If a business invests in building an AI system, they now have an asset, and the value of the business reflects the fact the business owns that software asset now - if someone were to buy the business they would get the AI software and all its potential to monetize that asset in the future, so of course it has value.
If its value grows beyond the value the business originally invested to acquire it, it is quite literally a capital gain.
Why do you think Anthropic is worth $175-$350bn? Where did that capital value come from?
Yes, that’s how a software business works. The OP seemed to be talking about how we need to reform tax due to worker replacement. Everyone is talking past each other.
Correct, building an AI is an asset, which can then be rented to other businesses.
However the thread revolves around employers replacing employees with AI. Given that the number of AI creators is minimal, and the number of companies replacing employees is large, it follows that most companies replacing employees are renting AI, they did not create it.
Hence, for those companies, AI is not an asset, it is an expense.
One way of taxing those companies would be to tax AI producers based on revenue, not profits. If 50% of revenue was tax, then, the costs of AI to the end-user would go up to cover that. So revenue would "double", but half would go to govt.
I am not a tax lawyer though, but I expect such a scheme is so radically different to the current tax regime, that is has precisely zero chance of being implemented like this.
Of course businesses have always leased equipment to reduce the need for labor. This isn’t materially any different than paying your neighbor to borrow his ox and plow so you only need one guy to work your field instead of three.
I guess to put AI vendors on an equal footing with human intelligence vendors (ie humans). Workers are taxed on their revenue - their gross earnings - not their profits.
If the point is to tax AI consumers then AI providers can collect that tax on behalf of the IRS.
Taxing the profit of AI companies is useless since profit is a number that is easily manipulated to 0. Taxing revenue is much more direct. Prices have to go up to cover the tax. Hence the consumer oays "more" and that more is passed onto the tax man.
Taxing profit is exactly why businesses pay so little tax - it's trivial to make "no profit". (For example if the IP is held in another jurisdiction with a lower tax rate, and is "licensed" by the company which wants to make no profit. )
Aren’t we then just nullifying the productivity gains that could be had from the technology though? Obviously some people want this, they want AI to be less competitive with human labour, but don’t we just fall behind other nations who don’t tax that way and allow maximum productivity gains in all the AI consuming businesses?
From a practical end user perspective, being able to buy a device, and download and install binaries onto it to make it perform a specific purpose by plugging it in and dragging the file over, is considerably easier than installing an IDE, and downloading compiling and installing from source.
You missed the point. Selling an operating system at all was the innovation, rather than having it just come with the hardware. That the operating system they came up with the crazy idea to sell was someone else’s operating system is just an implementation detail, following the age old pattern of stealing others’ work industrialized decades earlier by Thomas Edison and thus requiring no innovation at all.
Good correction. This is the important point here. And there is a sub-point which is nearly as important:
The 8086 was out there and selling for years. AT&T ported UNIX™ to it, meaning it was the first ever microprocessor to run Unix.
But even so, DR didn't offer an 8086 OS, although it was the dominant OS vendor and people were calling for it. CP/M-86 was horribly horribly late -- it shipped after the IBM PC, it shipped about 3-4 years after the chip it was intended for.
The thing is, that's common now, but late-1970s OSes were tiny simple things.
Basically the story is that there was already an industry-standard OS. Intel shipped a newer, better, more powerful successor chip, which could run the same assembly-language code although it wasn't binary compatible. And the OS vendor sat on its hands, promising the OS was coming.
IBM comes along, wanting to buy it or license it, but DR won't deal with them. It won't agree to IBM's harsh terms. It thinks it can play hardball with Big Blue. It can't.
After waiting for a couple of years a kid at a small company selling 8086 processor boards just writes a clone of it, the hard way, directly in assembler (while CP/M was written in PL/M), using the existing filesystem of MS Disk BASIC, and puts it out there. MS snaps up a licence and sells it on to IBM. This deal is a success so MS buys the product.
IBM ships its machine, with the MS OS on it. DR complains, gets added to the deal, and a year or so later it finally ships an 8086 version of its OS, which costs more and flops.
The deal was very hard on Gary Kildall who was a brilliant man, but while MS exhibited shark-like behaviour, it was a cut-throat market, and DR needed to respond faster.
