* Posts by Michael Wojcik

12132 publicly visible posts • joined 21 Dec 2007

95% of NFTs now totally worthless, say researchers

Michael Wojcik Silver badge

Re: Just a nice site to browse...

Checking that site is one of the highlights of my day, to be perfectly honest. Not the only highlight, but whenever I see Molly's posted a new piece or two it brings a smile to my face.

I feel a bit guilty about not being sadder about the tremendous waste of resources – energy, capital, inventiveness – being wasted on cryptocurrency and DeFi and the like, and about how it's used to enable criminal proceeds and bring hard currency to North Korea and so forth. But I can't help chuckling at Every. Damn. Story.

(It's much the same with cryptocurrency research. I think blockchain is a Merkle Tree for babies, and cryptocurrency is an abysmal idea. But I've read a bunch of really fascinating papers about this area. Stupid in practice, intriguing in theory. It's the theology of computer science.)

Michael Wojcik Silver badge

Re: a revolutionary new model for ownership

It's a weird concept of "revolutionary", since getting people to pay for nothing is hardly new.

Michael Wojcik Silver badge

Of course there were a number of savvy folks who hot-potatoed them and sold them on as soon as the price went up (or perhaps even, as you suggest, sold them at a nominal loss to launder ill-gotten gains, though I'd be shocked – shocked – to hear such a thing was done by any member of the upstanding Trump family).

And there were a whole bunch more people who took the expedient route of stealing NFTs from poorly-secured wallets and immediately selling them on, then tumbling the cryptocurrency proceeds, and cashing out as much as they could. That's, like, one out of three stories on Web 3 is Going Just Great. So it's not like NFTs don't have any uses.

Michael Wojcik Silver badge

Re: “Now”?

At least Beanie Babies you could throw at people. Hell, with enough of 'em, you could make a bean-bag chair. Can't do that with BAYCs.

Michael Wojcik Silver badge

Re: Shocked

"I've been disappointed in get-rich-quick schemes before, but here's a scheme that will get me rich, and quickly!"

Michael Wojcik Silver badge

Re: 95% of NFTs now totally worthless

The same way other collectibles with little or no use-value get elevated exchange-value. The information on baseball cards, for example, is readily available elsewhere; it's hard to see what use-value they have, unless you use one as a coaster or to make a sound effect for your bicycle. Their value is all in exchange.

Same thing with, say, collecting stamps (particularly canceled ones, or others that couldn't be used for postage).

That doesn't make NFTs not stupid, of course. At least with baseball cards and stamps there used to be, and presumably still is, a pretty large community of collectors, and there were social events like sales and swap meets organized around collecting; and they're physical objects so you can perform physical activities with them such as putting them into albums. NFTs lack even those virtues.

Now IBM sued for age discrim by its own HR veterans

Michael Wojcik Silver badge

Re: Now IBM sued for age discrimination

Krishna's 60 or 61. Born in 1962.

Michael Wojcik Silver badge

Re: So the Decision makers get the shaft

Karma is a bitch !!!!!!

Counterpoint: Karma is my boyfriend [Swift 2022].

Michael Wojcik Silver badge

Re: PART 1625—AGE DISCRIMINATION IN EMPLOYMENT ACT

Fortunately, Kevin Underhill has written a classic article which explains everything you need to know about the difference between civil- and common-law systems, "Way Less Than You Need to Know About the Civil- and Common-Law Systems".

At least regarding bee swarms.

Michael Wojcik Silver badge

Re: It's time to help IBM

Some people associate IBM with, y'know, mainframe systems in the S/360 lineage. That is still very much an IBM business. Kind of odd for a Reg reader to forget that one.

They also have this goofy OS called "Red Hat Enterprise Linux" that I gather some folks use.

Michael Wojcik Silver badge

Well, yes, that's essentially what the doctor wrote in the post you're responding to.

