We make fun of people using an LLM to spellcheck an email, yet here we are asking a human to define a common word, right?

Curate means you choose what goes in. Like monthly.xls orgchart.doc and context, for example, "This field means employee is NLE, unless they were a contractor, then it means the contract ...". This ain't rocket AI neurosurgery my dudes. Anyone can do that individually with existing tools, but what Cloudflare is trying to do, I guess, is make it easier to do that from one place and share with the company. Like Jupyter/run book type stuff but with LLM, probably. Just guessing but that seems kind of obvious to me.

No you're conflating very different things, this "tokenmaxxing" crap is some Microsoft/Amazon/Google/Meta scale thing where we have our own data centers dedicated to AI compute that could be seen as a wasted investment if not driven at 100% usage. Maybe some startups going balls to the wall at $API rates with someone else's money, but that is not "a lot of corps"

Then you have rando fintech, edtech, *tech adjacent business paying $20/head a month for office workers to have AI access because that's what all the other cool kids are doing. There you have Sally from Accounting turning bullet points into emails because the CEO said everyone should be using AI. That is a lot of companies, and what Sally is doing is not insane. Neither is putting a $20/month high value information tool at everyone's fingertips and encouraging its use to stay competitive. That is not the craziest thing happening in IT, ffs. A lot of that usage will go down the same way online video training services that the company bought and encouraged everyone to use.. queue up an Intro to Java video and go to lunch to show you used it. This isn't REMOTELY the same thing as "tokenmaxxing".. some AI subs will get dropped, a lot will stick, because it is genuinely useful every day for a ton of people.

You aren't forced and you aren't subsidizing anything, you're being priced out by people who want something more than you do and they're paying for it. It's supply and demand. It's like living in a gentrifying neighborhood, I don't know what else to tell you but nobody owes you a damned thing here. When people with more money than you do buy up all the inventory that's what happens.

Jesus dude, if we were actually talking about housing prices or closer to the issue, electricity rates, I'd have a little empathy for you, and I do think people are owed affordable housing and power, and other essentials etc, and there are ways of addressing those during spiking market demand.

But your PC parts, your PlayStations, I mean fuck off, I guess you've never lived through hard drive shortages, ram shortages before, $2000+ basic desktop PCs, internet access by the minute. Even if you think AI subscribers aren't yet paying for the full cost of AI services, you aren't subsidizing a damn thing unless you're the man behind the scenes providing private equity to some AI service provider.

... when in the past twenty years has replacing data center equipment been more economical? I don't get where you're coming from. We write down deprecation on IT equipment, that doesn't mean its value changes. Servers got bigger and bigger while workloads stayed about the same. It's what allowed VMware to grow into the behemoth it became. It also made replacing physical servers less and less necessary. You can run them well past their support life and replace only when truly dead or unreliable, or give up and virtualize it finally. The last big spend money to save money in IT WAS VMware. Well maybe it was mid tier SSD arrays replacing enterprise spinning disk ones, but anyways.

It's kind of like you're talking about depreciation of a minivan vs the value of a minivan full of cargo where every new day you sell a run full of cargo it's exactly as valuable as yesterday which was as valuable as it was on day one. Selling inference compute is like the selling cargo space, I mean sure you depreciate the equipment for tax purposes, but every day that thing runs 100% is money.

Catching up is great but there will still be a premium for higher and highest quality outputs for a long time because the bar where we all agree "this is smart enough, I do not need more intelligence" is really fucking high and will take us a very long time to get there.

Another threshold is when you can run an acceptable level of output from a local system that doesn't cost an arm and a leg. But then why would you not back that up with the highest quality remote model, pay the subscriptions but not rely 100% on them. Use a model router or toggle it. That only does those companies a favor and I don't see that going away until you hit that other stupidly high benchmark of no more intelligence desired.

I mean no amount of temu lawyers will make the best legal representation cheap. All Anthropic and OpenAI have to do is continue to slowly leapfrog each other. The demand for good inference isn't going away, and the demand for better and best isn't either. But yah one way or anither these valuations aren't going to math out in the end.

Our former FBI director is currently facing federal charges for posting a picture of "8647", apparently regarding a president that has already been impeached twice. The allegations are that 86 is a call for violence and not impeach and remove from office for example.

Mental gymnastics and fascist knee jerk solutions are kind of this administration's whole thing. Talk about out of the frying pan and into the fire. Maybe do a bit more research.

