mistercow

I suspect they used an underpowered model, or did it a while back. I would be surprised to see 6.1 Sol or Astra produce those bugs, even one-shotting. Instead, I'd just expect it to have broader product issues rather than out and out bugs.

I've wondered for a long time why ear clipping + optimizations has won on triangulation over stuff like Seidel's trapezoidal method, which has O(nlog*n) time complexity and handles holes naturally. I know big-O isn't everything, but that's essentially linear versus quadratic, and it seems like we care about triangulating more and more complex polygons as time goes on.

Is it just a triangle quality thing? Does Seidel lead to too many hard to deal with slivers?

I think it's basically an alignment problem. LLMs tend to think it's better to just do something small and bespoke than to just pull a mature library and use it correctly, which, to be clear, they are perfectly capable of doing.

And this is in part a reflection of our own bad preferences. This is what always annoyed me most about the left-pad Discourse a decade (oh god, a Decade™) ago; "why do you need a library for something so simple in the first place!?" people would scoff. And of course the answer is that even the reference class of left-pad-like problems includes a bunch of examples where the problem has bizarre and sticky edge cases you'd never consider, so in general, "just pop in a popular and well tested library" is actually a pretty good policy (assuming your package repository doesn't allow anything really silly like deleting entire packages and breaking everything downstream).

Of course charting is more obviously complicated, but again, it might seem fairly simple if all you're thinking about is one kind of chart in one style with limited settings. The difference is that for a human coder, "just whip up some chart code" is enough of a pain in the ass that they'll reach for a library by default, whereas a coding agent can just spit something out which mostly works in a few seconds.

> when it can't be derived from the training data

This sounds like a goalpost on wheels. Can you define clearly where your stake in the ground is?

It's not a dichotomy, though. I loved coding for the sake of it. I found it relaxing, meditative, and I was very good at it. I also like building things where coding is a necessary step.

And while I don't go all in on the notion of revealed preferences, a few facts stand out to me:

1. Nothing is actually stopping me from writing code by hand for fun, but I don't

2. There is no practical reason to manually write homebrew SNES games in assembly language, but some people do

3. I still have no deficit of fun problems to dig into, to sit and ponder, and then experiment with. They just aren't coding problems, and I have far more powerful tools to help me navigate prior art, test out hypotheses, and so on.

I think this is really selling the severity of "SHA-1 is a Shambles" short. The author distinguishes between second preimage and collisions, but Shambles is a chosen-prefix collision, so it's sort of in between the two ends of the spectrum he explains.

The difference here is between me being able to send you a benign file and then swap out a safe file (the scenario the author describes) and me being able to swap out your own familiar, benign file except this time it has malicious content at the end (along with a bunch of garbage).

That second scenario is obviously much more dangerous from the perspective of human review and noticing that something is wrong. If you have a large, rarely changing file already, then content silently smuggled onto the end of it could stay undetected for a long time.

(In practice, it's not that simple; you'd have to actually figure out how to make multiple pieces line up with chosen prefix collisions, and I haven't worked through how you would do it. Maybe it's impossible without further weaknesses. But that guarantee is feeling pretty threadbare.)

Anyway, this doesn't matter because git switched to sha-1dc back in 2017, and afaik, this addresses the shattered/shambles problems. I find it odd that this isn't the point the author focuses on; instead, the only mention of sha-1dc is in a footnote saying that it could be replaced with his proposal.

And sure, maybe that's true, but why isn't that the entire argument of the article then?

Feynman for sure could have. In fact, I think he sort of did, although I'm not sure he meant to. Multiple generations of nerds now have taken books of his anecdotes varying in plausibility and obvious exaggeration as some sort of weird physics cult of personality gospel. But I don't think he was really setting out to curate his legacy so much as he was a good storyteller and he liked to entertain.

But could he have conned people on purpose? Absolutely.

I think it was on qwen 3.5 9B or thereabouts that I saw a ~30% decode rate improvement on my M2 Mac vs MLX. That’s probably the best I’ve gotten.

Keep in mind though that this is also sometimes use case dependent. Off the shelf implementations are generally pretty good overall, but can have pathological behavior on specific workload shapes you care about. So do this as a somewhat later optimization, and particularly when you see performance characteristics that don’t seem to make any sense.

It's been pretty ad hoc, but my prompts are nothing special. Things I generally do:

1. Top end model on high/xhigh thinking (last time I did it it was Sol xhigh I think)

2. Make sure it creates some representative fixtures of different sizes and sets up a good testing, profiling and benchmarking loop that doesn't require my input.

3. Make sure it has access to reference implementation code

Edit: Oh and one obvious pitfall that for some reason I still have to remind even smart models of from time to time: make sure it knows not to try to parallelize its benchmark runs. I've occasionally had an agent struggle to figure out absolutely nonsensical data because it tried to run multiple tests on the same compute hardware simultaneously.

On the SDK front, you tried just having an agent reimplement the model you're interested in and just use the weights? I've taken to treating off the shelf implementations as reference implementations anyway, because I can often squeeze out significantly better performance for my configuration and use case by having Codex hammer at it for a few hours.

I’m not sure how you can really make either statement work anymore. Now that smaller models are actually broadly usable, “behindness” is no longer a scalar and at the tails, where no open lab seems to be trying to compete at the >10T scale and no closed lab seems to care about <400B anymore, it’s just apples to oranges. It’s like talking about whether Qualcomm is “behind” Nvidia.

Tom7 doesn’t know the meaning of the phrase “low effort”.

> I'm fairly certain this will defeat 99.9% of malicious users, many of whom won't even know it exists until someone points out it's missing.

