Posts by doublelayer

Boy riding bubble realizes what he's on, asks for more air

You appear to have a point you're trying to make, and you're hammering the reality of the markets until you can make the point. I'm confused because I don't think I know your point, so I have to keep pointing out where your claims are wrong. You've just made another one:

"Companies that do well after being taken over by someone else don't count, because as a former shareholder in the taken-over company, you don't get exposure to the success after takeover."

Companies being taken over in stock deals is very common. If they do well afterward, your stock in the company that took them over will rise. That rising did get mixed with whatever else the company was doing, so if it was a massive company swallowing something tiny, then it wouldn't make so much of a difference, but people have frequently been exposed to the results of a takeover unless they chose to sell. If the takeover was at the peak of the bubble, it's likely those results are something you don't want to be involved in, but people still had to actively make that choice. But as I said, calculating that would be hard, so I didn't try to make any claims based on it.

As for private companies not counting, well, duh. That's why I mentioned them as the biggest problem in finding counterexamples; I'd think something like "Netflix was founded around then, and their shipping videos around probably would have made them bubbly", but I'd check and find they only went on the market in 2004, so that doesn't prove anything because they weren't in the bubble, or at least the public part I get numbers for. When I thought of a name and it was in the bubble, the chances were relatively good that it was up from its peak. The seven companies I listed were the result of checking about twelve, with two that didn't make it back and three that weren't on the market. I didn't have to pick and choose to get them, and even if I did, it would still be very different from your "two and only two" inaccurate report.

The fact remains that what you said was wrong with multiple clear counterexamples. You first attempted to downplay this by focusing on one of the seven and a time period of your choosing, even though using the then-to-now time period you originally used works with that one too. You ignored plenty of examples that were even clearer disproofs. And now you're putting effort into explaining how the things I didn't count don't count. At the start of this process, I assumed that you were simply mistaken. Now, it appears that you've got a point to make and are telling the rest of reality to get in line. I don't know where you heard the "only Apple and Amazon" claim before, but sorry, they were wrong. If you keep defending them, you will also be wrong.

That's true enough regarding Microsoft, but not all the other companies I listed let alone the many others that we could find*. However, when you say that nothing is up now, people will generally be considering the values now rather than the ones at a carefully chosen time in the past, fresh off Microsoft's wasting lots of money trying to copy Apple's portable devices and failing miserably. I agree with you about most of the predictions about Microsoft, because they're very embedded into the AI bubble and cloud will probably keep being profitable but won't grow infinitely. Investors really need to learn that nothing can grow infinitely and very little will come close. That is not the same as what you said.

Let's consider eBay. What could be more .com-like than that? It's just a website, a startup at the time, no un-duplicatable software, and it follows the .com pattern very well. It peaks in March of 2000 and then falls like a stone for several months just like everything else. How long until it beats mattress money? A little less than four years. How long until it was at 5% annual return all the way back to buying at the peak? Six months after that. It didn't grow massively using your arbitrary 15-year period, sticking around 5% annual return, but in the last decade, it has grown quite a lot faster. So if your definition doesn't work with that, the problem is much bigger than the now versus a decade ago problem.

In addition to all the companies that have continued to do well after that bubble, there are many others who were bought by someone else who has. They aren't being considered because it's impossible to tell how well the buyer would have done without them, but it happens. Bubbles end up losing and gaining people lots of money, but there sometimes is value under them, none of which proves that there's value under this one.

* To find the companies I listed, the primary challenge was naming tech companies that existed in the late 1990s and still existed today. Once I came up with a name, the chances were pretty good that they would fit the "up" pattern. There were some failures; my test of Cisco didn't work out well, but the bigger problem was naming companies that weren't publicly traded until after the bubble ended.

"Remember, the only two companies that are up on the dot com bubble era are Apple and Amazon."

That's not right. Several companies other than those two have had significant gains and were among the .com stocks. For example, eBay, which is 900% up on its 2000 peak, Microsoft 700%, Match (online dating) 450%, or Adobe 700%. That's just if I limit myself to software or services, and since you included Apple, I don't have to. Hardware companies have plenty of examples, from AMD to Nvidia (no surprise there) to Qualcomm (although their 2000 chart is really weird and their growth isn't too fast though they are up on their peak).

What does this mean for OpenAI? Nothing at all; there are lots of examples of companies which lasted and didn't increase in value and many more which fell dramatically and either went bankrupt or subsumed into something bigger. If we're going to disagree with a statement though, we have to do it on its merits. Altman's statement about the .com bubble is correct. Not only was it correct about some specific companies, far more than the two you named, but it was correct in general because a lot of the things that bubbly startups had in mind have been done profitably by someone else after those hyped companies failed. Technologies that were going to revolutionize things sometimes did. The problem with the statement isn't that it's inaccurate but that there's little reason to believe that the LLM craze will follow the same pattern. Altman's comparison is about as useful as if I said that many people have managed to summit K-2, so you should expect and bet on me doing it; the problem is not with the first part.

Re: I have a question

If they've finished building them, then they'll stay in operation because they're much cheaper to run than they are to build and people want power for something. If they haven't finished building them, they'll likely jam on the brakes and leave some half-built junk sitting around unless they can find someone to sell the incomplete project to.