Of course, even if fines were levied against the C-suite-sitters personally, they're likely indemnified by the corporation, and it probably backs that with an insurance policy. At least in the US. (We sometimes see that in academia when, for example, a dean is a defendant in an employment-rights case and loses; the school, or its insurer, ends up picking up the tab.) So the court would have to order that contractual provision vacated, and that would get appealed and quite possibly overturned – and perhaps the entire lower-court decision, or at least penalty phase, with it.

In America executives of major corporations have a lot of power and freedom to abuse it. Sometimes they overstep – Elizabeth Holmes, Ken Lay, Bernard Ebbers are examples. But most of them get by unscathed.

Michael Wojcik Silver badge

Or, if they've proven themselves really incompetent, into corporate-officer sinecures like board seats and fluff positions in industry lobbying groups. Look at where Apotheker is today.

Michael Wojcik Silver badge

Re: HR...

get rid of the old iron and nobody is writing new code for it

Oh, they most certainly are writing new code for z systems. Supporting that effort is a substantial chunk of our mainframe-offload business.

EU right to repair updates pass latest hurdle

Michael Wojcik Silver badge

Electronic parts can wear out. Capacitor electrolytes can degrade, for example. Conductors and insulators in resistors can degrade due to heat dissipation, changing the resistance. Metals can migrate across bimetalic joints.

Sure, there are plenty of people on Hackaday and the like coaxing old computers to life, but often they refurbish basic electronic components in the process.

And phones have mechanical parts too. They have cases and screens. They have buttons – I've yet to see a phone with no physical buttons at all. There's that USB charging port and its contacts.

Phones won't last forever. Forever is a long time. Personally, if manufacturers were forced to support them even for, say, 5 years, I'd be happy to see it. 10 years would still be reasonable. But "forever" is nonsense.

Michael Wojcik Silver badge

Re: That's Phillips screwed then.....

That would be nice.

I'm not particularly fond of USB-C – mechanically it's not much more robust than earlier USB variants, particularly not compared to e.g. dead-simple barrel connector, and bad parts are a concern. (Knowledgeable consumers may make an effort to avoid those, but if we have to rely on knowledgeable consumers, we might as well give up now.) But at least if everyone's using it for everything, that makes charging much easier for consumers.

Authors Guild sues OpenAI for using Game of Thrones and other novels to train ChatGPT

Michael Wojcik Silver badge

Re: All authors started as readers

The parody part of the Fair Use exception in Title 17, and jurisprudence around it, are so, so much more complicated than that.

Michael Wojcik Silver badge

I am sure the courts will find your anecdotal argument completely persuasive. Indeed, we should probably just junk the whole judicial system and ask instead what veti can do.

Michael Wojcik Silver badge

That is pretty much how I thought it would be 'reading' the training material. Nowhere in the AI's 'brain' is an exact replica of the original work, just an essence.

To be perfectly frank, this sort of gloss is not terribly meaningful. It's far enought from any actual technically accurate or precise insight into how transformer models work that it's not very useful for drawing conclusions, practical or legal.

Is there, in the model, a sequence of bits that correspond to the text of a given novel-length work in some encoding that the model can reasonably be held to have an algorithm for decoding into, say, Unicode?1 It's true that's unlikely.2

However, particularly for works that the model has seen often enough in the training set to somewhat overfit on, it's entirely possible that there are positions and gradients in the parameter space – which is very high-dimensional, after all – that reproduce substantial parts of a given work, and possibly all of it.

Any CTT-compatible computation can be reduced to some form of compression (just as it can be reduced to Boolean algebra, or the operation of a Turing machine or a Post machine, etc). What you refer to as "essence" should be called "information entropy", and LLMs (crude though unidirectional transformer stacks are) are capable of storing quite a lot of it – how much depending on how large the model is, the pre-compression parameter precision, how much compression is done, and so on. It's not necessarily going to be true for any given input (assuming it's much smaller than the model size) in the training set that not all of the information entropy in the input will be captured by the model. And, of course, the output doesn't have to be complete, or bit-for-bit exact, to be infringing in the legal sense. An ALL-SHOUTY copy of the first half of A Game of Thrones with Ned Stark referred to as "POOR LITTLE NEDDY" throughout3 would still be viewed dimly by the court.

