My bad, I mistook the previous response to be for me, not a different response.

Are you arguing the US system, with its horrific track record of recidivism, is an improvement over e.g. the Scandinavian model that's focused on reform? The results are not on your side, in that case.

Neither would the data be, if you were to look it up, with regards to spending time on preventative measures to prevent youths from turning to crime in the first place.

But do you punish a person, with no previous history of epilepsy, for having an epileptic event and killing someone? You don't, right? Around here, you do lose your right to drive, and only get it back after a number of years with no more incidents, which makes sense. But we don't blame, and we don't punish.

That's not far form an example I've heard from Sapolsky. In that case it was a previously "normal" person suddenly starting to do horrible things. On examination, a brain tumor was discovered, and upon its removal he reverted back to his usual self. Years later, his bad behavior returned, and as it turned out so had the tumor.

This is a cause and effect we can all understand. It's the concept that everything is like that, just distributed over a whole lot of time and compound causes, that we have a hard time grasping. E.g. according to Sapolsky (as far as my memory can be trusted) it has been shown that the greatest predictor of outcome in parole hearings, is the time elapsed since the decision makers last had a meal. A particular scent in a room affects how homophobic you are. The hormone levels (stress hormones, I believe the example was) in your mother during her pregnancy, impacts your brain size. It just goes on and on, from seconds before you make a decision all the way back through generations and millennia.

I've heard mention of an example where, during open brain surgery, the surgeon poked at a part of the brain that caused the patient's leg to twitch. When asked why his leg moved, the patient came up with a story for why he'd decided to move his leg. Another example was a split brain patient, where an image was shown to one eye, causing him to use his other hand to draw a depiction of the image, and when asked why he'd drawn it he came up with a whole story of how he'd seen something on the drive over and he remembered that now and wanted to draw it. fMRI scans during decision making show that we make decisions up to several seconds before we become consciously aware of having made them.

The evidence would suggest that this is what our consciousness does all the time. It does not determine, it observes and rationalizes.

For me, what really sealed it was physics. We know the stuff our brains are made of, and it obeys the laws of physics just like everything else. Everything in physics is cause and effect. Given a certain state of a system, it's entirely predictable what will happen if we introduce something into that system. It's only complexity that means we can't practically calculate the outcome. No one would argue that if you toss a rock down a hill, it has a choice in how to bounce. Most would agree that if you could reset every atom to just what it was, and throw the exact rock in the exact same way, it would bounce and roll in the exact same way every time. The only way it would change, would be by changing some parameter in the system. Given that our brains operate within the same set of physics, there's just no room for consciousness to have any say in what goes on. The only reasonable explanation is that it's a thing that arises out of, rather than has any influence over, the brain's activities.

At first glance quantum mechanics might look like a potential savior, but it's hard to see how. It's random, so if it had a significant effect on our brains' outcomes, we'd not function very well. Even if it did, how would adding a quantum coin flip into decisions make it any more of a conscious choice?

I had a rough year or two trying to find room for free will anywhere, because I also was not buying it. But based on the evidence it seems the only rational conclusion is to accept that we are nothing but complex deterministic machinery. In principle entirely predictable, given our state at any given time and the inputs we are subjected to. It's just such a complex system we can't help but see it as otherwise. For me, I find that it takes a toll to try to keep this conclusion in my mind. So, for the most part I ignore the knowledge and live my life like Cypher wanted to, in blissful ignorance. But, sometimes I find it quite useful. Both in increasing my understanding of others who've had a less fortunate life than mine, and as a humbling agent when I remind myself I really had nothing to do with the outcome I'm currently busy patting myself on the back over.

(For the younger crowd, that was a Matrix reference in the final paragraph.)

That we have no conscious choice is not the same as it serving no purpose to discuss our justice system. Our conscious decisions are illusions, but that a mass of humans discussing things affects the decision making process of all the brains involved, is very much in line with what he's saying. We just have no conscious control over the outcome.

Put another way, we are free to "choose" the flavor of ice cream we prefer, but we have no control over what we prefer. That the choice is an illusion doesn't mean everyone should stop enjoying ice cream or stop picking the flavor they feel like, nor that it's important that we never phrase a sentence without including the caveat that we understand we have no real choice. Our brains still think and make their inevitable decisions even if if our consciousness is a mere observer rather than the controlling agent we feel it is.

With regards to criminal justice, as I understand it his argumentation is pretty much to remove the imagined agency from a criminal's action. If a car kills people because its brakes are bad, we fix the brakes or make sure no one drives it. If it keeps on happening we might look at how cars are built, or institute mandatory maintenance programs. We don't talk about how cars should make better choices.

