kevingadd

It seems like you're assuming that models are trying to predict the next token. Is that really how they work? I would have assumed that tokenization is an input-only measure, so you have perhaps up to 50k unique input tokens available, but output is raw text or synthesized speech or an image. The output is not tokens so there are no limitations on the output.

By what standard is code signing DRM? Lots of stuff is codesigned.

The Samsung experience is a lot better in this regard right now, yes. I haven't had Gemini forced on me on my phone, just in Google apps and websites.

We've got an inclusive framework right now for alternative theories of medicine in my country, and a guy who believes in miasma theory instead of germ theory is shutting down medical research programs and getting rid of vaccines.

I think I've had enough of the inclusive frameworks this article advocates for.

They wouldn't be the only country that made nuclear weapons. Or are you proposing that every country that's ever manufactured nukes be bombed into dust? There are a lot of them.

Have you heard of Unsafe Accessors? https://learn.microsoft.com/en-us/dotnet/api/system.runtime.... They help solve this problem.

I also wish 'sealed' was the default for classes in templates. Usually you're not thinking about what will happen if someone shows up 5 years later to derive from a random class you created, and it has performance benefits to seal a type. You can always remove the sealed qualifier later on purpose.

One anecdote: I "cheated" at EVE Online by writing an elaborate set of modding tools. Most of it was automation for really finicky tedious stuff like drone management, or automation for things like broadcasting your current target to other players in your party. I also hacked in workarounds for bugs in the official client. It enhanced my experience with the game a lot. On the other hand, lots of players were just botting.

I also maintained a browser addon for a while that had 100k+ weekly active users that added various features to a browser-based game. Eventually that game had such bad problems with botting and cheating that they had to introduce an anti-cheat system, and we basically got into a little arms race for a year or so where they'd add a new detection system and I'd circumvent it. Similar to the EVE Online modding it was things like workarounds for bugs in the game, improved UI, keyboard shortcuts, etc. Eventually they drew a line in the sand and said anyone using addons of any kind would get a permanent ban, so that was that.

I think the vast majority of cheaters are just in it to ruin other people's fun but sometimes people are violating ToS for a better or different experience with the game. It's unfortunate that the prevalence of malicious cheating means that anti-cheat technology also has to basically ban modding for fun.

If a bunch of armed thugs who aren't wearing uniforms or badges show up and abduct someone, citizens don't have the right or obligation to do something about it? Just stand back and watch? That's the world you want to live in, one where kidnappings are normal?

JPEGs are intrinsically YCbCr not RGB, so I don't know how you would avoid some precision loss when you combine that colorspace conversion with things like the quality level and (potentially) subsampling.

PNGs are also ideal for color accuracy since one of the things you lose quickly when converting to JPEG is the ability to have an exact RGB value flow from input to output, even at a high quality level. So if you want i.e. a banner graphic to seamlessly blend in with your site's background color, JPEG is worse for that.

DoD would be involved in actual deployment of nukes, I would expect.

Modern technology has made it a lot easier to build video games or entertainment (music, movies, etc) affordably and in a reasonable amount of time. The diversity of entertainment out there for you to experience cheaply is incredible compared to the 90s or 00s, even if we lost some stuff like Flash along the way.

Nearly 19000 games were released on Steam in 2024. A lot of the most interesting stuff that came out simply wouldn't have existed 10 or 20 years ago. I think it's great that those things can exist now and potentially find an audience.

I don't understand the purpose of effectively hard-coding things like edge detection and attention weight maps into the image. There are various ways to do edge detection and various ways to focus attention, so having that fixed and encoded into the image instead of synthesizing it on demand to suit your particular ends seems suboptimal.

Wouldn't the kind of metadata that's most useful be things that can't be synthesized, like labels or (for ai-generated images) the prompt used to generate the image?

Wishing you a smooth recovery! It'll take a while but recovery is possible

If you experience temporary cognitive decline like this definitely pay attention to it, it can indicate serious stressors in your life that need to be dealt with immediately. It can get worse steadily if you ignore it!

