jwilber

Taking a contrarian stance on everything (exactly what you’re doing here) is not a signal.

Explaining why research seems sus (what the comment does) is some form (strength proportional to the explanation) of signal.

Very surprising that this article does not mention trucks, which in America have become larger, more common, and an ever-growing contribution to the rising number of traffic fatalities.

“ Pickups also tend to be more dangerous in collisions between differently sized vehicles — car drivers are 2.5 times more likely to die when colliding with a pickup as compared to another car” [0]

Pickup trucks’ weight increased by 32% between 1990 and 2021 [1]

As a bonus laughable fact, the first generation of F-150s was 36% cab and 64% bed by length. By 2021, the ratio flipped to 63% cab and 37% bed. So much for the “rugged” weekend warriors.

0: https://www.tandfonline.com/doi/full/10.1080/15389588.2019.1...

1: https://www.nature.com/articles/d41586-021-02760-8#ref-CR1

See another interactive article explaining linear regression and gradient descent: https://mlu-explain.github.io/linear-regression/

This tired take is in every thread now. The sort of behavior better served by a Reddit bot, and just as annoying.

Housing price examples in regression are much, much older than Ng’s ML class.

A stat no doubt brought to us by genai-automated scraping.

FWIW, this post seems longer than most of OPs usual posts.

I’ll also add: as a longtime user of em-dashes, the constant low-effort dismissal of any writing using an em-dash as “must be genai!” is super annoying. So much so that I’ve made an effort to stop using them in my writing.

There’s some poetic irony in using genai to dismiss someone else’s work for perceived use of genai.

I agree that LLMs are hurting the general population’s capacity to think (assuming they use it often. I’ve certainly noticed a slight trend among students I’ve taught to use less effort, and myself to some extent).

I don’t agree that this will affect ML progress much, since the general population isn’t contributing to core ML research.

I’m making no judgment as to whether or not it’s smart. Just an observation on the state of affairs.

I do think a zeitgeist where it being openly done is applauded as smart is a sad state of the country.

Only 100 days on and we celebrate open crony capitalism as a smart business decision.

Hired left and right != interviewed left and right != interviewed quite a few at Crowdstrike.

Maybe you’re contributing to the narrative with the posts like above. It’ll certainly drive engagement.

The key is the “…over a certain size” solely benefiting the richest man in the world, who just so happens to be heavily involved (despite no election) in the very government setting the policy and determining the size.

LLM/state space models have been popular for some years now, see: https://arxiv.org/abs/2212.14052

More recently, hybrid architectures that utilize attention plus other operators are gaining traction.

See https://arxiv.org/abs/2503.01868

Any leading model. Ask it to generate the ui from the screenshot. Follow up and describe the required animations, and it will give you the css/js. Do it all the time.

This is just selling components someone else wrote, right?

Five years ago I could understand the appeal and appreciate the effort required. Today, it’s a matter of seeing others work, taking a screenshot, asking ai to recreate it, and then packaging it into a “library” and selling it for $50.

Maybe I’m alone in the sentiment, but it just rubs me the wrong way.

Amazing work. Motherduck and the duckdb ecosystem have done a great job of gathering talented engineers with great taste. Craftsmanship may be the word I’m looking for - I always look forward to their releases.

I spent the first two quarters of 2024 working on observability for a build-the-plane-as-you-fly-it style project. I can’t express how useful the cte preview would have been for debugging.

You’re conflating social taboos with societal taboos.

See also My War Gone By, I Miss It So - a book about the author’s time as a war correspondent in Bosnia during the Yugoslavian War. After the war, he became addicted to heroin for lack of the excitement that the war brought.

https://septemberpublishing.org/product/my-war-gone-by-i-mis...

There’s an interactive visual of Entropy here in the Where To Partition section (midway thru the article): https://mlu-explain.github.io/decision-tree/

The implication is they're all in the same club. And if you know any family members of members of Congress, you know this to be true!

Maybe it's a friend of Trump. Maybe it's a friend of Pelosi. Might even be a member of Congress!

"Rules for thee not for me."

Can you provide an example? I’ve been using d3 on-and-off for almost 10 years, and the core docs have just gone from README’s to static-hosted pages with the same (version-specific) content (eg https://d3js.org/d3-selection/selecting). What is significantly worse here between the two? To my eyes, they’re basically identical: https://github.com/d3/d3-selection/tree/86

My best guess is that you’re referring to bl.ocks.org, which hosted d3 examples as standalone html examples, but this was independent from the d3 docs.

Anecdotally, I get dramatically better sleep after good cardio. Weightlifting doesn’t seem to have much of an effect.