Re: People Do Love Bullshit

People are very excited to see computers "thinking" (although I'm not sure why?)

Because, if this stuff worked, it would actually be pretty exciting. If we had a program that acted like current models did but gave reliable answers, there would be many things we could do with it. Some of those could be dangerous, for example putting it in charge of things it could cause damage with, but it would be helpful at finding information, dealing with noise, and all sorts of problems we spend a long time on. We'd have to balance the risks with the benefits, but there would be real benefits.

The only problem is that the software we actually have pretends to do that but can't manage it and won't improve as long as they keep doing the same thing, the people making it don't know any way to do better, and they're not trying to because people are still willing to keep buying the crap stuff. But when there are people who want the answer badly enough, some of them decide to trust this bot instead of realizing how unreliable its output is. There are some who want the conveniently packaged answer so much that they choose not to think about whether the answer is right or how they can trust that the next one will be.

Should UK.gov save money by looking for open source alternatives to Microsoft? You decide

Re: Extra option?

"Look up how many employees Canonical have, it's less than a 5th of that."

Now look at how many things Microsoft makes that Canonical doesn't make. A lot of software you can install on Ubuntu isn't written by Canonical. They package someone else's work, which is exactly what they should do, but it doesn't explain how much you would need if you wanted the same number of devs as all the software you're using, which is the wrong number anyway because you won't need to duplicate things that you are going to use directly.

Re: Extra option?

It depends what job that software is doing. The best way to prevent it is to have a lot of users who can all contribute a little. That works great for things that a lot of people need, which is why I can use Linux and expect that I'll keep getting updates because a lot of people are going to write them because they need them. If it's less common tools, then we start to get less support. It ends up being the one developer, who gets bored or employed elsewhere or unfortunately dies, and then there is unplanned maintenance cost. There are ways to try to blunt this, from donating some money to the developer to having an employee brush up on the code and make some fixes to finding a company willing to have a commercial support contract for the thing, but each has its costs and benefits that need to be considered by someone who can make a reasonable choice based on the consequences of needing it. This is a thing that people should consider when they adopt a certain piece of software, and it is a solvable problem, but it is not an automatically solved one.

Re: Mix because...

I introduce you to one of Microsoft's unsung heroes: Nijam. While not quite as prolific or active as Microsoft's sales agents, there are people like this there to make sure open source attempts fail and Microsoft comes back if it ever left at all. Let's see how they do it.

User: The specialist software I use requires ...

Nijam: Almost certainly false.

Reality: Probably true. Lots of software did, and some still does, integrate with Microsoft Office. It would have been better if that wasn't how the programmers wrote it, it would have been better if that had meant people didn't buy it, but it was and is frequent that both of those didn't happen and that dependency exists. How could this be handled:

1. Check whether you can emulate something so it works with the replacement.

2. Check whether you can get a different integration from the writers.

3. Check if you can easily find an alternative.

4. Allow Microsoft Office use for those who need to run this until you achieve something 1-3, but use something else for people who don't run it.

Or you can use the method of the friend of Microsoft, deny that the problem exists without doing the least effort to understand it, anger the users, get complaints, and torpedo the open source replacement because it can't do the job, which might be true or might not, but the people in charge understand that disruption is bad far more than they understand how different Office programs differ.

User: We extensively use Access databases/Excel macros that are untranslatable.

Nijam: You mean "that are broken"

Reality: Maybe, though those who like to claim total compatibility tend not to use anything complex and therefore have no basis for saying it. But maybe this is the user's fault. However, from their perspective, it works with Microsoft and doesn't work with your thing. You could help them fix it, help them replace it, train them to know how to do that themselves, or you could blame them and do nothing else. Rewind to that "disruption is bad" point again to see the result.

Codeberg beset by AI bots that now bypass Anubis tarpit

Re: While Anubis is a good concept

It may not even require that much. Anubis is built to only bother people so often, meaning that setting cookies and continuing to send them is enough to prevent it from coming back every time a URL is requested. For any bot author too stupid to know how to do that, it makes it much slower and more expensive, but for anyone willing to put a little effort into the bot, it's only slightly more. Unfortunately, the kind of people who operate training collection bots tend to have plenty of cash since they're budgeting for lots of GPU time to use the result, no scruples since they're going to use data they clearly don't have permission for, but lazy because they're grabbing everything rather than trying to maintain a logical set of training data, so they can and do manage this.

Alexa hits snooze on basic functions as alarms and timers KO'd in UK outage

Re: one man's ewaste is another man's ewaste.

I don't have one, but the features don't seem that hard to understand to me. You can configure it some more, so for example, as another poster said, you can configure it to give you a set of information when you turn off the alarm. No, turning on a preset radio station does not do the same thing unless you sit there until they get around to saying that information. As it happens, I don't want that, nor do I want something I can request music by voice from, so I didn't buy one. I'm not surprised that someone who did want that and could get that for a £40 purchase decided to do it. If they knew some of the privacy implications, maybe they wouldn't, but given the reactions when I've talked about other privacy consequences, there are many who can understand that and don't care. It would hardly be the most expensive alarm clock that someone chose to buy.

