HNHacker News
TopNewBestAskShowJobs

mistercow

11,087 karma · joined May 9, 2009

submissionscomments
mistercow··on A wonderful coincidence or an expected connection: why π² ≈ g
All coincidences involving the definition of units will also have this property. Once you’ve narrowed to that specific domain, invariance to change of units is completely uninformative.
mistercow··on A wonderful coincidence or an expected connection: why π² ≈ g
Can you explain what you’re taking issue with in the post, then? Because it specifically lays out how the historical relationship between the meter and the second does in fact involve pi^2 and the force of gravity on earth.

(Granted, from what I can tell, it’s waving away a few details. It was the toise which was based on the seconds pendulum, and then the meter was later defined to roughly fit half a toise.)

mistercow··on A wonderful coincidence or an expected connection: why π² ≈ g
Obviously it’s nitpicky. That’s what a quibble is. But I don’t think it’s unwarranted. How you reason your way to a conclusion is at least as important a lesson as the conclusion itself. And in this case, the part I quoted is a bad lesson.
mistercow··on A wonderful coincidence or an expected connection: why π² ≈ g
But that connection actually is a coincidence. From what I can tell, when they standardized the meter, they were specifically going for something close to half of a toise, which was the unit defined as two pendulum seconds. So they searched about for something that could be measured repeatably and land on something close to a power of ten multiple of their target unit. The relationship to a circle there doesn’t have anything to do with the pi^2 thing.
mistercow··on A wonderful coincidence or an expected connection: why π² ≈ g
What do you mean by “no mention”? The entire article is about why this is specifically due to how the meter was first defined.
mistercow··on A wonderful coincidence or an expected connection: why π² ≈ g
Aside from the fact that the post already explained what the actual historical connection is, your explanation requires some serious hand-waving about the mass of the Earth and the gravitational constant, neither of which were known when the meter was first defined.
mistercow··on A wonderful coincidence or an expected connection: why π² ≈ g
Ok, then by that thinking, you should find it really interesting that Earth escape velocity is almost exactly ϕ^4 miles per second.

In fact, adding exponents here objectively makes it less interesting, because it increases the search space for coincidences.

What makes the case in the post most interesting to me is that it looks at first glance like it must be a coincidence, and then it turns out not to be.

mistercow··on A wonderful coincidence or an expected connection: why π² ≈ g
Yeah, I read the post. What I’m saying is “this relationship vanishes when you change units, so it must not be a coincidence” is a bad way to check for non-coincidences in general.

For example, the speed of sound is almost exactly 3/4 cubits per millisecond. Why is it such a nice fraction? The magic disappears if you change units… (of course, I just spammed units at wolfram alpha until I found something mildly interesting).

mistercow··on A wonderful coincidence or an expected connection: why π² ≈ g
This is interesting, but I have to quibble with this:

> If you express this value in any other units, the magic immediately disappears. So, this is no coincidence

Ordinarily, this would be extremely indicative of a coincidence. If you’re looking for a heuristic for non-coincidences, “sticks around when you change units” is the one you want. This is just an unusual case where that heuristic fails.

mistercow··on Grace Hopper, Nvidia's Halfway APU
I think this is the big point of uncertainty in Nvidia’s future: will we find new training techniques which require significantly less compute, and/or are better suited to some fundamentally different architecture than GPUs? I’m reluctant to bet no on that long term, and “long term” for ML right now is not very long.
mistercow··on Zero calorie sweetener linked to blood clots and heart disease, study finds
Yes. Read the actual study. And ask yourself, why on earth, with all the possible vehicles out there, someone would choose 20% of the typical RDI of glucose for a study on the effect of a sweetener on glucose tolerance. If they were going out of their way to find the most confounding vehicle possible, they’d have been hard pressed to find a better choice.
mistercow··on Zero calorie sweetener linked to blood clots and heart disease, study finds
> each bulked out with glucose to an equivalent size

Meaning that the packets were more or less pure glucose.

mistercow··on Zero calorie sweetener linked to blood clots and heart disease, study finds
Sucralose is also extremely potent, for what it’s worth, so that you only have to use a tiny amount of it.

