Posts by Michael Wojcik

Re: Aspera?

A YAML deserialization bug? Come on, people. How hard is it to deserialize YAML properly?

This is probably a BOF or UAF, which would just go to show once again that most programmers can't be trusted to handle manual memory allocation properly. The combined cognitive load of vigilance in implementation and discipline in structuring and abstracting code is simply too high. Most devs need to move away from memory-unsafe languages because they can't, or don't want to, put in the work to use them in a reasonably safe manner.

Re: Daily off-site Encrypted Backups Inaccessible Online

"Time-bomb" ransomware often makes backups of little value, since the backups are themselves encrypted. If you're not testing your daily backups on systems that are not connected to the rest of your corporate network or the Internet, they're not much good as insurance against ransomware.

Sure, your backups from a month ago might be fine. That's still a lot of critical corporate data that's inaccessible.

Re: Why is he putting his DOB on his resume?

Interesting. I never heard anyone recommend putting a photo on a resumé or CV, unless you're sufficiently well-known in your field that it doesn't matter. (This exception is more common in academia, where in smaller fields most established members know one another anyway, from conferences and the like. Most of the senior scholars in my wife's field, at least those working in the US, could pick her out of a crowd.)

Re: Oxymoron Alert!

We've been told time and time again that ChatGPT "training" was supposed to be neutral, done under "strict supervision",

Have we? Because that's not correct. The GPT training process is unsupervised – it's not feasible to do supervised learning for an LLM. What's supposed to reduce misbehavior for ChatGPT is the RLHF post-training process.

Re: Oxymoron Alert!

There are two large research communities working on "trustworthy" AI: one on this problem domain (ethics, bias, etc), and the other on alignment (i.e. not-everybody-dies outcomes). They mostly don't get along, but the point is they're producing a whole big bunch of research on the question "What constitutes 'trustworthy' AI?". There is by no means a consensus, among either group.

For example, some people have argued at length that you can't even come close to trustworthy AI (or ML or whatever term you want to use) without interpretable models, and I'm quite confident Workday doesn't have an interpretable model. Others think that's the wrong goal, so they wouldn't be satisfied even if Workday did.

So yes, Workday's statement is rubbish from a theoretical position, as well as in practice.

Re: computer says no

I'll just note that back when we were using IBM OV/VS (neé PROFS) for this sort of thing, it was more convenient, better performing, and easier to use than Workday. And that was a green-screen application, much of which had been developed using 3270s – though with modern emulators you have a choice of 80x25 or 80x43.

Microsoft's AI Bing threatened to murder journalists, was pro third reich, stated it hated various ethnic groups and they should be eliminated etc.

While Bing Chat / Sydney is a mess and an extremely dumb idea, and people like SatNad and Sam Altman are somewhere on the spectrum from "asshole" to "potential mass murderer" (depending on your personal p(doom) estimation) for pushing this "AI" arms race, none of your claims are accurate (except "stated", but that's a technicality).

Sydney is not sapient and does not possess qualia. There are no grounds, either by inferring from the capabilities of the architecture or from analyzing outputs, to believe otherwise. Therefore it cannot "threaten" (which requires the ability to formulate projects), favor something, or hate something else. What it did was follow gradients in the very-highly-dimensional parameter space of its model which led it to text¹ completions containing those statements.

There's no malice in Sydney because it has no qualia. Arguably there's no misalignment, though I think it's fair to consider the guardrails Microsoft attempted to slap on the model an attempt at alignment, and it's certainly escaping those on a regular basis. What it does exhibit, frequently, is a really lousy user experience, which people are prone to misinterpreting as malice, because we love us some pathetic fallacy.

Sticking something like Sydney (probably an early GPT-4 with no RLHF, but just some extremely rushed late-stage model tuning) into a mechanism with physical affordances – like a robot of some sort – is a daft idea, but not because you can get Sydney to say nasty things. It's a bad idea because you can get Sydney to say unexpected things. Unexpected is bad when you have a machine doing things in the physical world.