His argument is that it makes no more sense to do that for a human than for a car, because there's no more agency there. It doesn't take much thought to see the point. If you stop viewing crime as something that people choose, and instead take the perspective that it's something that happens to people based on past events, you skew the handling of crime from piling on more police and building more prisons, to preventative programs. Succeeding at the latter is a much better outcome for everyone, and it has been shown to work. Brains are programmable, and how we structure our society is a large part of what kinds of brains we end up with as inhabitants.

I watched a whole lot of videos, and read a lot of articles, and none of them made a claim like yours. They all explained, in varying levels of details, how dumb LLMs are when boiled down to bare bones, and at the same time how impressive they can be. Not to mention that we can't entirely explain how they arrive at the output they do.

There's a video available on something I think was dubbed "glitch tokens". Some examples were a few reddit usernames that were unutterable by GPT for a while. Asking it to repeat the name to you, would cause it to do the LLM equivalent of having its mind blown. The reason, or the hypothesis I remember at least, was that the token had made it into the model, but the source was later removed from the training data as it was from a counting subreddit (each post = a number one higher than the one before, and that's it) and deemed not useful as training data. So the model now included a token that occurred nowhere in the training data. Why would that cause an issue, when all you're doing is asking it to repeat the exact string you gave it? If it works with hitting random keys on the keyboard, what made those specific words break it? It boiled down to LLMs doing what LLMs do. Turns out strings of text do not necessarily get its own token, that only happens for really common words, more rare cases are built from a sequence of common tokens that end up forming the full word (or random string). The quirk was that these usernames did get their own full unique token. So now you have a huge 10000 or whatever dimensional distribution matrix thing going on, with this lone token that's not really connected to anything because there's no training data for it. Does the model "understand" that this is a word it has no clue where is from or what it refers to? Nope, it just keeps doing what it does, trying to match it to related words (of which there aren't really any, so presumably it has to pick from a huge set of equally likely alternatives) and the result was entertaining hallucinations like long sequences of "I say 'e' You say 'i'" and such, or a definition of a seemingly random word. Because, again, it doesn't actually have any understanding. It's equally impressive, as it is (still) dumb as a stump. Don't be fooled by the fact that it outputs sensible stuff a lot of the time, and we can't really explain how it does it.

I guess we've taken this as far as it can go. I still disagree with you, but enjoyed the conversation.

For the record, I wasn't posing the question to trick it. The original reason was because someone claimed that we did have residential gas lines, which I was certain was incorrect. When regular searches turned up nothing useful, I tried ChatGPT. Sometimes it is really good at digging up info that's hard to find via regular search engines, and its response usually gives sufficient data to go on to find credible sources.

I also don't doubt the responses will become better. They have a planet of people giving these things reinforcement learning now, after all.

As for it knowing how to go into detail, or that it has some underlying concept of the questions asked, I believe you are wrong. I'm not aware of any other process than the one I described. It just includes the previous response in the next prompt and goes from there. But I don't have anything to back that up with other than that it seems like something that would have been mentioned in the material I've consumed on the topic.

For all the doomsday prophecies regarding this technology, my biggest worry is that so many people genuinely believe these things are capable of actual thought and understanding. It leads to a misplacement of confidence that can cause a lot of problems. Sort of like how some people do not understand the distinction between the terms "autopilot" and "fully autonomous".

Oh my.. so, I decided to rerun the query regarding gas infrastructure, to see if anything had changed. It was obviously hallucinating the answer, so I asked it for sources. Of course, it said, and produced a list. I checked all of them. For context, my query was regarding a city in Norway that's had hydro power since the late 1800s.

- A link to an Aikido club. It claimed this was the memo of a city council decision from 1992, to discontinue the gas infrastructure.
- A newspaper article about a debate regarding China and freedom of speech in Hong Kong
- The Wikipedia entry for the government of Norway
- A newspaper article in a US newspaper about how San Diego is looking to phase out natural gas by 2035.
- A research paper on the impact of draught in Europe

I do understand why it's getting it so wrong. There are no sources out there talking about a city's non-existing gas infrastructure, nor a plethora of questions about it with negative answers. But all the words are, of course, used. So the output is semantically sound, but its contents absolute gibberish. Because it has no concept of the question that's being asked.

I agree, it can be quite impressive. The problem is that it will answer with the same level of assuredness and often, at least at a glance, sensibility even when what it has cooked up is absolute nonsense, because any actual understanding is an illusion.