It’s one thing you have to look out for in studies on the health effects of a lot of artificial sweeteners, particularly in animals. The doses might look reasonable at first glance, but are often orders of magnitude greater than anyone would actually consume.

(This doesn’t apply to sugar alcohols, mind you, which tend to have a similar or even lesser sweetness to sugar.)

mistercow··on Zero calorie sweetener linked to blood clots and heart disease, study finds
You can get neotame online, including on Amazon. I really dislike it though. Its sweetness has a very strange attack and decay, if that makes sense, so that there isn’t really a right amount to use where it won’t taste some combination of too sweet and not sweet enough over the course of a bite or sip.
mistercow··on The real "Wolf of Wall Street" sales script
Yeah, that was a startling admission to put in the post. If I were a shitty person who wanted to do scam people, I would probably just do it. It isn’t rocket science. But I have this weird quirk where I don’t like causing other people misery.
mistercow··on Show HN: Free e-book about WebGPU Programming
Yes, and not to mention automated testing. Apparently, deno has headless WebGPU support, which I might look into. Right now, I have this hacky build mode in my project which launches my Electron app, and sends IPC messages to launch tests (using my own, terrible minimal test framework), and then send the results back to the terminal.

On the plus side, it makes me appreciate how wonderful it is to have modern dev tooling when I’m working on literally anything else.

mistercow··on Show HN: Free e-book about WebGPU Programming
This is great to see. I’ve been working on a WebGPU project for a couple months now, learning as I go, and it’s rough how many things there are that take tons of digging to find even remotely straight answers about. There’s the basic WGSL language and JS API, which are, strictly speaking, well documented, and then there’s the stuff where you have to skip past documentation and find the one post answering a tangentially related question by someone with a singleminded interest in GPU programming.
mistercow··on Could AI robots with lasers make herbicides – and farm workers – obsolete?
Laser resistance can take a lot of forms. In the case of weeds, I’d bet that cat briers are already effectively laser proof. You can burn them down, but they send off shoots and store up massive amounts of energy in fragmented little potatoes. Miss any of them, and they come back up and keep going.

So that’s one solution, and it’s a big space for biology to play in. The bigger question will probably be whether or not there’s a route where weeds can make incremental progress toward better resistance from where they are right now.

For insects, as someone else pointed out, the easiest route is probably mimicry. But we might also see insects capable of recognizing the laser bots, noticing the chemical signals of burning bugs, etc.

mistercow··on Could AI robots with lasers make herbicides – and farm workers – obsolete?
But they’re basically guaranteed not to poison your food or local wildlife.
mistercow··on Could AI robots with lasers make herbicides – and farm workers – obsolete?
It’s going to be fascinating to see what new laser resistant weeds and bugs come out of this.
mistercow··on CrowdStrike representatives issue trademark infringement notice to ClownStrike
> I’m most definitely not trying to put CloudFlare in the middle on this… so I told CloudFlare that I will take the site off of CloudFlare; however, it is staying on the internet…

I mean that’s kind, but the whole point of DMCA safe harbor provisions is that they aren’t in the middle of this. They send along the notice, you file a counter notice, and that’s it for their involvement, yeah? If CrowdStrike wants to press the issue, they go after you, not CloudFlare.

mistercow··on How I got my laser eye injury
Maybe they meant the plastic wheel well liner? I don’t know if that makes sense, I’m just googling around looking at Jeep Cherokee images.
mistercow··on After 10 years, Yelp gave my app 4 days
A hard lesson I learned after several times getting burned in my time long ago doing indie stuff is “never depend on a single company without a backup plan”.

Sometimes this can be pretty difficult to even see. If you were a Mac shareware dev depending on VersionTracker and MacUpdate to drive downloads, with no actual marketing budget, you might not have noticed that you were implicitly relying on Apple not creating an App Store which would turn the entire ecosystem upside down.

This is one of the reasons that it’s so tough to make a business work at a small scale. In some ways, you have more flexibility than a big company, but the lack of capital means that certain events that a larger company would shrug off can totally upend everything you’re doing.

mistercow··on Alexa is in millions of households and Amazon is losing billions
I actually do use Alexa to buy stuff pretty often, but it’s stuff that I’ve already ordered on Amazon, and which I can’t conveniently buy locally, like pure sucralose powder and aeropress filters.