And this last points to the real crux, which is that copyright law (i.e. Title 17) in the US, and the courts adjudicating upon it, are unlikely to care much about what is "stored" by an LLM and how it is represented. They're going to care about actual and plausible effects. Will LLMs have a chilling effect on creator revenues, and if so to what extent is that an actionable harm under the law? Can the LLM guardrails against reproducing portions of copyrighted works plausibly be bypassed, now or in the future, and how infringing would the output be? Is substantial information from copyrighted works incorporated (in any representation) in the models, and if so is that incorporation transformative or otherwise allowed under Title 17?

1It should be obvious that trivially a given LLM has a bit-sequence corresponding to any given extant novel under some arbitrary encoding, because LLMs are large enough to represent any single given novel, and you can just invent such an encoding on the spot. Thus we have to distinguish between arbitrary encodings and reasonably plausible ones.

2Not impossible, though, given the size of these models, for some relatively small set of works, particularly given the low information density of natural languages. Model compression would tend to eliminate these, but if you figure that, say, Moby-Dick has around 222 bits of entropy – quick estimate by deflating the plaintext version from Project Gutenberg – and a GPT-3-class LLM weighing in around, oh, 233 bits, then if those bits were evenly and randomly distributed (they're not, but let's pretend for a moment) you'd have around a 1-in-2048 chance of finding a target bitstring with the right information. Assuming I got the arithmetic right. Of course you'd need to decompress it, so that's not really a fair estimate.

3Actually, does Poor Little Neddy survive to the halfway point? I don't remember.

Michael Wojcik Silver badge

Re: It doesn't store the original, just 'interesting' features of the original

it is very specifically designed not to allow the text to be retrieved

"It" (i.e. ChatGPT-x and other major unidirectional-transformer LLMs currently in vogue) most certainly is not "specifically designed" to avoid reproducing copyrighted work verbatim. That is a guardrail tacked on very late in development. Frankly, judging by the published research, neither OpenAI nor any other LLM team have any idea how they would "design" a transformer model to avoid reproducing copyrighted material. That's a difficult outer-alignment problem.

Michael Wojcik Silver badge

Re: It doesn't store the original, just 'interesting' features of the original

I read a book and being of average intelligence if asked, could provide a reasonable synopsis...

LLMs fail the commercial-use test. LLM vendors are seeking to monetize their models, so if their models display similar behavior, that's very distinct from what you're describing.

LLMs fail the financial-harm test, if LLM activity does indeed reduce commercial demand for the existing or future work of the creators whose works they've been trained on. That's also very distinct from what you're describing.

Michael Wojcik Silver badge

unless your prompt is as complex as the original work

In a strict, information-theoretic sense, this is almost certainly wrong.

In terms of information entropy, considerable entropy is stored in the model; the prompt just has to elicit it. It's equivalent to a form of dictionary compression with a very large dictionary. Therefore there's almost certainly a prompt which contains less information entropy than the source document which can elicit the source document from the model.

As a practical matter, it is almost certainly possible to identify recurring template phrases in the source document that can be elicited multiple times, with replacements, in the correct locations, using a prompt shorter than the total length of those realized templates. That's one mechanism whereby the prompt becomes both absolutely shorter in length and less in information entropy than the source document.

Would creating such a prompt be easy or useful? No. But it's not true that the prompt must have at least as much information entropy as the desired output, as it would with, say, a general compressor that does not contain any prebuilt dictionary.

Michael Wojcik Silver badge

Re: A Song Of Ice and Fire

Dude has a railroad to run.

(Yes, the subhead of that article is deliciously ironic.)

