Why aren't we training llms to load the data into R/pandas/polars/duckdb and then interrogate it iteratively that way?

It's how I do it. Why not our pet llm?

I had a lucky experience taking HS calculus the semester before as physics. I saw other students torturing themselves memorizing the formulae from the physics text and even then struggling to apply it to novel problems.

For the most part, knowing basic calc, it was possible to just draw a free body diagram and either integrate or take a derivative to get the answer. Didn't memorize much beyond f=ma and v=IR, for better or worse.

I still firmly believe that physics and calculus should be introduced together to provide a tangible and practical base to understand the mathematical theory.

Non-calculus based physics is violence.

It's common in European universities to not have "service courses" that the math departments provide to other departments. The other departments teach the math that is part of their field!

I'm not convinced that this is more than the market having a jump-scare at an incremental leap in model building technique.

Say DeepSeek has worked out how to do more with less - that's great! I don't think it means that the market for Nvidia's silicon (or anyone else who can hop the CUDA moat) is going to shrink. I think that the techniques for doing more with less.. will be applied to _more_ hardware, to make even bigger things. AI is in its infancy, and frankly has a long way to go. An efficiency multiplier at this stage of growth isn't going to reduce the capital needed, it will just make it go further. There may be some concern about scaling the amount of training material, but I don't see that as the end of the road at all. After all, a human's mental growth is hardly limited by the amount of available reading material. After all written material is trained upon, the next frontier will just be in some other mimicry of biological metacognition.

At least that's my armchair analysis.

I feel like diffusion algorithms are going to make headway into the procedural generation space soon.

With WFC you can have your rules for how a city is credibly composed of certain tile selections at some scales. But I'm imagining a future where a game designer can prompt a diffusion model with something like 'generate a mid size industrial city with a classy commercial district in the north and abandoned factories in the west, using the pre supplied positions of important buildings."

A diffusion model could be trained on the structure and semantics of a real city.

Great take. Wave function collapse is like a Markov chain in terms of complexity. As we know there's plenty of opportunities to apply more sophisticated techniques. The downside of course is that more complex techniques often require orders of magnitude more data. while I believe there's technical room for growth, I'm not sure there's practical room for growth or at least it's harder to imagine outside of "cool tech demo".

One thing I was waiting for the article to mention is that basic has a REPL.

I started on Atari basic in the 80s and having a REPL, even though I didn't know to call it that, was revelatory. Writing a line of code or a little for loop and drawing lines on the screen was a very tight learning experience.

Python approaches that with its REPL and notebooks but doesn't quite match the pedagogical ergonomics I enjoyed 35 years ago with that basic cartridge.

For beginners, python also has some confusing foot guns too, like understanding which objects pass by value and which by reference. It's all part of learning, but when a beginner is still trying to get a handle on variables and functions, it can confound the fundamentals.

Everything is by value.

Passing by reference means this is possible:

    def change_param(ab):
        ab = 768


    x = 12
    print(x)
    change_param(x)
    print(x)

Expected output in reference parameters, 12 followed by 768.

Otherwise it is always a value, even if the value happens to be a reference to an heap allocated object.

Python has no mechanism to change the contents of the x variable used as parameter on call site.

> even if the value happens to be a reference to an heap allocated object

IIRC in CPython, the value is always a reference to an object. Other implementations might pass immutable objects directly by value, but the semantics is the same as passing a reference to the object by value.

You are correct. (I'd be surprised if other implementations passed immutable objects by value, partly because representing an "object" coherently requires several words of memory and partly because of the importance of certain object identity semantics.) Python uses pass-by-value (if you prefer, pass-by-assignment) of variables which have reference semantics. Just like non-primitive types in Java, or class types in C#.

> I'd be surprised if other implementations passed immutable objects by value

I'm thinking of MicroPython here - it supports several pointer-tagging schemes, in which "references" might be actual pointers or might instead be immutable values stored directly.

For example, using the NaN-boxing scheme, MicroPython references might actually be inline `float`s, small `int`s, or constants like `None`, `True` or `False`.