Microsoft pushes Pull print, so you don't have to dash to the printer to grab the 'Fire everyone' memo

Re: S-poo-ler

I know you're a diehard GNU person. So with your encyclopedic knowledge of things they've ever said, whether or not they matter, how did you come to the conclusion that they have something to do with CUPS? Oh, I know, it's because:

GNU originally wrote CUPS. Oh wait, no, Easy Software Products did that.

GNU now maintains CUPS. Oh wait, no, Apple does that.

GNU operates a modern, serious fork to CUPS. Oh wait, no, OpenPrinting does that.

CUPS at least uses a GNU-written license? Well, no, because the FSF wrote those licenses so if it was one of those, ascribing that to GNU would be flawed logic, and also because CUPS uses the Apache license, which neither GNU nor FSF wrote and I know you don't like.

Re: Er, hasn't this been a feature of grown up printers

Not generally the printers themselves, but networks of printers have had lots of ways to allow users to get their job printed at any printer. In my experience as an infrequent user, not an admin, they never work right. For example, the last time I remember using a system, the process worked like this:

1. Print the document to the geographic work queue because the business had offices on multiple continents. The geographic queue was actually named for a city I wasn't in, but the intranet document said that was still the right one, next to all the warnings about what print settings you could use.

2. Go to the nearby printer and scan my badge. The job should print automatically. I don't know what you would do if you had multiple jobs because there was no screen on this printer to select them from, but I only had one, so that shouldn't be a problem.

3. Try scanning your badge again in case that wasn't working.

4. Look closely at the LEDs on the printer, all of which are not illuminated.

5. Try to look to see if the printer is even plugged in (yes, but it takes a while to see that).

6. Ask the colleague who walked by if they've ever used the printer. Colleague repeats steps 3-5.

7. Walk with colleague to the next closest printer. Repeat steps 2-5 on that printer.

8. Ask team members if any of them have ever successfully used a printer here. They all report no, and a couple of them also try steps 2-5.

9. Accept teammate's offer to print this on his home personal printer and bring it back tomorrow.

10. Try to find the original job in the queue to delete it, but fail because it doesn't work like a print queue; every job sent to it is immediately removed from the simulated print queue and stored somewhere on some server, probably, who knows.

I've seen similar things happen to others. Earlier, a university I knew had printers with a similar scan process, each equipped with a USB cable to plug laptops into because that was by far the most reliable way to get something printed. No doubt the thing works sometimes, and probably those who print frequently know how to make it work, but when multiple IT and software people can't make it work, there's a problem. If the Microsoft version manages to work, that would be great, but I have a feeling it will be as broken as any other version.

I started losing my digital privacy in 1974, aged 11

Re: Why would you ever delete patient data?

"In today's world - you will be asked proof of identity for many different reasons - even more then before."

Exactly my point. If identification is universal, it makes it very easy for anyone I interact with to demand it. I am of the opinion that a few things, yes, including getting a loan, justify collecting it, and most other things do not. If other places demand it, I do not approve and would like to prevent it, especially when they store something which, if it turns out they leaked it, will cause significant problems for me.

"If you don't trust your governent enough, ask you why? Maybe because you and others ara voting to ensure there are crooks at the helm so you can ignore rules too?"

An interesting theory, and we could probably debate why I approve or disapprove of various governments for a long time. It's irrelevant to this, though, because we often seek to restrain governments before they do dangerous things, not because we mistrust the people running them now but because the dangerous things are too prone to abuses.

"in dictatures, people get the politician and goverments they vote for"

I think you may need to learn what dictatorships do again; it seems you missed out something the first time.

"Can't understand why people fear government more than opaque financial and tech companies with have even more interests to exploit people - and upon which people have no power."

That's an interesting interpretation of my post describing private users of the identity as the bigger risk than government, stating that you don't trust the private users and that you might, I even assumed it, trust the government. But to argue the point I never said, I do have a little more power over private companies because I have some choice of whether I interact with them. Not a lot, certainly, but if some business demands my passport in order to use their services, I can leave and try to find someone else's services, whereas if the government demands it, I probably don't have a choice. Hence, I will try to improve privacy in my connections with both types of groups to the meager extent available to me, and if you agree with any of those areas, I'd welcome your collaboration where we agree.

Re: Why would you ever delete patient data?

One intrinsic option if you have universal identification is that others start to ask for it, and now you have two problems:

1. People you don't know or trust, assuming that you do trust your government, have copies of your identification information and may be storing it with terrible security procedures.

2. It is now much easier to link all activities you've taken with that single identity, whereas there are various methods available for somewhat anonymizing other identifiers for you if you're motivated to do it.

Back to being FOSS, Redis delivers a new, faster version

Re: Sorry, but why do we need so many Open Source licenses ?

From the statements of those changing the licenses, that's the problem. It's usually unconvincing when we look at what was contributed from those evil hyperscalers, generally a lot more code patches than you'd think. Because of that, the fact that an open source license explicitly allows anyone to host the software without paying the copyright holder, and that no company who has done this has yet rewarded all those who contributed code they're making money from, I tend not to agree with their arguments.