"Observable is obnoxious if you want to add a D3 pie chart to your Vue application and have to untangle calls to D3’s API from reactive cell values, which look like ordinary JavaScript, but are not, and will cause compilation and runtime errors when copied."

Yep - as I wrote: "If you want to just blindly copy and paste d3 code, you may have issues with the docs being hosted on observable."

If instead you learn the core api (selections, bindings, and update patterns), any chart is really just a matter of using the correct layout/generator.

Here's d3 pie: https://d3js.org/d3-shape/pie

No need for going into Observable at all, as the pie generator is completely documented.

But if, as you mentioned, you prefer to copy the code from Observable, the linked pie example is extremely simple to do so: https://observablehq.com/@d3/donut-chart/2

Zero reactive cells, you can literally just copy the code inside the cell statement (e.g. the curly braces).

What makes you think this?

Separate, possibly unrelated question: what was trump’s goal when he created and rug pulled the Trump coin?

"You can use as many italics as you want but I don’t think it’s a wildly bold claim and your counter example is simply silly."

I don't understand the relevance of any of this, but I think I've done a fair job outlining my points above. Let me give my best summary: the execution environment used by a library's (example) documentation is independent from learning the API of the library itself. I agree that newcomers to JavaScript may find plenty of confusion there, and I'm sure a decent chunk of D3's users may be new to web-programming in general, but it's not the job of D3 maintainers to account for that.

I actually think our back-and-forth is a perfect example of why open-source is so painful to work on.

D3 is one of the best documented libraries out there is. There are multiple books, hundreds of hours of youtube videos, and most importantly, dedicated maintainers (Mike Bostock, Philippe Riviere, etc.) who've poured hours into additional sources of documentation and are very responsive on GitHub issues.

The unfortunate outcome here is that users have come to expect this sort of high-quality support and documentation (again with the italics, who does this guy think he is???). Every D3 submodule has standard api documentation, sure, but thats expected of all libraries these days. However, the additional example documentation (again, nobody got paid to make this material) for the most recent releases has been migrated from bl.ocks.org (a now defunct open-source service users didn't pay for it) to Observable Notebooks. Now, Observable is a VC-backed business, yes, but the documentation is still completely free. In this thread, you mention you don't like this, so in at least one online conversation where D3 comes up, you actively advocate against using it out of principle! (Couldn't resist!).

Of course this is just my viewpoint on what has transpired, and I'm likely articulating it in a more-inflammatory-than-reality manner. But I'd prefer to have D3 documentation in the form of free, interactive Observable Notebooks rather than to have no documentation at all. Even more so if it helps out the authors of the open-source library.

As a tip for those who have difficulties going from the reactive Observable model to vanilla js -> you can always just take an observable notebook and ask AI to convert it to a set of vanilla files for you.

PS - Sorry for the novel, at an airport!

This is a very bizarre take. If you want to just blindly copy and paste d3 code, you may have issues with the docs being hosted on observable. But absolutely zero part of the d3 api (or core design patterns of use) have anything to do with observable.

It’s like saying, “I refuse to use PyTorch because their docs are built with mkdocs.”

Moreover, even if the coupling were limiting (which, again, it is not), it’s odd to attack observable since everything they put out is fully open-source. It’s not like they’re hiding docs behind a paywall, they’re actively contributing to the viz ecosystem through basically everything they release (eg Observable Plot, Observable framework).

As far as echarts goes, it’s a great tool. The declarative syntax for charting always feels a bit odd, but it’s easy enough to wrap into component libraries. AFAIK it still powers many of the big BI viz tools used today (eg AWS Quicksight).

Convolutions are used in many non-image applications, including language (eg dilated convolutions have been popular for some time) and 1D cases. The paper I linked references the hyena operator, which is literally a convolution replacement for attention (though it’s often used in hybrid architectures like the one I linked).

There’s no one-size-fits-all answer here, but in my experience, for long contexts, perf for conv-based methods outperforms strictly attention-based methods. See evo2:

“With the current implementation of Evo2, we do not have the heavily optimized kernels in place for convolution operators like we do for attention layers in a model like llama2. Even with this shortcoming, we see that the benefit from including more convolutional layers makes up for the earlier stage of optimization at around the 64k context length. Beyond that point we see an improvement in performance even compared to a highly optimized transformer model.“

https://docs.nvidia.com/bionemo-framework/latest/models/evo2...

Achieved by “applying convolution operations over queries, keys and heads, allowing nearby queries and keys to affect each other's attention weights for more precise attention”

Cool to see convolutions making such a comeback lately in the llm world. See also the recent striped hyena2 architecture, which uses the conv-based hyena operator to great success:

https://arxiv.org/abs/2503.01868

You absolutely do pay to be on that list, definitely an anti-signal.