And that’s the problem. I’m not shopping with Alexa, because that experience is excruciating. I’m reordering things which I was already going to order through Amazon. They’re not actually generating new sales there; they’re just saving me from having to pull my phone out.

mistercow··on Show HN: Convert HTML DOM to semantic markdown for use in LLMs
Absolutely. I'm just making a more general point that "the same information in fewer tokens" does not mean "more comprehensible to an LLM". And we have more practical evidence that that's not the case, like the recent "Let's Think Dot by Dot" paper, which found that you can get many of the benefits of chain-of-thought simply by adding filler tokens to your context (if your model is trained to deal with filler tokens). For that matter, chain-of-thought itself is an example of increasing the tokens:information ratio, and generally improves LLM performance.

That's not to say that I think that converting to markdown is pointless or particularly harmful. Reducing tokens is useful for other reasons; it reduces cost, makes generation faster, and gives you more room in the context window to cram information into. And markdown is a nice choice because it's more comprehensible to humans, which is a win for debuggability.

I just don't think you can justifiably claim, without specific research to back it up, that markdown is more comprehensible to LLMs than HTML.

https://arxiv.org/abs/2404.15758

mistercow··on Show HN: Convert HTML DOM to semantic markdown for use in LLMs
Well, markdown and HTML are encoding the same information, but markdown is effectively compressing the semantic information. This works well for humans, because the renderer (whether markdown or plaintext) decompresses it for us. Two line breaks, for example, “decompress” from two characters to an entire line of empty space. To an LLM, though, it’s just a string of tokens.

So consider this extreme case: suppose we take a large chunk of plaintext and compress it with something like DEFLATE (but in a tokenizer friendly way), so that it uses 500 tokens instead of 2000 tokens. For the sake of argument, say we’ve done our best to train an LLM on these compressed samples.

Is that going to work well? After all, we’ve got the same information in a quarter as many tokens. I think the answer is pretty obviously “no”. Not only are we using a small fraction as much time and space to process the information, but the LLM will be forced to waste a lot of that computation on decompressing the data.

mistercow··on Show HN: Convert HTML DOM to semantic markdown for use in LLMs
I wouldn’t be so sure on reducing tokens. Every token in context is space for the LLM to do more computation. Noise is obviously bad, because the computations will be irrelevant, but as long as your HTML is cleaned up, the extra tokens aren’t noise, but information about semantic structure.
mistercow··on Show HN: Convert HTML DOM to semantic markdown for use in LLMs
I haven’t found any specific research, but I suspect it’s actually the opposite, particularly for models like Claude, which seem to have been specifically trained on XML-like structures.

My hunch is that the fact that HTML has explicit matching closing tags makes it a bit easier for an LLM to understand structure, whereas markdown tends to lean heavily on line breaks. That works great when you’re viewing the text as a two dimensional field of pixels, but that’s not how LLMs see the world.

But I think the difference is fairly marginal, and my hunch should be taken with a grain of salt. From experience, all I can say is that I’ve seen stripped down HTML work fine, and I’ve seen markdown work fine. The one place where markdown clearly shines is that it tends to use fewer tokens.

mistercow··on Show HN: Convert HTML DOM to semantic markdown for use in LLMs
This is cool. When dealing with tables, you might want to explore departing from markdown. I’ve found that LLMs tend to struggle with tables that have large numbers of columns containing similar data types. Correlating a row is easy enough, because the data is all together, but connecting a cell back to its column becomes a counting task, which appears to be pretty rough.

A trick I’ve found seems to work well is leaving some kind of id or coordinate marker on each column, and adding that to each cell. You could probably do that while still having valid markdown if you put the metadata in HTML comments, although it’s hard to say how an LLM will do at understanding that format.

mistercow··on Polychromatic Pixels
Yeah, and both techniques can be combined, which common with LCD screens, although it does sometimes lead to visible moving patterns when viewed close up.

There’s more flexibility with tunable wavelengths, though, since there will often be multiple solutions for what colors and intensities can be combined to create a particular photoreceptor response. By cycling through different solutions, I wonder if you could disrupt the brain’s ability to spot any patterns, so that it’s just a very faint noise that you mostly filter out.

← PreviousPage 7 of 34Next →