Re: Sorry, but why do we need so many Open Source licenses ?

"Public Domain: mainly an artefact of the US higher educational system: stuff written in/by students as part of their studies belongs to everyone."

Where did you get that from? Whatever the answer, it's wrong. Public domain has nothing to do with education or the US. Public domain is what automatically happens when copyright expires, meaning that anyone can do whatever they want with the covered work. The closest connection to the US is that the US government can't hold local copyright to things it writes except for a bunch of exceptions. Some people choose to put their work in the public domain, which is the most permissive they can be. For example, the SQLite database and surrounding software is set up that way rather than using any particular license. Students are not required to do that, even for their homework, though if they try to assert copyright and charge their educators for access, they can expect to receive a failing result.

Microsoft wares may be UK public sector's only viable option

Re: Grand Enshittification

I've seen that on occasion, but my personal laptop, if I boot to Windows, has it started and logged in within 20 seconds and my work machine in about 40, most of that being the dual login system. So I don't think Windows can be fairly blamed for that.

Re: I read this as propaganda

That works better with the Oracle example than it does with the Microsoft example. The difference is that in the case of Oracle, there's still a substantial amount of stuff specifically developed for Birmingham's particular needs by someone, whether that's Oracle themselves, some other business using them, or Birmingham's hired programmers. On that, I agree with you; calculating costs is very difficult for both approaches and trying to claim that one is easy and the other isn't is unconvincing.

But if we're looking at something like Office365, there's a lot less code being written for the individual user. People already know most of the features involved in this, what they cost, and, if they're already using 365, whether they need them. Figuring out how much it costs to add Intune for device management is mostly a licensing question with some calculation of how long it should take to make the policies that will be enforced. Compare that to developing or adapting an open source device management system. The costs for that will be hard to determine, because if you are building it for the new government Linux distro you're also creating, you'll have more things you need to build in but more control over the system so more ability to integrate them during development, whereas if you have to manage fleets of lots of operating systems, you have to develop many different clients and deal with OS makers breaking things you relied on. That is, in fact, more difficult to calculate and it's likely that, if you only consider finances, it's more expensive. A proper comparison needs to also compare the expected features, maintenance cost, risks, and all those complicated things which makes both prices, but especially those for the system that doesn't exist yet, hard to calculate.

Another reason this is hard is that we're not just comparing two alternatives. We haven't decided whether this should be a single government body making all the software they no longer want to buy, a process where other businesses are asked to do it and sell the result to the government, a process where businesses are given cash and asked to make the result open source, a process where existing software is supposed to be used unmodified, or separate methods for different parts of the government. That makes the calculations hard, and although they're well worth doing, many of those who would have to would prefer not to.

Re: Do we have to explain the difference between Open Source and Free Software again?

It appears we do, because we've got another person who doesn't know and is making up some definitions. Welcome to the class, coderguy.

When we talk about "open source", we're referring to stuff that is either completely compliant with or very close to the OSI's definition. We may disagree about certain parts of that, but most of that is considered required to qualify. Your definition: "You can read the code. You may not do anything useful with it though. Maintenance is usually done by a single entity." is not that. Those who try to pretend their software is open source when it would qualify only as your definition will earn our scorn because it's not open source. In case you're interested, they can also lose a lawsuit because our definition, the one where you have the right to fork and distribute, is considered so correct that courts have ruled that those who don't intend to provide it are lying about being open source. Now, we've explained.

Re: I read this as propaganda

The problem being that deployment costs with something open source can follow the same pattern that the Oracle thing did. The reason for that cost wasn't Oracle coming to Birmingham and telling them they had to pay ten times as much. It was systems not functioning properly, processes being delayed for months or years, having to build the custom parts over and over again. Probably Oracle's business or software can be blamed for some of that, but not all of it. A bad plan can absolutely obtain the same level of breaking things around an open source database, and while that would be less expensive because you're not paying for licensing the database, it doesn't help with the rest.

In some ways, this is an argument against the one from the article, because having a commercial software provider doesn't prevent that kind of implementation cost. You can still have a disaster when changing something, even when Microsoft or Oracle is involved, because neither of them is making sure processes aren't interrupted. However, given the frequent responses implying, or sometimes outright stating, that open source would fix this kind of thing which it has nothing to do with, it's a point which seems to go against both attitudes to what kind of software should be used.

Crypto-crasher Do Kwon admits guilt over failed not-so-stablecoin that erased $41 billion

Re: erased $41 billion?

Not entirely, although that value is partly notional money. About $18b was invested in the TerraUSD token, and that was supposed to be stable, meaning that the value lost when that token collapsed wasn't the result of prices going up. The rest was mostly based on the collapse of the floating Luna token, so most of the $23b remaining was based on a notional trading price. So if 2/5 counts as a tiny fraction, you have it correct. In terms on who got that money, it's not just the people running the systems involved. Some people who figured out how to break the not-so-stable coin were able to convert their holdings to Luna and cash it out at that high price. A lot of funds ended up going to speculators and blockchain hackers, but the operators of Terraform managed to keep some, probably more than we know.

