Bayesian software such as brms, Stan, pyMC, become a flexible way to integrate may sources of uncertainty. Although, I suspect methods like SEM still dominate for their use cases.

Personally, I find myself using Bayesian methods in a similar bag-of-tricks way that I use Frequentist methods mostly because its difficult to believe that complex phenomena is well described by either, so I use whatever makes the case best.

I consider myself an applied Statistician amongst other things, and I find this to be an ideological take mostly.

When we do Statistics, we are firstly doing Applied Mathematics, which we are secondly extending to account for uncertainty for our particular problem. Whether your final model is good will largely depend on how it serves the task it was built for and/or how likely its critics believe it is to be falsified in its alternative hypothesis space. That is, a particular uncertainty extension is not necessary nor sufficient.

For less usual examples, engineers may use Interval Arithmetic to deal with propagation uncertainty, quants might use maximin to hedge a portfolio, management science makes use of scenario analysis (deterministic models under different scenarios): all deal with uncertainty, none necessarily invoke either Frequentist or Bayesian intuitions.

So, in my opinion, the most useful thing to teach neophytes is how to model with Maths. Second, it is how to make cases for the model under uncertainty.

Statistical modelling is largely unrelated to machine learning in its ideology. If you're a professional Statistician then you're most likely working as part of some function heavily utilising randomised experiment design, or less frequently, observational designs. This would include hard sciences, actuarial sciences, finance (risk), manufacturing, poll/census research.

The main commercial opensource language for serious Statisticians is R. You can Google for the sorts of jobs requiring R as a marker, if you're interested in applications of Statistics unrelated to LLMs.

To answer your own question about classical ML, you can Google for jobs requiring the specific classical ML technologies in which you are interested as a marker.

I think this is accurate but mostly because statistical modelling aims for interpretable parameters. That very strongly regularises complexity.

I think the SWI Prolog clpBNR package is the most complete interval arithmetic system. It also supports arbitrary constraints.

https://github.com/ridgeworks/clpBNR

Interval/affine arithmetic are alternatives which do not make use of probabilities for this these kinds of calculations.

https://en.wikipedia.org/wiki/Interval_arithmetic

I think arbitrary distribution choice is dangerous. You're bound to end up using lots of quantities that are integers, or positive only (for example). "Confidence" will be very difficult to interpret.

Does it support constraints on solutions? E.g. A = 3~10, B = 4 - A, B > 0

The correct answer is duckdb with the vec extension :-)

NN computations -- in the main -- are high school Maths. Architectures are a topology governing chains of high school Maths. This is so much the case that the building blocks can neatly be expressed in a spreadsheet, which may or may not help some people establish more mechanical sympathy for what is going on.

Set theory, mathematical logic and proof makes sense to learn early on. Maybe even after strong command of arithmetic. These things are very portable to thinking generally. They'll change the way you see in a way you can't unlearn.

To my understanding, an LLM -- and similar models -- have a Markov chain equivalent.

There is an old argument from philosophy that any mechanical interpretation of mind has no need for consciousness. Or conversely, that consciousness is not needed explain any mechanistic aspect of mind.

Yet, consciousness -- sentience -- is our primary differentiator as humans.

From my perspective, we are making strides in processing natural language. We have made the startling discovery that language encodes a lot about thought patterns of the humans producing the text, and we now have machines which can effectively learn those patterns.

Yet, sentience remains no less a mystery.

I suggest searching job adverts by tech rather than title since OR tech is very specific to it.

If your aim is to evaluate an effect size of your treatment because you want to know whether it’s significant, you can’t do what the article advises.

I think this is what you need:

https://github.com/facebookarchive/planout

It’s just the bones of a factorial design framework for online experiments.

It’s simple enough to copy/paste and roll yourself. Many have translated it into their preferred language.

Yeah I think so, especially 2nd time around. You kind of have to refuse to think about your problem computationally; that’s really important when doing prolog I think, else you’ll end up with functional programming.

I try to think about what the solution of the problem implies , and then test each such interpretation against a prolog program to express it.

I used Haskell to solve AoC 2022, and in the midst of it I read lots comparisons to Lisp, which in turn turned me off Lisp. E.g. "why calculating is better than scheming".

I'd suppose this is because I have a strong bias to mathsy looking aesthetics.

I’ve been tempted in that direction too. Or using something like “Forth”. Both strike me as a “solve AoC with an abacus” style approaches, requiring bigger levels of problem understanding.

That sounds about right if your aim was to learn the language. I had the same experience with Prolog.

Just on this topic, would it be possible to make a whistle to do the same thing? I.e. crafted so it emits both ultrasound, and the audible counterpart which interferes with it to make the return audible?

Perhaps it could be such that the ultrasound warbles whilst interfering sound does not (or vice versa), which would make the sources easier to distinguish also.

See here: https://www.youtube.com/watch?v=PD3Y1l8XyUw

Another route would be to mix the ultrasound with another sound closer to the ear, then there is no need for an electronic ear at any point. The interference between sound can cause the inaudible frequencies to become audible.

Has anyone used Datalog with Datomic in anger? If so, what are your thoughts about Logica, and how does the proposition differ in your experience?

I may be misremembering but I think that at the time, Logica was the work of one developer who happened to be at Google. I'm not sure that there was an institutional push to use this language, nor that it has significant adoption at Google itself.