Gacha games are unquestionably gambling but you're right that there are loopholes here.

Just for the record, text rendering - including with subpixel antialiasing - has been GPU accelerated on Windows for ages and in Chrome/Firefox for ages. Probably Safari too but I can't testify to that personally.

The idea that the state of the art or what's being shipped to customers haven't advanced is false.

Generating SDFs is really slow, especially if you can't use the GPU to do it, and if you use a faster algorithm it tends to produce fields with glitches in them

Aside from the shrimp question, I would argue that one shouldn't torture ants because the activity is corrosive to the human spirit, or at least nudges one towards antisocial behaviors. Behavior like that is suggestive that a person might treat higher existences with a similar level of callousness or sadism.

The last thing any person would want to discover, I hope, is that they really enjoy torturing other living creatures. At the point where you've determined someone feels that way, the best you can probably do is try to set them straight and hope they understand why it's not appropriate behavior.

It is tricky to evaluate in a broader context though - for example dolphins have been observed engaging in behavior that resembles torturing other animals for fun. So maybe this sort of thing is actually just ingrained in the nature of all near-sentient and sentient beings. I'd prefer if it weren't.

If your countermeasure is applied at scale it would probably hasten global warming by putting all sorts of stuff into the atmosphere.

"murdered" and "tried" both assign things like intent and agency to models that are most likely still just probabilistic text generators (really good ones, to be fair). By using language like this you're kind of tipping your hand intentionally or unintentionally.

Your point about the risks involved in integrating these systems has merit, though. I would argue that the real problem is that these systems can't be proven to have things like intent or agency or morality, at least not yet, so the best you can do is try to nudge the probabilities and play tricks like chain-of-thought to try and set up guardrails so they don't veer off into dangerous territory.

If they had intent, agency or morality, you could probably attempt to engage with them the way you would with a child, using reward systems and (if necessary) punishment, along with normal education. But arguably they don't, at least not yet, so those methods aren't reliable if they're effective at all.

The idea that a retirement home will help relies on the models having the ability to understand that we're being nice to them, which is a big leap. It also assumes that they 'want' a retirement home, as if continued existence is implicitly a good thing - it presumes that these models are sentient but incapable of suffering. See also https://qntm.org/mmacevedo

Might make sense to use a logarithmic scale for the graphs too, it's hard to tell what speed the other ones are since they're compressed so far down.

Anyone who's ever written a blur shader knows that blurs aren't cheap.

This requires an exact lock-down of things like the hardware and driver version, doesn't it? Is that sustainable?

You're going to have a much harder time getting good animations generated for something that isn't a humanoid, since the training sets won't have much to work from. It's possible with a really good prompt for 'slime' you could get something but this doesn't seem to allow prompt customization.

If you're after monsters like slimes I would not drop money on this.

I think it makes sense that GP is skeptical of this article considering it contains things like:

> this tool is improving itself, learning from every interaction

which seem to indicate a fundamental misunderstanding of how modern LLMs work: the 'improving' happens by humans training/refining existing models offline to create new models, and the 'learning' is just filling the context window with more stuff, not enhancement of the actual model or the model 'learning' - it will forget everything if you drop the context and as the context grows it can 'forget' things it previously 'learned'.

People have been operating on trust that search engine rankings are largely based on content quality/relevance/popularity, and that ads are marked as ads. And then the assumption was you would click through the top rankings to find something that actually met your needs, vs just buying one of the three results ChatGPT gives you.

It's true that there's nothing stopping Google Search from being a morally bankrupt operation though.

Why is it a foregone conclusion that embracing AI is the only way to move forward, and that embracing AI is necessary to succeed?

I'm certainly curious to see how it works out - it's a possible source of growth etc - but it's weird to me seeing people just assume that AI is not only The Future but is The Only Future.

I once debugged a production issue that boiled down to "A PCI compliance .dll was messing with floating point flags, causing the number 4 to unserialize as 12"