A library which patches a security vulnerability should do so by bumping a patch version, maintaining backward compatibility. Taking a patch update to a library should mean no changes to your code, just rerun your tests and redeploy.
If libraries bump minor or major versions, they are imposing work on all the consuming services to accept the version, make compatibility changes, test and deploy.
The problem with having a shared library which multiple microservices depend on isn’t on the microservice side.
As long as the microservice owners are free to choose what dependencies to take and when to bump dependency versions, it’s fine - and microservice owners who take dependencies like that know that they are obliged to take security patch releases and need to plan for that. External library dependencies work like that and are absolutely fine for microservices to take.
The problem comes when you have a team in the company that owns a shared library, and where that team needs, in order to get their code into production, to prevail upon the various microservices that consume their code to bump versions and redeploy.
That is the path to a distributed monolith situation and one you want to avoid.
Yes we are in agreement. A dependency on an external software repository does not make a microservice no longer a microservice. It's the deployment configuration around said dependency that matters.
No. Special relativity postulates that special relativity is preserved at all scales. It's an axiom. Comes from nowhere. It's assumed.
This is what a theory is: assume XYZ is true, and see how much of the world you can explain. Why is XYZ? That theory doesn't explain it.
Theoretical physics is: what is the smallest set of XYZ assumptions that can explain other theories. So if you can come up with a theory that's internally self-consistent that _predicts_ something which is postulated by another successful theory, that's a very convincing result.
Pardon, but, huh? I very much thought that Lorentz invariance was built into the assumptions of string theory.
Concluding from “A AND B” that “A”, while it does reach a conclusion that is distinct from the assumption, is not impressive.
If string theory does not bake SR into its assumptions, wouldn’t that make the way it is formulated, not manifestly Lorentz invariant? Don’t physicists typically prefer that their theories be, not just Lorentz invariant, but ideally formulated in a way that is manifestly Lorentz invariant?
Of course, not that it is a critical requirement, but it is very much something I thought string theory satisfied. Why wouldn’t it be?
Like, just don’t combine coordinates in ways that aren’t automatically compatible with Lorentz invariance, right?
If you formulate a theory in a way that is manifestly Lorentz invariant, claiming to have derived Lorentz invariance from it, seems to me a bit like saying you derived “A” from “A AND B”.
If string theory isn’t manifestly Lorentz invariant, then, I have to ask: why not??
Lorentz invariance is built into some descriptions of some stringy theories. For example chapter 1 of the Polchinski, you have the 26-dimensional bosonic string which is constructed to be Lorentz invariance. Obviously in this case it's not a "prediction", but then again, it's just a toy-model. Our Universe doesn't have 26 dimensions and doesn't have only bosons.
Ok, so I looked into it a bit, and here’s my understanding:
The Polyakov action is kinda by default manifestly Lorentz invariant, but in order to quantize it, one generally first picks the light cone gauge, where this gauge choice treats some of the coordinates differently, losing the manifest Lorentz invariance. The reason for making this gauge choice is in order to make unitarity clear (/sorta automatic).
An alternative route keeps manifest Lorentz invariance, but proceeding this way, unitarity is not clear.
And then, in the critical dimensions (26 or 10, as appropriate; We have fermions, so, presumably 10) it can be shown that a certain issue (chiral anomaly, I think it was) gets cancelled out, and therefore the two approaches agree.
But, I guess, if one imposes the light cone gauge, if not in a space of dimensionality the critical dimension, the issue doesn’t cancel out and Lorentz invariance is violated? (Previously I was under the impression that when the dimensionality is wrong, things just diverged, and I’m not particularly confident about the “actually it implies violations of Lorentz invariance” thing I just read.)
You understand that this have nothing to do with actual Lorentz invariance, yes? It sounds like you don't really understand the meaning of those terms you're using.
Do you understand what "manifest Lorentz invariance" means?
Of course. Did part of what I said suggest I thought otherwise?