Marc Andreessen wades into the UK's Online Safety Act furor

Re: How Andreessen might have avoided this situation...

"If there was already a widely-adopted solution, regulators could just point to it and tell websites to get with the program."

But they wouldn't. They wouldn't because any OS would come with a place somewhere where you disable your PICS settings. That thing would be locked behind the admin password. Some parent would make the point that their child could get that password somehow, just as you've just suggested your child will use ADB to disable your controls, which they probably won't, you could probably prevent, and is not the kind of problem you think it is, but that was your argument why another thing was needed. More importantly, that's still an opt in solution that needs a parent to think for a few seconds. The people who passed the OSA do not want that. They want a blanket solution which not only isn't opt in but doesn't have any other options. That's why they passed the law they did. We could have that, because it would be just another parental control mechanism like the many that aren't easily bypassed, and it would be as effective or ineffective as those have been.

Meet President Willian H. Brusen from the great state of Onegon

Re: I was surprised that ...

My guess is that the "Willian H." comes from "William H. Harrison", an actual president, and the "Brusen" I have no idea. Some of the labels they use clearly have a connection to the right answer, whereas others, if they have one, are far less clear about it.

How OpenAI used a new data type to cut inference costs by 75%

Re: "smaller, faster, and more importantly, cheaper"

No, I speak as if FP16 is the original precision, because you're right that you can't go any higher than that. Whatever the original precision they used before quantization, that's as high as you can go, just as my zoom with resolution method won't let you get any more resolution than the camera originally gave you. With the camera, you can degrade it and then recover information later by keeping those copies in parallel. If you can do that usefully with an LLM, then the high precision version will be whatever they had the first time.

Re: "smaller, faster, and more importantly, cheaper"

Not at all if you keep both things. Keep the full-res image somewhere, but only send the downscaled one. When a user zooms in, crop the relevant portion, downscale it less, and send that. Result: a picture that looks like you can zoom with a lot more detail than you'd expect but you don't have the bandwidth costs of sending the whole thing down the pipe.

Whether this works for inference is another question. If they train a model on FP16 and quantize it for use, then they still have the unquantized version around somewhere. I don't know whether there's any real ability to or benefit in cutting over from one to the other mid-inference.

The problem with answering that is that they have very little information about the quality of any answer and don't care because, if they did care, they would still be trying to improve their models rather than selling them. Quantization on its own means the quality will degrade, but a newer model, trained on better data, could still produce better output quantized than an older model did unquantized. This version of quantization will produce better results than normal FP4 does at the cost of the model being bigger; it's almost as if bits aren't infinitely compressible.

Also, you can't predict losses the way you did because burn rate is not directly proportional to revenue. Although it's expensive to execute their models, most of their expense is in the training process somewhere, so if more people pay them to use the models, their burn ratio will decline.

Your CV is not fit for the 21st century – time to get it up to scratch

Re: Surprising lack of pushback ?

Perhaps you're not getting much pushback because that question is mostly unanswerable and unimportant to us. I'm not convinced that the AI is better, but without running three methods in parallel (seeing candidate lists from the AI, human HR parsing, and reading them all myself), there's no way for me to know for sure. I'm not a hiring manager, so I can't do that, if I was I'd only be allowed to run one of those, and so that information is only known by people who did the studies which are probably mostly people who want to sell one of those approaches who are untrustworthy. It's also unimportant because, unless I'm setting HR policy, I have no control over what method is used. From the perspective of the candidate, the one you wrote from, I don't have a way to know what method they use and I can't make any choices based on it.

I'm not convinced because you don't appear to have done anything at all to try to prove or even defend your opinion. You just stated it. I'm not sure where you would get the information needed to attempt a proof of the opinion, and since you did not provide one, I assume you don't have any and that's why you didn't try. But since I don't have any either, who am I to tell you you're wrong, especially when I'm not sure you are wrong? Among other things, AI candidate filtering is probably the only available way to deal with the flood of AI submissions, so it could end up being better merely on the metric of not causing hiring managers to give up and never hire again, even if it is throwing away good candidates ten times as much. Better hiring managers might be the better solution to that, but unless we can make that happen, it's not a realistic one.

Re: Master and the slaves

You can try being the person who employers search for. Some people get to be that person. Others who try get nothing because they're really not that unique. When there is an open position, it's more frequently the case that the employer has money and the employee needs it than any of the other squares in that table, so the employer often has the stronger position and can expect candidates to come to them. There are exceptions. Companies known for mistreating workers, underpaying, or looking for unusual skills may find that they can't get anyone to apply, and now it's on them to try to find someone and convince them to apply. That also happens for intense jobs or ones where the employee isn't certain of a long-term option. Startups, for instance, often have to decide whether to do that or to pay large amounts and advertise that they will in order to get talent.

The fact that employees usually have to do more work to get an employer to notice them isn't the law. It's the inevitable result of one party to the interaction wanting something more than the other one. When employees are hard to find, it goes the other way and quickly. Any person can try any level, and they only need to increase the work they put in if they find that the level they're doing now isn't getting them the results they're hoping for. If we all decide to not bother with this, the companies with bad filters will get bad candidates and may eventually clean up their processes. However, for us to all not do this will require that some of us don't get jobs we would have accepted otherwise until the change happens.