I haven't ridden the trains m'self, but by sheer coincidence my wife and I happened to be in Santa Fe (a rare occurrence) and in the Railyard District (an even rarer one) on the day and time of Sky Railway's maiden journey. We hadn't heard about it yet – news from far-distant Santa Fe (nearly 80 miles!) takes a long time to reach the Mountain Fastness – so at first we had no idea why there were all these people and news crews milling about. Then we saw the train, and it was Cool.

We also shopped at GRRM's bookstore that day. I think that's where I picked up Jo Walton's What Makes This Book So Great.

Back on thread... Personally, I'd rather complete SoIaF myself, for myself, than read something a transformer comes up with. The whole point of the transformer architecture is minimizing information entropy, aside from whatever the temperature is set to. It'd give you the most predictable ending.

Uncle Sam names three Amazon execs as Prime suspects in subscription ripoff case

Michael Wojcik Silver badge

It's quite possible different accounts get different UXes. A/B testing is very widely used on large commercial websites, for example, and there's nothing to stop Amazon from having different UI controls, layout, and workflow based on region, account details, shopping history, or various other factors.

aerogems mentions the extra Prime nag for people Amazon thinks are students. My account isn't from the .edu TLD – it's a Gmail address – but Amazon has always dumped me into the "special student offer" page regardless. Something in that algorithm decided I'm a student and refuses to give up the idea. (I earned my most recent degree ten years ago.)

I have no intention of ever signing up for Prime; in fact, I try to avoid buying from Amazon as much as is feasible. Doesn't stop them from trying.

UK Online Safety Bill to become law – and encryption busting clause is still there

Michael Wojcik Silver badge

Re: At last!

Certainly we seem to have a low tolerance for authoritarian idiots.

Michael Wojcik Silver badge

Re: Why?

What Apple demonstrated was client-side scanning, which is not what Clause 122 calls for, at least according to the stories I've read. (I haven't looked at the bill myself, as I'm not a UK citizen or resident and have bigger things to worry about.)

Michael Wojcik Silver badge

Re: Blighty seems determined to cut itself off from the rest of the world

Yes, and we also remember Matt Blaze breaking the LEAF checksum.

As TikTok surveils staff's office hours, research indicates WFH is good for planet

Michael Wojcik Silver badge

Re: Well, then . . .

Alas, they would, but they're too busy ranting in the comments.

Michael Wojcik Silver badge

Re: What I miss from working in the office

I admit that if my employer had an office near me – even, oh, an hour away, let's say (and possibly further if public transportation were available) – I'd go in once in a while, if only for a change of scenery and a chance to interact with some colleagues who aren't on any of my teams and aren't routinely communicating with me for other reasons.

But they don't, and going to an office near me wouldn't let me see any teammates because they're all at least 1400 miles from where I live. The closest person that I even occasionally discuss technical issues with is ~350 miles away, and he's WFH too. The nearest office is ~285 miles but I've never been there and don't know anyone who works there. The nearest office I have been to is ~650 miles.

Aside from a two-year period when I did live near an office, I've worked remotely for over 30 years. I've enjoyed my rare visits to the office, but I think mandatory attendance is stupid. And, yes, my employer has also instituted a mandatory-attendance policy for people who aren't on a remote contract, and I find their reasoning utterly unpersuasive and completely unsupported by any sort of methodologically-sound research.

Michael Wojcik Silver badge

Re: Free Haircuts

"Homer, did you stick your head in the Shine-O-Ball-O?"

I'd like to point out that the article says nothing about "haircuts" being limited to the head.

Signal adopts new alphabet jumble to protect chats from quantum computers

Michael Wojcik Silver badge

Regev's algorithm

Right from the abstract: "It is currently not clear if the algorithm can lead to improved physical implementations in practice."

We frequently see new quantum algorithms being published. Sometimes they're simply wrong – someone made a mistake. Sometimes they rest on unproven assumptions, such as the assumption that QAOA is good for anything and not just a huge waste of time.

