Posts by doublelayer

Re: Hi, I'm Charlie Stross

Since they were covering and combining multiple opinions, I'm not surprised they chose to quote rather than asking to republish your post verbatim. I also don't see the out of context quote, even after reading your original post. If it was about the possible local LLM-style tool that you would find useful, I don't have any objection to you wanting it, though some in the comments decide they do, and I don't see the article's context as any different than your own. If you were referring to the LLM explanation for translation, I think almost all of us understood what you were getting at from the article and one comment's comprehension issues when they thought it was actual translation is their problem, not a lack of context.

Re: It's worth thank we think...

They probably did train on the texts for those, as they did so on any pirated book they could find. I think you'll find that there are also very voluminous discussions of a lot of science fiction that discusses all of that on a bunch of websites. Not just science fiction, but in my experience it's one of the more common genres to have that much online analysis and debate. That means the model doesn't have to read the book and some transcripts of the TV series and make the connections itself when it can paraphrase the many pages that already say this.

Re: Microsoft's patch Tuesdays

The reason we patch operating systems isn't because something got in and rewrote them and we need to undo that damage. It's because they had a problem and we want to fix it. If we can't rewrite that bit, we can't fix problems we find in it. All your OS versus software quibbling does is to lock part of it down, preventing us from fixing that part, and change what you're calling the parts you can write. Those things would be as vulnerable to rewriting attacks but as easily patched as operating systems are now. Your suggested change would have no positive effects and negative ones on any part you actually made impossible to change.

You can't exclude all hacks by design. You can include a lot of hardware security layers which are difficult to get around, and that would help a lot with security. As with most security, it would also make the devices a lot harder for people to use as they like. Security and convenience often oppose one another, and the solution that introduces tons of guards on security tend not to be purchased by people wanting to get their work done today. I speak as someone usually pushing the more secure option. A lot of attention is needed to keep those safeguards from preventing users from doing what they need the computer for, and as your solution already mistakes how the security part would be implemented, I have doubts you're considering the convenience elements properly.

It's somewhere between very difficult and impossible to "open source" something like Optane. There are only a few companies with the ability to manufacture, let alone advance the design of, something like that, and open sourcing makes almost no difference to them as they all have plenty of money with which they could purchase it, either as a license or to own outright. Funnily enough, not one of them, not even Micron who already has the manufacturing experience, seems to have decided this would make sense, even though obtaining a canceled product from Intel who was/is in financial distress wouldn't have been that hard.

Nor would it be the miracle solution that the article and your comment suggests. I agree with the author that, if it had still been here, LLM systems would probably want it. Other than that, almost no difference. It still needs fab capacity to manufacture. That still takes a long time to build. That still gets locked into one product for a while. At best, we could have had one more fab making Optane, but another possibility depending on exactly what hypothetical we're using is that one of the fabs now making HBM instead was configured to make Optane and the same capacity limit applies. Optane didn't have a simple solution to fabrication limits.

Re: Software was a problem

You speak as if the bulk of boot times is waiting for disks. Some of that is starting peripherals, including hardware in the same box as the CPU, which have actually shut down. No memory change will reduce that. We already can copy RAM to disk and shut down, meaning a relatively fast restore to the existing state when you power back up. Using Optane for that would make that process quite a bit faster. But it wouldn't be instantaneous because the hardware startup would be the same. In most of the systems where Optane would be financially viable, they don't care much about boot times because they hardly ever reboot unless they are changing something in the operating system and, if they are, they will need to reload it from scratch anyway.

Re: Words cannot

They aren't running scripts intentionally. If you just include JavaScript in a message, clients, including webmail, won't run it. You need to find vulnerabilities like this which allow you to bypass the intentionally not-running they do. Those vulnerabilities tend to be easier to find in webmail which is usually built in the same JS you're trying to inject rather than standalone clients which can and often do use HTML rendering without any JS support.

Re: Automated poetry?

I assume they used a rhyming dictionary for this, which helps the poetry. Otherwise, I'd like to see the dictionary that has six nouns alphabetically between cloud and crowd but only six more between crowd and shroud.

Re: RTFA

No, it won't, because all the words will be recognizable English ones and the grammatical parsing will also look like English. Language detection methods would certainly categorize it as English, so unless they have another model trying to guess whether the content is plausible, it would not be sorted as any other language using the tools they already employ.

I think the people who made this and those who will deploy it don't mind the damage they could do to LLMs and LLM users, if not seeing it as a bonus. I am not concerned about that damage, though I wouldn't go to the effort of intentionally causing it. I don't have concerns about misleading users because LLMs are already unreliable and likely always will be. I could argue that, the more nonsensical the output, the more obvious that will be so damage saves people from better LLMs which sound more plausible but still emit inaccurate text, but whether that is true or not, I mostly think that users of LLMs have made their choice, can change their mind at any time, and the consequences are for them to deal with. I am also unconcerned about damage to LLM providers because it is the responsibility of those who make the things to manage the quality of their training data, training techniques, and post-training refinement, not me as the writer of text they ingest without permission.