I guess the part about the “when you quantize it after fixing the gauge in a way that loses the manifestness of the Lorentz invariance, if you aren’t in the critical dimension, supposedly you don’t keep the Lorentz invariance” part could imply otherwise? If that part is wrong, my mistake, I shouldn’t have trusted the source I was reading for that part.
I was viewing that part as being part of how you could be right about Lorentz invariance being something derived nontrivially from the theory.
Because, the Polyakov action (and the Nambu-Goto action) are, AIUI, typically initially(at the start of the definition of the theory) formulated in a way that is not just Lorentz invariant, but manifestly Lorentz invariant,
and if there is no step in the process of defining the theory that isn’t manifestly Lorentz invariant, I would think that Lorentz invariance wouldn’t be a nontrivial implication, but something baked into the definition throughout,
so, for it to be a nontrivial implication of the theory, at some point after the definition of the classical action, something has to be done that, while it doesn’t break Lorentz invariance, it “could” do so, in the sense that showing that it doesn’t is non-trivial.
And, I was thinking this would start with the choice of gauge making it no longer manifestly Lorentz invariant.
I trust you have much more knowledge of string theory than I do, so I would appreciate any correction you might have.
It does, but a number of alternative theories of quantum gravity do not. So, if Lorentz invariance is shown to be violated, this would favor those over string theory.
> naming things after random nouns, mythological creatures, or random favorite fictional characters is somehow acceptable professional practice. This would be career suicide in virtually any other technical field.
Really? Have you specced a microprocessor lately? Seen what pharmaceuticals are called? How polymer compound materials get named?
The "Raptor Lake" codename in microprocessors is internal, the product ships with systematic designation. Engineers spec chips by model numbers that encode generation, tier, and performance class.
In Pharmaceuticals, Doctors prescribe "sildenafil," not "Viagra." The generic name describes chemical structure. Brand names are marketing for consumers, not professional nomenclature.
Mythology in chemistry/astronomy has centuries of legacy and connects to human cultural history. Calling an element "Titanium" after Titans carries weight. Calling a SQL replicator "Marmot" connects to... what, exactly? A weekend at the zoo?
"Raptor Lake" isn't an internal codename, it's very much external as it's what Intel actively referred to that generation as. How's a non-geek shopping for a PC going to know if it's better or worse than "Lunar Lake" or "Alder Lake"? Maybe they just think their machine is shipping with some game where your giant dinosaur bird thing has to stop off for a quick drink to regain energy.
But in any case, this isn't the real travesty with these names. It's that they're reusing existing common words. The article hates on "google" when actually it's a fantastic name - if you googled it when it was introduced, all the results were about what you wanted. By comparison, Alphabet is an awful name, because if you search for Alphabet only a tiny subset of the results are going to be useful to you.
Naming schemes in consumer marketing serve a function. They are easily identifiable, unique, and memorable. All of these properties serve to identify the thing by associating a unique name with a unique set of services/function/effects on use.
Medical and chemical terminology is built on the history of latinate terms and compounds whose simples follow the same pattern. Latinate terms, I might add, which reference mythical, fantastical, or unusual things. Consider the planet Mercury, for example. The only difference? The centuries of time it took for scientific evolution to turn these unique names into a taxonomical language with its own logic.
There is no such taxonomy for computer science. But in the course of the evolution of such a taxonomy, it will be built out of the mess of names like the ones we like to use for our programs and tools like Rust, Ocaml (notice combination of interesting and technical), git, npm, bun, ada, scipy, etc etc.
Depends on the location, I guess. I've had doctors prescribe trade names, which I don't understand if there are alternatives with the same dosage, route of administration and similar inactive ingredients. Not even talking about the "do not substitute" prescriptions which are also based on dubious information most of the time.
As for "sildenafil" - I don't think generic names are usually meaningful. Usually the suffix relates to the category of the drug, but the first letters seem as random as the letters in trade names. I could imagine a world where the generic name is viagrafil and the trade name is Silden.
Generic drug names don’t describe chemical structure, they allude to purpose but that’s all. ‘-afil’ is used to apply to a particular class of drugs, although when it was discovered, ‘sildenafil’ was the only example of that class, so it didn’t mean anything already.