I once asked it to explain the usage of two words (similar sounding, spelled differently, often confused in writing) in sentence constructions, upon which it proceeded to give me what looked like well thought out explanations and examples. But, as it turned out, none of the examples it provided used either of the two words. I asked it about a city's gas distribution infrastructure, upon which it gave me the whole history of it, years, company names etc. The only issue was that it never had a gas infrastructure.

It's important to remember that, behind the scenes, what ChatGPT does is find the next word that is a reasonable continuation of the preceeding words, with some randomness thrown in to make it feel more human and not answer the same exact thing to the same prompt. It doesn't even really know the words, the first thing it does is convert those to tokens. Everything after that is matrices and transformers and all the things I have no understanding of. But, at the end of it, a "this is a reasonable next token, given the training data" is produced, which is converted back into a word. This is added as the first word of the response to your prompt. Then it does the same thing again to find the next word.

The fact that it ever produces anything sensible at all, is really impressive. That it seems so real and human in doing so, is awesome. I love talking to LLMs. But I use them as a starting point, then I do research to confirm. Because I know it's just building sentences based on statistical distribution of words, with no understanding whatsoever of what it's asked or what it outputs. If I'm using them for anything that matters, I don't take anything they produce at face value.

I have many issues with how the abilities of this supposed "maintainer of code" comes across based on these citations, but let's chalk that up to a need for brevity and me being too lazy to RTFA.

A more important issue I have, is he seems to believe ChatGPT understands how WordPress handles hooks. Unless something's drastically changed in how ChatGPT functions, that's not at all what it does. It answered what people have previously followed similar strings to the question with. That's all.

Not that I don't think LLMs can be helpful tools, but for the foreseeable future it seems firmly seated in the "make a suggestion or two and have a human with knowledge take those into consideration" department, as well as a slightly more fancy snippets engine. I'd not worry about my job anytime soon. Then again, I'm old, my odds of being retired before AI takes over programming is above average.

I really don't. I was sharing my anecdotal reasons for sticking with WinRAR – in response to someone offering not even an example of what made one better than the other – not pretending to have produced a research paper comparing the two. I'm sure someone's done that at some point, but I don't have that high on my list of things to spend time on in life. My very limited testing is plenty for such a trivial decision for me.

I will add that I have previously spent some time tweaking the settings of both programs, with the same conclusion.

It's a thread about WinRAR and your subject said "WinRAR? Why?" so, yes. I'm talking about WinRAR. I should also add that an extra nail in the coffin is that 7-Zip's very slight compression advantage isn't a constant in my experience, sometimes WinRAR does better on that point as well, but the speed disadvantage is pretty much consistent.

I'll definitely keep 7-Zip in mind for when the year of the Linux desktop arrives. Or the heat death of the universe, whichever comes first. :p

I've found 7-Zip to sometimes produce slightly smaller files, but at a huge time penalty. Case in point, I just did a test where WinRAR produced a 349MB result in 21 seconds, while 7-Zip ended up at 347MB in 38 seconds. That's slightly superior on one point and far inferior on the other. For my use cases, the speed is much more relevant than that tiny storage win.

Over the years I've found that result to be consistent. Whenever I spend a little time doing some tests I end up concluding the switch isn't worth it and stick with WinRAR. Would I have paid for WinRAR today if I was already using 7-Zip? Unlikely. But that train left the station some 15 years ago.

They had the foresight to make it so simple that really anyone can use it. None of the elders I know have issues with it. In fact, they have a lot more issues using their online banking on a computer. The amount of information and menus there is bewildering compared to a very focused "does this one thing and one thing only" app.

I never use my phone for any actual typing. It's just too slow and annoying. If someone sends me a text, unless it's time sensitive they'll generally get a response once I'm home and can use phone link on the computer to send a response. I know people who will literally have a laptop on their lap, and still use their phone for Discord. Boggles my mind.

We mostly use a bank endorsed app here these days. Transfers are free for sensibe amounts (a number of times more than what any sane grandmother should be giving her grandkids). There's even optional gift box popup and on-screen confetti and such available.

I did a search and apparently checks _are_ still a valid form of payment. Or at least it was back in 2017, when that article was from. It's just that no one uses them. I found a story about a guy who got a course in selling check scanners, and in the two decades since he'd sold zero. At this point I'd half expect a bank clerk to have to call a retired colleague to figure out what to do if someone were to show up with a check.

I can't answer that as I don't know how backwards individual states are considered, but it's near Orlando, Florida.