Re: On the other hand...

I wonder how they know whether it was run. I know there are people who pull any new package pushed to some repositories to check it for indications of malware, but they tend not to execute the ones they're testing. If they only checked download count, that could explain some of this.

[edit] No, as it turns out, they did collect data from them, so 15 people or automatic systems did actually run it. At least one of them was one of those malware scanners, and Anthropic claims that they nonetheless executed the code on something that contained credentials, but they aren't being specific about how or who any of the other 14 were.

Re: Delay

If people are willing to buy stuff from you at a ridiculously high price, you make even more money by selling a lot more of it, even if you take your prices down a bit, than if you just take the existing sales. That's what any company, monopoly or not, would do. Memory is limited because it takes a long time for them to set up new manufacturing capacity and they have good reasons to worry that, by the time they complete that, people won't be willing to buy it at a high price anymore. But if they reduced production when they didn't have to, they would get a lot less in total, even with a higher price, than if they continued to make and sell as much as they could while demand is high.

They're not banning them, at least not in this particular campaign, due to mental health risks, and not completely. They're only doing it during school hours and they're doing it so that students don't ignore their classes. Adults usually are not taking classes and, if they are, it's their responsibility whether they're good students or not. Specifically in education, we take a lot more responsibility for trying to educate children whether they feel like learning or not but leave it to individual responsibility by the time they become adults. So far, I think that's likely the correct approach.

I wouldn't be surprised that phone bans are already active but enforced by teachers. My teachers wouldn't allow students to be distracted by things they weren't supposed to have in class, and I don't have any reason to think that has changed. Teachers who didn't care probably still don't, so children can still get around enforcement if they're careful about which classes they break the rules in. But even if something already exists in practice, sometimes politicians feel that they really must do something and here's a thing that isn't likely to make any voter very angry.

Re: "Should you use AI for a task?"

Using LLMs on earlier assignments it's more likely to succeed at isn't in the students' interest either. Some of them are stupid enough to do it, either getting caught because they aren't good enough at using it, or succeeding and getting passed when they don't understand the topic, which they will likely regret later if they ever need what they were supposed to have learned.

But you're ignoring the rest of my comment. In responding to Pete's comment, I've explained why students are first assigned things with known answers, but if they were not as he suggests, then you would be testing many more students on their ability to produce research, some of them being the same ones who cheated with LLMs this time and others being ones who would not have cheated but, since they weren't taught, are incapable of doing the research. That does mean you'd have more fake research to detect. I expect you would catch quite a bit of it, but would you catch it quickly or might you have to do rather annoying verification to be conclusive?

Re: "Should you use AI for a task?"

If you care about teaching people who haven't already self-taught themselves most of the topic, you'll have to assign things that train the basic skills and knowledge so they gain the ability to do something more advanced. That means that the solutions to introductory courses are easy to find online, and therefore easy for LLMs to reconstitute without checking. Your argument is a similarly bad one as if I argued that arithmetic problems given to young students are trivially computed by a pocket calculator, so we might as well not teach them that at all.

That's how a lot of courses work. They train students on problems whose answers are already known, and usually known publicly, meaning you can find or create solutions with the right software. That software isn't always LLMs, for example if you want a proof generated you're more likely to get one using software specifically designed for doing that, but it exists. The only alternative to that is not bothering to teach students, just give them research papers and see if they can independently do valid original research in the field. Many of them won't be able to without some instruction, and LLMs will cheerfully produce fake original research too which would be much more annoying to detect.

Re: Continuing the tradition

Which is a thing that good historians spend a lot of time dealing with. A lot of their research is repeatedly asking the question "How much can we trust this source when it talks about this topic", finding and defending reasons to answer it, and using that to enhance the knowledge. But also, your implication applies better to things before the modern day where records are much more plentiful, even when historians still have to extract truth from biased or mendacious records.

Re: Maybe not all the data came from America?

It's good not to overstate historical fact when you're trying to make points. Such facts including that the people running the US at the time had no debate about whether to side with the Nazis and wanted to, and when possible did, support the UK. The two largest camps were those wanting to support the allies and those who didn't want to fight at all. There was indeed a group of people who thought the Nazis were great people to emulate, and even most of them didn't argue that the US should enter the war on their side because they knew how impossible that would be to sell to the rest of the US populace. Their primary campaign was to not fight in the war at all because that's what they could possibly accomplish.

The same allegations you're using apply to the UK. It stood by and accepted as Germany committed serious crimes against other nations because it really didn't want to fight either, and it too had its set of BUF-linked organizations who thought signing on with Germany sounded fantastic. They weren't close to a majority either, so if I cited their high-attendance meetings and pretended it counted for more than it did, I'd be exaggerating or outright lying. If you want to complain about the US or I about the UK, we've got plenty of true things they've done. Let's use those.