Now, Regev is a legitimate researcher and reliable, and Aaronson referred to the paper as exciting. The algorithm looks good so far, as far as I can see from informed commentary (it's beyond my level of expertise to critique it), and it does lower the complexity of the quantum part from O(N2) to O(N3/2), which is indeed smaller – significantly so once N gets large. (Regev's paper actually talks about quantum circuits of size proportional to O(N3/2), not time, but since you have to set up the circuits time becomes proportional to size, I think.) But it's still going to require a lot of fast qubits to work. And it depends on a number-theoretical assumption which remains unproven.

So why even mention it? Either Shor's Algorithm is already sufficiently bad news for RSA, DH, and classic ECC (on the assumption that one day large QCs become physically and economically feasible), or irrelevant (if that assumption doesn't hold). Regev's algorithm might advance the traditional-problem-asymmetric-cryptography doom clock forward just a bit, but in practice it doesn't really make a difference. We're still far away from it being practical or sensible to break most traditional asymmetric cryptography, and by the time we get there, most traditional asymmetric crypto will have been replaced by PQC just in case.

Michael Wojcik Silver badge

Re: Weird structure?

The post-quantum cryptographic algorithms are believed to be QC-resistant because they use problems that 1) are believed to be trapdoor, i.e. worse-than-polynomial time to solve but polynomial-time to confirm;1 2) don't resemble any problems known to be in complexity class BQP, i.e. the ones amenable to a quantum-computing speedup; and 3) are believed for theoretical reasons not to be in BQP.

CRYSTALS-Kyber is in the family of lattice-based PQC algorithms. It uses learning-with-errors over modules. LWE-type algorithms go back at least to the original NTRU proposal, though NTRU and most others were LWE over rings rather than over modules.

There are tons of papers and blog posts and the like describing the various NIST PQC candidates as they moved through the various rounds of the competition. I'll link to Soatok's, because his crypto expertise is solid and his explanations are generally easy to follow. And Soatok's whole fursona thing irks people I dislike, as an added bonus.

1This requires P≠NP, of course, or the complexity hierarchy collapses. If P=NP everything is (even more) terrible and we can't have cryptography. OK, it's a little more complicated than that; Impagliazzo's classic "Five Worlds" paper is a good introduction.

Judge sides with Meta and Google, puts California child privacy law on hold

Michael Wojcik Silver badge

Re: This may be too little, too late

How do you keep a kid from using an adult's account, full stop? It's a stupid, stupid law.

Michael Wojcik Silver badge

Re: Money

Yes, the law is almost certainly unconstitutional. But it's a terrible idea regardless.

There's no practical way to do online age verification without tremendous damage to privacy. (There are cryptographic information-concealing schemes that could avoid directly revealing PII beyond a proof of under/over 18, but only under generous assumptions, and they'd still be vulnerable to de-anonymization through correlation with other data.) People trumpeting the think-of-the-children line while breaking everyone else's privacy can fuck right off, in my opinion.

Michael Wojcik Silver badge

Re: Fuck me

Doesn't mean they're not right.

Michael Wojcik Silver badge

Re: Imagine

It's an abysmal law, and there's no justification for defending it. You're not on the side of the angels here.

Australia to build six 'cyber shields' to defend its shores

Michael Wojcik Silver badge

Re: Balancing act

Yes, it's easy to be critical of stated ambitions that are infeasible and vague, but at least it's an attempt to move in the right direction.

In the US, the Federal government has issued hundreds of these sorts of statements, and only a handful have led to actual progress – but a handful is better than none.

Certainly this is more sensible than some past Australian governmental meddling in this area.

California passes bill to set up one-stop data deletion shop

Michael Wojcik Silver badge

Re: Registration probably required

I suspect that the State of California already has PII for every California resident who has PII in the databases of any of those 500+ data brokers. I am not a resident of California myself, but if I were, I would use this portal; having a relatively small amount of PII held by a single government entity (which, as I've already noted, almost certainly already has it) is much, much better than having who-knows-what held by hundreds of completely untrustworthy commercial operations.