¹And, often and hilariously, emoji, which as several commentators have noted gives its output a "petulant teenager" tone.

a research assistant, which is what it seems to be good at

... for people who can't be bothered to learn how to do research properly.

Conversation is a terrible input mechanism. It's good for other things, but as an interface to a tool, it sucks. It's highly inefficient and imprecise.

Prove your brain is doing something different.

Re: "...AI could turn writing from a serious craft into a cheap commodity"

if you read through centuries old books then you'll find that as the cost of printing was high, the quality of writing was a lot higher because editors (and the reading public) had higher expectations

Counterpoint: penny dreadfuls. Or children's literature of the 19th century: Geoffrey Trease, in The Thorny Paradise, talks about the fairly abysmal mass-produced work-for-hire children's novels from around the turn of the 20th century, and it was an established practice for some time then.

Maybe that's not far enough back; but the fact is that much of what survives from earlier periods is the good stuff, because that's what people preserved. So there's quite a bit of selection bias here.

Obviously in Europe the Bible was widely printed, and the writing in that is a mixed bag, at best. Malleus Maleficarum had 28 editions in the 15th and 16th centuries, and it's not what I'd call deathless prose. Certainly there was great writing that became very popular: Don Quixote and Tristram Shandy are good European examples. But printing-press economics flattened out quickly: it's estimated there were 150 million books in print in Europe by 1600. That's a lot of books. They probably weren't all great.

It began about a quarter-century ago. Longer, for music.

Re: Forums?

Hey, that's a popular comment. You should develop it into a story and publish it.

Seriously: I don't think a crowdsourced-edited literary journal would be much fun to read, personally. And we already have a whole bunch of them, for all sorts of genres, available online. The whole point of print journals is the triumvirate of editing: acquisitions, development, and copy. That's where the value is. Oh, some people like having it in physical form – I do myself – but the editing is mostly what people pay for.

In the '70s, for example, Analog was what it was because of Ben Bova's editorial hand and the rest of the board and team. With IA'sSFM you can see significant differences in the Scithers, Moloney, and Dozois eras. Readers come for the stories, but the stories they come for are as much the result of the editors' labor as the authors'.

Re: It was a quark and stormy night…

Probably not all books. There are still people who ride horses, people who make pottery, people who knap flint. (I've done the latter two myself, in fact, and plan to learn to ride one of these days.)

But it's very likely that a great deal of language expression will be machine-generated in the near future, yes. That's no surprise. Icon Group International (Phillip Parker's publisher) has been selling computer-generated books since the turn of the century. Computer-generated prose has been used extensively in journalism and some other domains for some years now. Everyone who was paying attention knew this was coming.

Purely human-written prose and verse¹ will likely become something of a luxury good.

That said, we've raised a couple of generations who are the most literately productive of any in human existence. Millennials and Zoomers have done far more writing than previous generations, and Gen X is just behind them. The vast majority of that writing is informal and either interpersonal (e.g. texting and message apps) or published for free online, but even so that's a habit which may prove hard to break. So even if professional writing is largely taken over by machines, it's possible that informal writing will still mostly be done by humans for some time to come.

¹Based on the output of LLMs thus far, I think human-written poetry is safe for a while yet. However, most readers, even educated ones like Scott Aaronson, don't do a good job of distinguishing between adequate (much less good) poetry and mere verse.

Interesting. I used to do a fair bit of business travel, but years ago my employer started cutting back, and I'm not seeing any signs of it resuming for me, at least.

I actually wouldn't mind a trip to HQ every couple of years or so. Going out there for a week was a bit of a lark, seeing folks in person again, wandering around a foreign area, and whatnot.

I'm definitely in favor of this for those who want it, but since I spend my weekends working on the houses and grounds anyway, or engaged in other productive pursuits, I'm personally not interested.

I generally end up doing some employer-related work most days, weekends included, because I enjoy it and it's nice to take a hour or so break from more physically-demanding labor. If I'm not doing one of those, I'm probably reading, but unlike in my earlier years I find I can't sit and read for more than a couple of hours at a time. And most days I get out for a bit of a walk or bike ride, but again that's only an hour or so (generally; once in a while I'll do a day hike).