Re: "Include every damned language, tool, protocol, mathematical technique..."

Because they won't, and finding the amount they're willing to pay requires doing interviews. There's no point doing interviews in and for a thing I don't want to do in the hope that afterward they offer some ludicrous amount of cash that changes my mind. Anyone paying £2.5k per day for that can say that up front. I won't believe them, but that's a bridge we can cross the first time anyone makes the claim.

Intel chief Lip-Bu Tan to visit White House after Trump calls for him to step down

Re: Ironically, this is the right thing for the wrong reason.

What would you suggest as Intel's strategy, given that intensive R&D on everything they used to do would, following your analogy, be taking someone with a serious disease and enrolling them in all the olympic sprint events? I don't exactly like Intel's current strategy either, but they are not in a good position where there's an obvious path of success ahead.

The International Obfuscated C Code Contest is back for 2024

Re: Devious

It will work on a big endian machine. By the time that the character is sent to putchar, it's already been converted to an integer in the host's preferred byte order by the compiler. The putchar function will cast that to unsigned char, which will take the least significant 8 bits no matter where the CPU chooses to store them. Programmers sometimes learn that to their detriment when trying to use non-Unicode-aware functions like putchar with Unicode strings.

Humans make better content cops than AI, but cost 40x more

Re: Complete waste of money.

Advertisers have long decided that there is value in the positioning of their advertisements. They try not to have their advertisements positioned next to things that would degrade its image. Are they right that there's a difference? I don't know. Probably they don't either; advertisers seem uninterested in proving the kinds of things they spend their time and money on, and it's not the easiest question to answer even if they wanted to. But this is far from new, as they've been doing this ever since the time when advertisements were placed out in public.

From their perspective, it doesn't matter if the advertisement is shown to a pro-Hitler person among pro-Hitler imagery; that's probably a positive. The problem is when it's shown to an anti-Hitler person among pro-Hitler imagery, where the viewer's disgust at the surrounding content might, and they're not sure if it does, convert to disgust toward the brand. You may be correct that this doesn't matter, but it's not a new thing that advertisers care.

Tech support team won pay rise for teaching customers how to RTFM

Re: Netware

Not necessarily. On the same archive, the version 2.2 image consists of 19 1440k 3.5-inch floppy images. If those also shipped on 360k floppies, that could be almost 80 of those. Any number you may state can be backed up by some of these versions. It seems they added a lot of something between 2.0 and 2.2.

Re: Netware

I have never used any version, but to try to answer the how many disks question, I found an archived image of NetWare 2.0 which consists of 13 360k 5.25-inch floppy disks. I'm not sure if that's the only option they had, but if you could use 3.5-inch disks to have fewer, then you could cut that to 8 720k or 4 1440k disks, either of which would probably make for a rather small package if trying to show someone clueless what a lot of money was spent on.

Mexit, not Brexit, is the new priority for the UK

Re: The elephant in the room is Brexit

Scale has a lot of advantages. One of the reason Microsoft has so much of the market this intends to replace is that they can write the same software and sell it to lots of people. If the intent was either to encourage Europe-based businesses to compete by only agreeing to buy their software or building it and making it open source, then having more people who can develop it scales very well. One of the nice things about software is that the cost of one more copy are very low, but the downside is that the costs of the first copy are massively high, as are the costs of developing additions and fixes. The more people you can divide that cost among, the lower it is for everybody.

If you tried to do this as a multinational project, then you would have more problems adding to the cost. The UK's version could be English-only as long as Wales's local government isn't using it, but a European version will need better localization. But that kind of thing is a lot smaller than the set of features which would be common. This also depends a lot on exactly what this new software is intended to replace, whether it's Office365, Windows, databases, or any number of other products, because some of those will scale better than others will.

Re: Alternatives

Of course M365 or Google Workspace doesn't give you everything you need. They never said it did. What they said is that it gives you more things than LibreOffice does, so if you want to replace it with something, you need more than LibreOffice to manage it. They went on to say that, if you need more stuff, there are two necessary steps:

1. Finding software that does all the things that you used from what the previous thing did

2. Making all those separate pieces of software work well together

Neither of which is impossible, but both of which take some effort. There are some people who insist that it's really just a drop-in replacement, usually only naming Linux and LibreOffice as needed software, and anyone using that argument is unconvincing when they don't have to be. There is a lot of open source software out there that does most of what Office365 provides, but as long as we tell people who know what Office365 provides to use LibreOffice and then stop talking, we look like we have no idea what we're talking about.

Re: Alternatives

They didn't say it was. They were pointing to a problem with something which definitely was FOSS, and something that could happen to any other piece of it, and describing it as a problem. Debian looks strong, and let's assume that it is and will stay that way. They asked for far more than an OS stack, and are you confident that any given open source program that is part of that stack is as strong as Debian is? Your simplistic answer is exactly what they were cautioning against.