Everyone realizing that photos don't prove anything would defeat 100% of malicious users. I don't understand what people incorrectly trusting photos is supposed to achieve at this point, in your view.

Aside from the fact that this was obviously never viable and the entire problem is clearly unsolvable if you sit down and really probe it for fifteen minutes, what I find most frustrating about this is that the false promise of preserving photos as reliable evidence is actively harmful.

You will not build a perfect system, or even something near perfect. The best you're going to do is make it so that it's hard to casually present AI photos as real, leaving only the cases where it really matters. In the "best" case, you've just made the public more trusting of photos in general, so that when there's actual money or power on the line that makes jumping through the hoops to fake authenticity worth it, the public is more susceptible.

The best outcome at this point is for everyone to get on the same page that photos have roughly the same probative value now as drawings. Poorly thought out snake oil efforts to prove authenticity are only going to delay that.

Most old books that are rare and unpreserved are so because their value is marginal, so nobody has bothered to collect and preserve them.

But where did you hear that they’re buying “all copies”? And to what end?

I think the technology also dictated it. It’s easy to forget that in those days the idea of window chrome being full color, full transparency, with live window dragging and visuals not locked to tiling 8x8 or at best 32x32 elements was incredibly novel. And that was as much about preemptive multitasking and expecting and allowing every application to keep its windows up to date as it was about Quartz.

Aqua was garish and in your face, but it needed to be in order to be a symbol of what the system had achieved, which I think was especially important because that came at the cost of being a buggy, immature OS that had to run crucial apps in a hideous compatibility mode while it was getting established.

Edit: and if I can be a cranky old codger for a moment, I think that Liquid Glass unintentionally serves just as well as a symbol for what Apple has become: a company grasping to relive the memory of innovation without remembering how to innovate.

But the intuition based on the constant gravity assumption correctly illustrates the point, and adding in the more complicated picture of non-constant gravity creates new terms that exactly cancel out to give the same answer. Why would you talk against intuitions that work correctly?

And to be clear, building intuitions that fail at certain scales is still a useful and important thing to do. But you haven't shown a scale where these intuitions fail. That's not insightful, it's just throwing smoke bombs for no reason.

If you're talking about intuitions, you have no firsthand intuitions about lifting effort decreasing with distance to the Earth. We can intuit about constant gravity, and the math of constant gravity works fine for this description.

And while the real situation at scale is more complicated, the math is going to come out to the same answer, albeit with extra terms muddying everything up.

If someone says that something true can be illustrated intuitively with a thought experiment, "sure, but what if we take that to a scale where our intuitions fail" is a sort of odd place to take the discussion unless you're genuinely curious how the math is going to shake out.

IANAL, but isn't most of that unenforceable in the US? Typefaces have their own weird place in IP law which is different from copyright.

And I think the answer to the "doing the math" question is, until you've actually collected the data, "what math?" Until someone actually puts a bunch of six-figure value hardware through its paces, pushes the previous limits, and sees what that does to its lifespan, there's nothing to meaningfully calculate.

WebP generally preserves higher visual quality at the same compression ratios as JPEG. You can make bad looking images with either codec by demanding smaller file sizes, but all else equal, WebP usually looks better.

> More so than any other image format, JPEG takes advantage of the quirks of human perception to throw away information while largely preserving quality.

That is a very odd thing to say in a post that also mentions WebP. I would say that of all the commonly supported lossy image codecs released since 1992, JPEG takes the least advantage of human perception quirks.

Well, it’s one flaw. I would argue that the bigger flaw, which you alluded to, is that the cost of computing the cache yourself maxes out in the single digit dollars even very large frontier models, and that’s a one-time cost. Even if you imagine all the logistics are free and all the transfers are instant, what are we even talking about here from an economic perspective?

KV caching is a super interesting engineering space, especially when you’re talking about local models where compute and memory bandwidth are highly constrained and you’re trying to trim fractions of a second everywhere you can by flipping between different ICL prefixes. But selling caches for specific documents just makes no sense at all.

> Then the part that matters: where the KV lives

When your abstract was clearly generated by an LLM and not curated to at least make it sound human, it does not make me want to read your paper.

> And I didn't learn anything.

So first off, I mean, of course you didn't, right? You had already learned most of what you were going to learn from this specific project by doing it by hand.

But yes, if you use AI to do projects that might have been at the edge of your abilities pre-AI, you will learn a lot less. The solution, I've found, is to get more ambitious until you're back to the edge of your abilities even with AI. You should be asking the agent questions like "has anyone tried?" and keep pushing until it isn't sure, and you're not sure you have any idea what you're doing. Then ask questions until you feel like you sort of know what you're doing, and verify your understanding by directing the agent to build.

I'm pretty sure they just vibe coded the site in half an hour on a lark. High gloss without any real direction or intent behind it is the new "vanilla html page with no CSS file", or maybe less charitably, the new clip art.

And that's fine, right? Like people can do what they want. We're just going to get used to it.

Are your old hard drives encrypted using asymmetric cryptography? If not, I'm not sure who made you that promise.

My current hunch is that that benchmark captures most of the relevant gap between Anthropic and the rest. “Can’t distinguish truth from fiction” has long been one of the deeper complaints about LLMs, and the bullshit benchmark seems like a clever approach to testing at least some of that.

I don't mind if people use AI to help them write, but when I see this kind of thing, it implies to me that they're barely even skimming it before posting. Surely people don't want this super cliche AI-hype-man tone in their blog posts, right? And if they haven't taken the time to at least skim through it and iterate on basic style, why should I assume it's worth my time to read it?

> Maybe its because I'm using not-frontier models to do the coding

IMO it’s probably that. The difference between where this was a a year ago and now is night and day, and not using frontier models is roughly like stepping back in time 6-12 months.