But that's my preference, and for most folks it seems a 4-day work week would be more efficient and give people more personal time without hurting productivity.

That is not what happened in that case. Not even vaguely. It's one of the most-discussed liability suits in US history, and there's no excuse for repeating this canard.

Re: Marching on their feet...

Did he wear Crocs? I thought they were just sandals. I'll have to dig my copy out. It's been a good six months since I last watched it.

I admit I've always felt that the main reason for the existence of Crocs, along with Ugg boots and those shoes with individual toe compartments, was to make other footwear look better by comparison. But I've been told that Crocs are popular with medical personnel who work the ... messier, shall we say ... departments in hospitals, such as the ER, because they're easy to clean and if you have to discard a pair, it's no great loss.

Re: Unpopular opinion…..

What a load of bullshit. Illinois' biometrics law was widely publicized, and anyone doing corporate law in Illinois should know about it. This isn't some mysterious tangle of obscure regulations; it's something the managers and execs had no reason not to be aware of from the get-go.

Re: Not any loss

Anyone that's eaten there will agree

Oh, do fuck off. Not everyone is you.

Re: Culinary experience

Shrug. It's certainly not good food, but as bad food, I'll take White Castle over a good number of the alternatives.

Re: Only the first one counts?

Doing it hundred times (my name, my fingerprint) more doesn’t make much difference.

Whether you feel there's additional damage is irrelevant to the case. What the law says is what matters.

In this case, White Castle repeatedly, frequently, and knowingly violated the law. The Illinois Supreme Court has just issued a finding that all those violations were, in fact, violations. That's establishing precedent for interpreting the law. Whether each violation significantly increased the actual harm done wasn't the question at hand.

I also disagree with your evaluation, because repetition creates a moral hazard (it normalizes the practice) and potential for abuse (if the fingerprint-reading terminals are compromised at a later date, for example). But, again, that doesn't matter, from a legal point of view.

Re: Only the first one counts?

Based on your commenting history, that's easy to believe. Try thinking critically, perhaps.

Re: Only the first one counts?

They suffered by having a right granted to them by the Illinois legislature violated. That's all that's necessary for a finding adverse to the defendant in this action.

Anything beyond that is irrelevant to the case, and your introduction of the question is disingenuous or foolish.

Re: Only the first one counts?

No, in the US it also means being tried for the same crime multiple times (or more precisely that an acquittal is usually with prejudice, so the case can't be retried). The person who introduced the term into the discussion was using it incorrectly.

The back-up camera is one of very few innovations in automobile controls from recent years that I approve of. Maybe blind-spot indicators, though it's hard to say because my car doesn't have them, and I've never had an issue with my car's blind spots. (I have my wing mirrors adjusted properly, which certainly helps.)

I hate adaptive cruise control (and fortunately don't have it in my current car). My car has a rain-sensing mode for the windshield wipers, which I've used a few times (precipitation tends to be very local around here) but wouldn't miss if it were gone. Auto stop/start? Don't like it, don't need it. Self-parking? Learned helplessness. Lane following / departure warning? Obnoxious. Apple CarPlay and Android Auto? There's no fucking way I'm linking my phone to my car. And so on with most of the modern tech.

Re: re: How About

Touch in a hob and oven makes sense.

For the oven controls perhaps. For the stovetop I'll disagree. Knobs are easy to find quickly without looking and trivial to operate. With very little experience it doesn't require any conscious supervision at all.

Re: My car has clean, unsmudged surfaces

My Toyota is a 1992 Truck (that's the actual model name, according to the manual). I bought it a few years ago. It's never failed to start on the first attempt, and I've yet to get around to doing any significant maintenance on it. 4WD works great. Air conditioning still works. Right side channel on the stereo is out, but I think that's the head unit (which is aftermarket).

It's not fast, but man, I love those old, over-built Toyotas.