The moon has a left wing bias to these folks, because if it's not a very narrow and unintelligent subset of American culture, then it's woke or left wing bias or something like that.

Why the fuck shouldn't you be able to sell a book for more than five years. Authors do that AND continue to write.

If you're looking at it as if they will be forced to make money somewhere else and you'll have more creative works because lazy writers will now be forced to write more and more to make enough money from writing? That most definitely will not work out. When do authors stop writing, for fuck sake man, you didn't think that through all the way did you.

Holy fuck people complain about sequelitis NOW, eeeeezzzuz... In no possible fucking way does your plan increase creative output. That's all you'd get is increasingly more derivative slop, more and faster.

Yah I mean eventually I was going to ask a remote server about every little problem on my computer so it might as well just take a full copy without my knowledge, fuck it yolo, hahahaha. Why would just the chat transcript on my screen that is obviously sent to the server be the boundary, copy everything else too, same diff.

My point is we have lots of utilities that use remote servers to do work, and it's generally understood that the inputs sent to the server are now on someone else's computer. An AI model is no different than how "itranslate.exe --spanish lemonade stand.txt" might work. There's no excuse for copying everything else to some server along with the actual request, that's malware plain and simple. There may be a fine line between that and say you working on *.txt without thinking it through all the way, but that is a very solid fucking line.

JFC, it's even worse, it's like running itranslate.exe in interactive mode and then having it send all your files out because you asked how to say hello in Spanish. Imagine hearing dickheads on /. say well... the translation services are online so ... the fuck you say?!?!

I would say we can chew bubblegum and walk at the same time, but that is definitely not even a technical problem in need of a technical solution.

Dude... how do you compare reading a published book without paying for it to copying the contents of your computer to a server without your knowledge? The problem isn't if a model was trained on your tax records it's that it copied them all without any indication and they're all available to some random grok engineers if they weren't already thanks fucking DOGE.

One IS a copyright question whether you like it or not, the other is straight up malware. Due to the nature of how these things work we know anything in the chat session is on their server already just like anything you paste into a Google prompt is. The checkbox to help them train their model is a pinky promise and we all know it but we know what we're sharing and what we're not. Just because it's in a git repo doesn't mean it's public or ever will be.

The same author who's published books were read and used without his permission (the horror) just had his unpublished works and private notes copied and uploaded to grok because he was smart enough to track things in a git repo and used the tool for a creative brainstorming session. Whatever, same thing right, lol.

No, not all models hallucinate in the way you mean to the same degree or frequency. Like the other person said, a more useful way of looking at it is that's all they ever really do. They converge on the truth by reaching a logical consistency with the facts. They don't know what the facts are but they can generate some from their training, like how we say something from memory without thinking about it. They can look up potential facts by reading things, same as you can. Or you can tell it. With external lookups, your input, and its own output as it's next input, it iterates and generates more output that is hopefully logically consistent because that's how it's been weighted to use our language. The same way we use our language to think, you don't need it but you get a lot of utility out of it that way, that's what powers the intelligence of a LLM, distilled human language logic.

So it's not something that ever gets solved, not in an LLM, not in us. We never really know anything, we just say shit that makes sense, and if it doesn't we say it again until it does. We have a more complete model, of the non-language world around us and that gives us an advantage, but that's all any mind is and can ever be, an incomplete model of reality.

It's not a technical word, that's easy.

There's a technical reason to say sanctioned, untrusted, denied, prohibited, blocked, etc. there is no technical reason to call anything a blacklist. It's pretty simple, you can't explain what it is without using a technically better word, so call it that.

Master is kind of like that, but it can be used technically correct. It often isn't and master/slave definitely never is. If you have to ask what it really means, biggest, first? Then maybe it's a dumb name.

What do you think blacklist or master and slave technically mean?

He got a blow job, got impeached, and was still a great president. Bill Clinton is both a better human being and a better president than Trump will ever be. That's a low bar and he easily steps over it, no argument.

There is just no excuse for being as shitty a human being, in absolute terms, as Trump is. None. Lots of people would make bad presidents because it's a tough job, but for electing a known shitty human being, Trump voters have no one to blame but themselves and should feel ashamed. No, Bill Clinton is not an excuse.

Grabbing and giving back is a really funny way of looking at copying something. Pay for a copy or don't.