However, I'm not convinced by their reliability point either. It is a problem for an open source deployment, but it's also a problem for a commercial one. Commercial software gets dropped as well, as schools who used Windows 11 SE devices just found out to their detriment. There is little any user can do to guarantee that the software they rely on won't become unsupported. Between the two, open source has the advantage that others can resume development of it if it is dropped and commercial has the benefit of legal agreements that prevent it from being dropped a week ago and you didn't see the Mastodon post announcing it, assuming you read those agreements in the first place. Neither approach will prevent old tech from needing replacement with something else because development has stopped.

Re: On a related note

"Firstly I was talking about new projects so there is no previous work to re-use."

There is almost always stuff to reuse. A lot of tools have been developed specifically to make writing reusable code and then reusing that code really easy, and that happens. And yes, quite a lot of that will be the code the provider wrote, so they are able to assign that copyright to the UK government, at which point they will never be able to use it again because, by assigning copyright, they no longer own it. They certainly won't be able to reuse anything for other places, but they might not even be able to do it for other government software unless they can get whoever controls that to grant them a license to reuse it. Having tried to get permission to use code that the copyright owner wasn't doing anything with and clearly didn't care about, I know from painful experience that this is not easy. If they were going to reuse code, expect prices to go up to deal with that.

There are some cases where there is much less code to reuse, in which case this works better, and although I'm not too familiar with the UK's procedures, government owning code for custom applications developed for them is not very unusual. What is unusual is for them to own the code to other things those applications were built around, such as databases or operating systems. If they were on an open source thing, that will be better, but if they had Oracle databases and specified a program that connects to those, then they'll still be paying Oracle even if they own the code that's writing there. If you want them not to, a company owning the code for a smaller application is not your problem. You want to make government departments design around and specify for your accepted set of software, and whether you can manage that or not, it will only happen if you know which part of the problem you should focus on.

OpenAI’s new model can't believe that Trump is back in office

Re: ollama rm gpt-oss:20b

They didn't train it on material saying Trump didn't win. They trained it on material from June 2024 and before, meaning the model has no clue who the president is now. But instead of acting like a human would and saying "I don't know", the model is effectively required to give you an answer. What you are getting isn't the material it's trained on. It's a guess from random jumbled data. And that's what you get from any other model on any topic. Some models will give you the right answer because their data is up to date, but if we copied the model, went to 2029, and asked that who the president was, it wouldn't know, wouldn't find out, and if we manage to convince it to try to answer, the answer would be useless as well.

Kind of, and sometimes, that works. Generally, this is implemented with some initial prompting. The prompt isn't as simple as this, but effectively, it ends up saying "If the user asks about something that happened after June 2024, tell the user that training data ended at that point and decline to answer the question". Sometimes, it successfully identifies that the US presidential election was in November, that's after June, and that prompt is honored. Other times, it doesn't get enough weight on that and proceeds down the normal approach which guesses the most likely words.

The problem with any LLM is that they don't know things. I don't mind calling it "lying", but it does suggest that the LLM is aware of information it's not providing, and it is not at all that far. It's using clever random number generators to guess words. If the words make truth, great. If they make falsehood, great. As long as they look like natural sentences on topic, the model part is fine with them, and it doesn't always get that much.

It's more likely that the problem is what the model said it was: OpenAI trained it on data ending in June 2024. Why they did that is another question, because there's reason to wonder when they started and stopped training the base model as opposed to the layers of prompts and protections around that which make the thing we have access to. Either they trained this a while ago and have been taking a long time to build the rest or they just didn't bother to update their data when they started, but it means there's no information for the model to even know that Biden had been replaced on the ticket.

Combine this perfect lack of knowledge with likely prompting to prevent it from lying about the outcome of the 2020 election, and you have the perfect way for it to come to a conclusion. Combine that with whatever is causing it to not change its mind, maybe some reaction to other models which change their minds on command, and we have the situation described in the article. All this should indicate yet again that there are many things an LLM can't do and many others which it might do but can't be trusted, so verification is required.

GitHub CEO: Future devs will not code, they will manage AI

Re: Fusion

Except that they do spend lots of money on fusion. If that works, it will be wonderful for everybody. If it doesn't, well that's what's happened so far. Be careful whenever making a statement that something is definitely possible with enough money. Most who do find that they're wrong because it takes more than cash. Lack of funding can cause lots of problems, but having funding can't fix all problems.

Re: If future devs "will not code"...

That fails for two reasons.

If my C is not right, I have to fix the C by knowing what it does. If my LLM-generated code fails, unless I can fix it only by telling the LLM to do so, then I will need to know what the underlying code does. If the envisioned process is (1) send prompt to LLM to generate Python, (2) test the Python, (3) fix the minor bugs in the Python, (4) release the Python, then unless I can read and write Python, I will fail at step 2. Somehow, that seems to be the general process recommendation from people predicting the demise of programmers as a career.

But the other one is more obvious, which is that occasionally, my Python isn't working, and I try to fix it, and I fail, and eventually I look into it further, and it turns out that the Python is actually fine, but the C that implements the interpreter isn't. I can't do that unless I know C. I can't do the same to a C compiler unless I know assembly. Most of the time, I don't have to do that, but that's because others are doing it for me. Who is going to make the LLM produce working code? So far, they haven't been too active at preventing it from screwing up.

Re: If future devs "will not code"...