Having read the room, Unity goes back to drawing board on runtime fee policy

Michael Wojcik Silver badge

Yes. That doesn't mean the announcement wasn't also planned months ago.

The reasoning could simply have been: This thing we're planning to do in six months could boost the stock price or hurt it, or not affect it much at all. I'll sell a bunch of shares as a hedge.

Sometimes it's easy to demonstrate insider trading. Sometimes it isn't. Things like pre-trade filing, waiting periods, and blackout periods help reduce insider trading on short-term surprises by executives, but they don't really impair anyone's ability to plan for slightly longer-term events.

Michael Wojcik Silver badge

Re: what an "apology"

Also, I'd be interested to hear them clarify what "confusion" they believe they caused. Seems to me people understood them quite well.

Unity closes offices, cancels town hall after threat in wake of runtime fee restructure

Michael Wojcik Silver badge

Re: On false equivalences

When you produce an image with Photoshop, or a movie with Premiere, or build a house with a hammer, the resulting product doesn't contain Photoshop or Premiere or the hammer.

s/hammer/nail

Probe reveals previously secret Israeli spyware that infects targets via ads

Michael Wojcik Silver badge

Re: Insanet only selling to Western nations?

In the case of buggy image-rendering libraries, it's not a Harvard/Princeton issue, and strict code/data separation wouldn't help with many exploits. Buggy code is buggy code. Yes, strict separation eliminates in-process shellcode injection, but it's hardly a panacea; you still have ROP, for example, and if you can get some arbitrary data written to storage and subsequently executed all bets are off.

Activist investor to GoDaddy: Cut costs, improve sales, or sell

Michael Wojcik Silver badge

Ah, GoDaddy meets an investor with similar motives

Starboard: "We know you're rubbish, but surely you can be more rubbish?"

Ex-Twitter employees pull Musk back to money table over missing severance

Michael Wojcik Silver badge

Re: Who are you catering to who has been living under a rock?

Exactly. "Twitter" is a stupid and annoying name, but "X" is at least an order of magnitude more stupid and annoying. Also, including "Twitter" in the story makes it easier to search for; "X" is an abysmal search term.

Portable Large Language Models – not the iPhone 15 – are the future of the smartphone

Michael Wojcik Silver badge

For negative values of "better", perhaps

Yet smartphones are about to change for the better – thanks to the current wild streak of innovation around AI.

Ugh. All the ugh. There is absolutely nothing I find appealing about this prospect – and I've worked in ML, and I follow a fair bit of the LLM-related research, so it's not like I'm simply ignoring all the supposedly wonderful crap and claims of utility from the AI superfans.

Nor is this simply about my personal feeling of repugnance for LLMs and imprecise hallucination-prone UX. This is accelerating learned helplessness, and burning compute resources (which have actual real-world costs) in order to optimize human ignorance and foolishness. I've rarely agreed with Pesce's articles here on the Reg, but this one I think is particularly daft.

Michael Wojcik Silver badge

Check the author.

Michael Wojcik Silver badge

They certainly aren't a good interface for any of the "services" I use my phone for. Or for anything else I do, ever, with anything.

GitHub alienates developers by force feeding them AI recommendations

Michael Wojcik Silver badge

Re: Where's Elon then?

Like many tech executives, SatNad aspires to Elon-nature.

Michael Wojcik Silver badge

New GitHub, now with 20% more suck.

GitHub was always a terrible idea. (A central repository is precisely the wrong way to use a distributed change-management system, and GH's implementation is particularly crap.) Microsoft are clearly doing their best to make it even more terrible.

When does tackling pandemic misinfo become censorship? US courts argue it out

Michael Wojcik Silver badge

Re: Fairly obvious answer.

I thought we were talking about Carlin's shoe.

IBM Software tells workers: Get back to the office three days a week

Michael Wojcik Silver badge

Re: Triple Productivity?

Nah, it's easy to triple software productivity. Pick a metric, and watch the teams Goodhart it.

"Productivity" by itself is a meaningless term.