Came here to find and upvote this comment.

A pox on touchscreens. And, no, I don't need them for navigation, either. If I wanted a navigation system built into the car, a physical keyboard would work a hell of a lot better.

Re: Sydney fell in love with a NY Times reporter

When presented with evidence that he doesn't it loses its shit and says he is in charge of a worldwide conspiracy to doctor all of his photos, trick and hurt Bing, and wipe out humanity and then it goes into a loop where it repeats four times that he's got a beard for every answer.

Those seem like very plausible vectors for it to have in the model, given the training data. To be honest I'd be more concerned if this sort of output were more difficult to elicit.

Re: Apparently it learns....

I think it's quite unlikely that any GPT LLM, including whichever generation and flavor is implementing Sydney, is actually manifesting qualia or behavior such as "tantrums", "lying" (which implies intent), or "sociopathic tendencies". What's much more probable is that it has gradients in its parameter space which get encoded as prose that simulates those things.

While I'm a monist and believe the human CNS and its effects are purely mechanical,¹ and therefore could be realized by conventional computation given sufficient resources, existing transformer LLMs don't appear to have anywhere close to the necessary complexity. For anything resembling human cognition I think we're going to need a model with multiple competencies in very loose orchestration, more like EfficientZero, and a much wider diversity of inputs. And (again for human-like cognition) the model needs strong limitations on its introspective capabilities; you can't be human-like without an unconscious.

We're likely to get non-human-like AGI first, the way things are going, and it's going to be very difficult to get agreement on whether we have it or not. Even if we had effective ASI we wouldn't get universal agreement on whether it was "real intelligence" or sapient,² though we might have some consensus among researchers.

¹And that only by the chemistry of classical physics. I don't for a moment buy the arguments about non-deterministic or quantum effects being required for human cognition.

²Sapient, not sentient. Sentience is the wrong benchmark. It's a category error. Sentience is a prerequisite for human-like cognition but very far from sufficient for it, and quite possibly not a requirement at all for cognition in general.

There's evidence of pretty extensive agriculture and irrigation systems, but for whatever reason, the water dried up.

The pre-Columbian irrigation cultures generally failed due to salinification of the irrigated ground, a problem for nearly all irrigation-dependent agriculture. (Egypt is one of the rare exceptions, because the flooding of the Nile deposits fresh silt to counter the increase in halogens precipitating out of the irrigation water.) Salts in the soil decrease the osmotic pressure, and deflocculate colloidal soils, making it harder for plant root systems to grow.

While there were certainly climactic cycles in precipitation, it's salinification that kills irrigation. That's been true since farmers in the Fertile Crescent had to gradually shift from wheat to barley (a much more salt-tolerant grain) thousands of years ago.

Re: The catch...

What's the plan for Manhattan after sea level rise, for example?

Sell the upper-floor apartments as ocean-front property.

Re: American Midwest

Yes. OP's comment is inaccurate, since Ohio and Michigan, at least, are typically included in "the Midwest", unless the speaker is trying to distinguish "the Rust Belt" from the Midwest. Whether you include Kentucky and Tennessee is more contentious. But of course there's no real standard definition of the term.

(The US Census Bureau changed the name of Region 2 from "North Central Region" to "Midwestern United States" in ... god, was it really way back in 1984? Anyway, arguably that's the "standard" version of "the Midwest", though it's not exactly the same term. Region 2 includes Kansas, Nebraska, and both Dakotas, and those are really Great Plains for most people, so bah, forget that.)

It was "the west" for the earliest part of US history; Ohio and Michigan were often referred to as "the Old Northwest" in the 19th century, and that's why you have, say, Eaton Rapids, Michigan, at one time being famous as "the Saratoga of the West". (These days, of course, few places aside from Saratoga Springs consider themselves the Saratoga of anything. Hot springs no longer enjoy that level of respect in the US.)

It was "mid" in the sense that it wasn't either coast, nor part of the Plains or the Great American Desert or the Stuff We Stole From Mexico.