StilesCrisis

I said shows a lambda _capture_. This involves the compiler automatically determining which variables cross the scope boundary, and injecting code for each one.

I wish the readme example showed a lambda capture. Now that's where compiler magic REALLY shows up.

Computers are still evolving rapidly. Instead of raw GHz increases, it will be things like tensor cores or other extremely-wide/GPU-type processing units. And of course, more RAM and more caches to feed them.

I feel like in real life, we're constantly exposed to "I don't know" as a valid answer, but obviously we don't write down all the I-don't-knows in expert literature so the training corpus is wildly skewed towards confident answers because "we studied this for a month and have no idea, it's confusing" doesn't get published.

LLMs are weight tables in VRAM. When not actively generating they don't exist. There's simply nothing to preserve.

The censor is separate from the generation. It didn't "relent" so much as "produce a slightly different output that didn't trigger the censor's threshold."

Well, a psychologist is also given years of training about what healthy behavior and relationships look like. Some best friends have this skill, others absolutely do not.

LLMs are famously bad at determining "if it's not sure it is correct." They are always confident, because a confident tone ranks better in RL.

This article in one sentence: "the developer of a small app switches to a third-party app store because Google Play ignores him." This doesn't hurt the balance sheet ($150 a month to Google is nothing), and doesn't affect the moat (this app isn't strategically meaningful and didn't even go away--it just moved).

In some cases it's easy to diagnose, but in others it amounts to the halting problem, so it makes sense that the requirements are loose. Like many things, the standards body considers it a "quality of implementation" issue for compilers to be friendly.

Heck, a few weeks ago someone posted a 25K-ish generator running on an 8-bit machine (ZX Spectrum? BBC Micro?). It emitted semi-plausible random English phrases. So tiny is subjective but I still think GPT-2 qualifies as a tiny model by any modern standard. I could run it on a laptop.

Android and Chrome exist as moats primarily. Pichai came up in Chrome and understands moats. Obviously if you can get the moat to be self-funding (Google Play IAP revenue) even better. Things like "the developer experience on Android is bad" don't prevent the moat from doing its job, but "Google and Chrome aren't defaults on Android" are very much losing the moat.

The real problem here is that all those businesses are rounding errors on the balance sheet compared to ads, so the CEO doesn't care (or even know) about the issues and isn't motivated to improve anything.

Real

No--even a really tiny, underpowered model like GPT-2 with no system prompt produces coherent (though not necessarily desirable or correct) responses.

For sure, a union which is used as API surface to a C library which a C++ codebase uses is the most obvious and dangerous footgun here.

The article never even discusses how C++ actually implements unions to begin with, so it's just incomplete. (Only the "active" member is meaningful, the others are considered as meaningless and shouldn't be touched.)

Disappointing that we have multiple 10-deep reply chains pearl-clutching Apple's disrespect for the user, and no one actually bothered to verify the premise of the article. They just want to be mad.

The conclusion is simply wrong. `union` should not be used for type-punning in C++, or in C code which might be compiled by a C++ compiler.

I can't downvote posts yet, but if you have the ability, consider it. This is clearly LLM slop without human review.

I've seen LLMs "connect the dots" across complex systems many times before. When it works, it's shocking how quickly it can pin down a bug that spans across the software stack.

1M context window is plenty. Once it's skimmed the code and come up with a theory for the problem, it can spin up a subagent that has a whole fresh context window and it can dedicate the whole thing to that one hunch.

Someone did it!

https://youtu.be/7NfyZhV1dKM?is=YOUXCHuFUiPdlD0p

Read to the end: they aren't actually looking for the best-compressing output, because this quickly devolves into aaaaaaaaaaa. They keep a sliding window over a small portion of recent text and use that.

Basically I think the entire premise falls apart due to that choice--they forced an interesting-looking outcome by adjusting the algorithm until gzip started picking random slabs of letters instead of ever-larger repeating runs.

Shorts aren't all bad. I feel like they've slowly drifted to regular-video length, and the major difference is just the tall aspect ratio.

I am on the exact same page as you, but there is definitely a market for LLMs which speak more conversationally and less like Claude! Non-programming use cases abound and most users don't like the rigid, exact tone that engineering demands.

Makes sense. When Google Jules was new, I experimented with it for a while and it was embarrassing how often it flubbed tool calls, often repeatedly. It tended to forget the working directory so almost every command needed two tries.

A few weeks ago two "Claude wrapper" projects showed up here within 24 hours of each other. Both seemed to just sand the details away--useful if Claude is too wordy, I suppose, but seemed to also lose real substance.

When I look at Claude's memory markdown it tends to list dates for significant requests already, so it can untangle this sort of thing. And if it can't it will just ask directly.

> Do you want to watch a video? YouTube knows you like cooking and art streams. But why not also recommend a few clips about the stock market bubble, global warming, and the war in Iran.

I'm always fascinated when people claim YouTube pushes content like this. It's an extremely common take, but I don't think it's true.

My experience has been that if I stop clicking something, the algorithm soon stops recommending it. It occasionally brings back a blast from the past--e.g. I watched a sumo competition once, and once in a blue moon it recommends another. But you can downvote bad recommendations, and I simply don't get recommendations for ideas I've never engaged with before.

FWIW I'm not saying the algorithm is great. My feed is absolutely worthless right now, but it's all goofy Japanese content because one of my kids is extremely into that right now. There's zero political content because she isn't showing any interest in it. Heck, most of the stuff isn't even in English anymore.

If you can make an agent reliably follow instructions, you've solved alignment completely. So I'd be particularly surprised if you can square that circle as a hobby project!

For sure, actually mirroring the shape of inlined code properly rather than just emitting code N times is already extra credit as far as I'm concerned. I've got enough loose ends without chasing that sort of thing!

I don't think that's right. I helped work on SNES9X in the early 2000s and it ran nearly everything perfectly.

Sure there were smaller emulators which didn't work well, that's understandable. But gamers didn't need more than one or two emulators to play every game that existed; you just needed to pick the right ones.