That's correct, but from the perspective of a user, the problem is in the first part of the analogy: they're not using AI to learn, they're using it to get something done. From that perspective, using a fork lift is a perfectly acceptable solution to wanting something higher than it was and not wanting to do it by strength of muscles. The analogy would need to be corrected to specify a forklift without any of the balancing or containment features that normally exist to prevent the load from dropping off one side and crashing to the ground. Somehow, that one doesn't have the same conciseness. I would use the analogy of outsourcing the task to an unsupervised child, except that with some of the people I'd be using it with, I'd take my chances with the child.

Re: Can't wait to be a former developer

But shareholder will care if the code doesn't work and people don't use it. A few companies may be in a position to release whatever buggy crap they feel like and remain powerful, although several have found that they don't after all, but a lot of companies don't have that. If Small Financial Company LTD. finds that their AI investment software has a bug which resulted in them buying far more of something than they wanted to, shareholder will care. If New Mobile App LTD. finds out that users are getting so confused by the account registration process their LLM generated and therefore aren't getting as far as the entering payment details process the LLM also generated, shareholder will care. It will likely take some example cases to get shareholder to be aware that's a risk, so some shareholders are going to find this out after it's done rather than preventing it.

AWS wiped my account of 10 years, says open source dev

Re: I'm confused

I could easily believe that they could screw up weirdly, but at their scale, I would expect that if they did, there would be more reactions. More than one account would have a complaint, the AWS response would be more generic to avoid saying anything relevant to many accounts. So far, I'm not seeing anything like that whether in the news or from other people complaining, and I know multiple people with personal, low or intermittent usage AWS accounts.

Meanwhile, there are a lot of parts to this story which suggest that the account user was perhaps using the account in an unusual way, what with the payment from someone else for some reason which AWS didn't like, and that's not explained either because AWS tends to be happy to accept money and, from the scant details available, was accepting it for a while before deciding not to. Combined with the implausible story of a dry run that didn't work because of a mythical missing language feature which isn't a language feature at all unless you misinterpret a different language feature which wouldn't have helped and is only a thing for the one language they were claiming it's not present in, and I agree with the original post. I think we're missing information from both sides of this story. I wouldn't be surprised if AWS ends up getting some blame if we got that information, but I would be surprised if any of the explanations we've gotten for how AWS screwed up end up being the way it actually happened.

Uncle Sam floats tracking tech to keep AI chips out of China

Re: AI chip as a service

Again, pursuing this as only a thought exercise, that's not impossible. There are many audit systems that do check for the identity of a company, and there are even audits that check where they got money they're now trying to spend. Nothing would prevent someone from trying to mandate that anyone renewing a license gets a full audit first, and the companies to do it exist. Their solution of having a bank validate a customer is much smaller than that. The problem is not creating a financial system that doesn't exist but instead that this would be costly and not necessarily effective because corporate proxies are relatively easy to create.

Re: AI chip as a service

The idea is stupid, but the implementation suggested is probably the closest you can come to it and not be making up technologies that don't exist. A location tracker that transmits is not going to be able to communicate to people to care. One that locks can be spoofed to not or would do nothing if it was simply operated outside China for Chinese users. But requiring cryptographic unlocks on a schedule is something you can implement with a little shim in the controller to be able to transmit them and some microcode to validate them. Of course, in addition to the annoying everybody problem the original comment correctly identified, there's also the problem that this method wouldn't work too well at preventing chips from getting used in China any more than the restrictions on exporting them prevented that, since the people who exported illegally can also proxy activation data. It's mostly a waste of time to try to design the crazy tech magic politicians demand, but sometimes, it can be an interesting exercise anyway.

Network scans find Linux is growing on business desktops, laptops

Re: Coluld be good, could be bad

Not a lot of those are kernel vulnerabilities, but few of the attacks on desktop Linux need to be kernel vulnerabilities either. It often starts with user vulnerabilities which may not need any tech vulnerabilities at all but rely on convincing users to do things that the OS was intended to allow but cause problems. But there's a large stack of software above any kernel which can be attacked and people will do it.

And there already are people attacking in the domains you listed, but the problem will get larger with desktop Linux because the attack surface gets larger. People attack Linux servers incessantly and successfully. Linux-based IoT devices find themselves in botnets routinely, and one category of those that's particularly targeted is networking devices. It happens, and it will continue happening.

Perplexity AI accused of scraping content against websites’ will with unlisted IP ranges

"if someone chooses to ignore robots.txt, all you can do is shake your fist and mutter curses"

Not quite. It's not too hard to implement a thing that enforced the thing that robots.txt can do. It filters on user agent and, if you've said that a given user agent is disallowed a path, send them a 403. You can do that. The problem is that any bot which wouldn't do it voluntarily is not going to make it that easy for you. Well-behaved bots announce themselves and check the restrictions. Badly behaved bots don't just ignore the restrictions, they also tend not to announce their presence. That means you have a much harder problem because you have to distinguish between a bot saying it's a normal user with a browser from a normal user with a browser. That is an arms race that's a lot less fun. I actually do find it fun the first three times, but when I'm trying to get a bot to go away for the tenth time, it's no longer any fun.