You can do that if the parameter is a list:

def change(l):

     l[1] = 'x'

 >>> foo = [1,2,3]

 >>> foo

 [1, 2, 3]

 >>> changel(foo)

 >>> foo

 [1, 'x', 3]

This does a fundamentally different thing with similar syntax. Assignment in Python is fundamentally a rebinding operation; slice assignment rebinds a reference internal to the object, whereas ordinary assignment rebinds the name.

The classic test: in Python you cannot write a generic "swap function" - you could only possibly write a function to swap the internal states of the arguments, presuming them to be compatible.

You can't change which list `foo` refers to, though. Only the contents of that list.

`foo` is a reference to a list, and that reference is passed by value.

Now make foo point to another list, not change contents of the list foo points to.

Passing by reference requires changing the variable itself, used on the call site.

> like understanding which objects pass by value and which by reference

Isn't that true for almost every "mainstream" language?

Not discounting larger cultural forces and momentum at play, but I have to wonder how big a role inflated healthcare costs play in parents' level of protectiveness over their young children.

Even adjusted for inflation, a visit to the emergency room for a broken arm or a few stitches is a way different financial proposition today than it was in say 1988 when I was a pre-teen running amok, alone, within a 5 mile radius of home.

I know I have a dollar amount floating in the back of my head when I'm watching my kiddo cavort around in the playground or jumping between big rocks the same way that I did at his age. Is there mortal danger? Not at all.. Buuuuut, is he going to split his lip if he face plants on a leap? How many grand is that going to be, even with insurance? It figures into the mental calculus, even if we can ostensibly afford it. Lots of families probably cannot afford it so well. My family definitely wouldn't have been able to if health care costs were the same in the 80s as they are now.

Likely not at play since you see the decrease in free play across countries with public healthcare systems. As a Canadian I do not think about the cost of healthcare whatsoever. If my kid falls and breaks an arm, the cost is $15 which is one day of parking.

Broke my arm a couple years ago in the US - total cost for me out of pocket after insurance was around $5k.

> one day of parking

Unfortunately you really will be there the whole day.

Do we have issues with wait times, yes. However, the guy above you broke his arm and had to pay $5,000 AFTER insurance covered their portion.

I worked with a Canadian about 10 years ago that broke his arm while visiting family in Winnipeg during Christmas. He waited 2 or maybe 3 full days to have it set and get a cast with no luck in seeing a doctor and finally drove across the border to have it done in the US. He was 60 or so and when he didn't even seem upset at that happening! That really opened my eyes on "free" healthcare.

N of 1 stories don't really work as I could easily find a ton of terrible stories of why the private system in the states is bad for people.

The Canadian system (or any public system) gets a bad rap for wait times, but it's not like everyone is just walking around with open wounds and broken bones here. I work in a hospital, and if you come in and need an MRI urgently, you get it. If it can wait, you wait. We aren't fully public, and if you want a private MRI next day you can pay for it (and in fact, the last time I compared local prices vs prices in Montana, we were cheaper).

Another side effect of the healthcare cost is the liability for the owner/ maintainer of the play space as naturally parents would want to offload the cost of the treatment by seeking a lawsuit.

The news has gone to 11 with the fear mongering since the 1990s and fear mongering about your kids in danger is the most profitable.

FWIW, my kid always wanted to go to the playground when she was a youngster. When she hit around 11, she was no longer interested. Computer took over.

Sooooo.. how about venues just be required to provide affordable black unisex chess mumuu's for players as needed? Problem solved?

We can just make an NFT of each one to make sure they stay unique too!

This is cool! A long time ago I wanted to make something a little like this for my 20% project at Google/YouTube - a page that rained thumbnails of uploaded videos as soon as they became available.

Unfortunately, the idea was nixed since it had a pretty high chance of exposing ugly stuff that would otherwise have been lost in obscurity and never seen.

Do you think your idea has any more risk than sites like these:

- http://astronaut.io/

- https://obscuretube.com/

Absolutely 100% yes. Difference is these projects themselves are obscure. Opposed to an official Google branded service that will see significant publicity.

I regularly use duckdb on datasets of 1B+ rows, with nasty strong columns that may be over 10MB per value in the outliers. Mostly it just works, and fast too! When it doesn't, I'll usually just dump to parquet and hit it with sparksql, but that is the exception rather than the rule.