Re: Linguistic Political Bias

Combined with that, the test they're running it against has already decided how to weigh those questions and assign them to two arbitrary axes. While I think two is far less bad than the one left-right axis, I think it's still incorrect as it is quite possible to hold opinions this diagram would put on opposite sides simultaneously without contradicting oneself. Internationalism versus isolationism is part of the questions they ask, but you can be internationalist in a "I will impose my views on other countries" or "I will try to collaborate with other countries" way and isolationist in a "other countries can do what they want as long as they don't attack me" or "I will try to separate myself from other countries to their detriment for no reason" way.

Sometimes, people do end up contradicting themselves. For example, one way to assess people's attitude toward the level of public services is to ask, as one of the questions does, whether the rich are taxed too little, the right amount, or too much. Even then, you probably have to ask what country and who counts as rich. Another way is to ask whether the government provides too little, enough, or too much services to the poor, again defining location and "poor". Some people I know will cheerfully answer that the government should provide more services and tax the rich less. It will be difficult to assign that to any location on a spectrum.

You need to check your brand names. Flock makes public surveillance cameras, sold to government for general surveillance. If you're talking about private house surveillance, even when it's also used or misused by authorities, you might be thinking of Ring or similar brands. As written, your comment is incorrect.

Re: Ah great

That's part of it, but I think there is also distrust of the users themselves. If I had camera glasses, and you knew with absolute certainty that they weren't sending the video to someone's cloud server for LLM training, but I could get the video from them and do with that file as I please including uploading it, would it be fine then? I think Defcon would think not. I doubt this is a problem we can do much about since covert cameras will only get more covert with advances, but I think it is part of the concerns as well as Meta's misuse of the data.

Re: differentiate someone using an accessibility tool vs a glasshole.

They have a more expensive model that does have one. No clue how good it is, and reports suggest it stands out more to observers, but it exists.

Re: I read it twice to be sure.

It wasn't too clear, but there's a reason it sounds like that. That quote/paraphrase wasn't from Defcon. It was from different people, but the information saying that was a few paragraphs up so it's natural not to associate them:

"Monopoly Events banned the same style of smart glasses from its shows, which include UK Comic Cons, in recent days."

Re: M$ has killed email

If you think megacorps are the reason the protocol you linked wasn't adopted, you're going the wrong way. The only place to attempt implementing that was Google.

The only anti-spam measure that protocol included has already been accomplished in email, namely sender domain verification. That was a problem behind a lot of spam when it was suggested in 2000, when spammers were able to send anonymous and impersonating mail and expect it to arrive in inboxes anyway. Neither is practical nowadays with enforcement of SPF, DKIM, and DMARC or a subset of those. The protocol would do nothing against spam sent the way modern spammers do it, from real accounts, usually free ones or ones they've broken into. It would have plenty of other implementation problems, hence nobody implemented it in the first place.

Email has a lot of problems due to its age, and a new replacement protocol would let us do several useful things. It would be good for us to be realistic about what we could really accomplish though.

Re: People issue is not to be overlooked

This is unfortunately a large part of it. I know two people who felt the pain of Microsoft breaking Office 2019 on Macs. While I think the right solution to this is Microsoft getting sued for theft of the product people paid for, unless that happens, these users needed some software to edit documents and spreadsheets. I suggested to both of them that they try LibreOffice, or since they were on Macs, Apple's Pages and Numbers. I prefer LibreOffice because it's not as insistent at making you use their default format, but either of those would probably have worked for what they needed and were free.

They rejected both suggestions without trying either of them. They suggested familiarity and compatibility concerns to explain that, explanations I don't count very highly since they hadn't tried either of them, but this is a problem even when the software's entirely capable.

Re: M$ has killed email

They didn't lecture you about spam and reputation. They lectured you about using the tools that, like it or not, are mandated by a bunch of mail servers. It's an open standard, and lots of systems have had to add support for it. You probably have to if you want your mail delivered, and Microsoft isn't responsible for that.

Re: We have all been there

The original example doesn't demonstrate any problem with disjoint file systems, where I can easily do a ls /mnt/sda and then rm * without recognizing that those weren't going to operate on the same files.

Your new version doesn't demonstrate Unix's superiority either, since if you wanted to delete all the files in the A: drive, you would obviously type del A:\*.* which would not care about the working directory. In Unix, leaving out the path separator would be no better rm -r /mnt/sda* could do even more damage and rm /mnt/sda* might do nothing or might do something but almost certainly not what you were going for. Neither operating system tries to or does prevent you from running completely valid commands that are not what you want to run.

Re: Anti-AI open source has an enemy in common, but almost nothing else

You have failed to consider, in your reading of headlines, that "Anti-AI open source" might not be one single party, and that their not being one might be one of the core points the headline was accurately expressing. But if you couldn't be bothered to read the article, thus having context and more information to use to back up your argument if you still thought the headline was bad, why should we care about your criticism?