This is like having the first tool of a particular type come along and call itself ‘Mosaic’ and then someone makes another tool of the same kind and calls it ‘Mozilla’.
Brand name pharmaceuticals are sort of a different thing. Brand names must comply with the naming guidelines of the FDA, European Medicines Agency, and HealthCanda simultaneously. In practice, this makes it tricky to use actual words. So my companies adopt an 'empty vessel' naming approach. The empty vessels are nonsense words that (1) invoke an emotion (wegovy is a good example), (2) can be trademarked, and (3) it can survive brand pressure.
This is definitely where I would this pattern - MS Office 97’s customizable toolbars necessitated this model where every single thing you could do in the application had an icon.
It then got copied into Visual Studio, where making all of the thousands of things you could do and put into custom toolbars or menus have visually meaningful icons was clearly an impossible task, but it didn’t stop Microsoft trying.
I assume Adobe, with their toolbar-centric application suite, participated in the same UI cycle.
By the time of Office 2007 Microsoft were backing off the completely customizable toolbar model with their new ‘Ribbon’ model, which was icon-heavy, but much more deliberately so.
I still regard Office '97 as the best UI it ever had. I spent a lot of time inside it, including a couple of years at a bank reconciling corporate actions before I got my first programming job. The ribbon version was awful in comparison.
Also, the Excel Labs formula editor. But it needs a way to tell it "I know I have too many cells! Just let me trace over the 100 nearest rows."
The old scripting language can still be handy if you can keep people from opening the online version of Excel. Especially if you have a certain debugger addin[1]. Excel's JavaScript features are of limited use, if you're offline.
I keep wishing for a spreadsheet to implement all its scripting and formulas in something like Forth behind the scenes, so that every time a competitor announces n-more functions, we can just be like "Oh, really?" and add it.
[1] Related to waterfowl of the plasticised yellow variety. I'm not sure I can mention the name in a post anymore, since ages ago when I tried multiple times to post a properly-referenced (overly-hyperlinked?) message while my connection was very flaky. Note to self: should probably mail dang about this, some day.
Is the baseline assumption of this work that an erroneous citation is LLM hallucinated?
Did they run the checker across a body of papers before LLMs were available and verify that there were no citations in peer reviewed papers that got authors or titles wrong?
They explain in the article what they consider a proper citation, an erroneous one and an hallucination, in the section "Defining Hallucitations". They also say than they have many false positives, mostly real papers who are not available online.
Thad said, i am also very curious of the result than their tool, would give to papers from the 2010's and before.
If you look at their examples in the "Defining Hallucitations" section, I'd say those could be 100% human errors. Shortening authors' names, leaving out authors, misattributing authors, misspelling or misremembering the paper title (or having an old preprint-title, as titles do change) are all things that I would fully expect to happen to anyone in any field were things get ever got published. Modern tools have made the citation process more comfortable, but if you go back to the old days, you'd probably find those kinds of errors everywhere. If you look at the full list of "hallucinations" they claim to have discovered, the only ones I'd not immediately blame on human screwups are the ones where a title and the authors got zero matches for existing papers/people. If you really want to do this kind of analysis correctly, you'd have to match the claim of the text and verify it with the cited article. Because I think it would be even more dangerous if you can get claims accepted by simply quoting an existing paper correctly, while completely ignoring its content (which would have worked here).
> Modern tools have made the citation process more comfortable,
That also makes some of those errors easier. A bad auto-import of paper metadata can silently screw up some of the publication details, and replacing an early preprint with the peer-reviewed article of record takes annoying manual intervention.
There are other issues. In January they claimed that a US health report contained "fabricated" and "AI generated" citations with the headline being a claim from a Cigna Group report. Their claim it's fabricated is based on nothing more than the URL now being a redirect of the type common in corporate website reorgs.
I did some checking and found the report does exist, but the citation is still not quite correct. Then I discovered someone is running some LLM based citation checker already, which already fact checked this claim and did a correct writeup that seems a lot better than what this GPTZero tool does.
The mistakes in the citation are the sort of mistake that could have been made by both a human or an AI, really. The visualization in the report is confusing and does contain the 73% number (rounded up), but it's unclear how to interpret the numbers because it's some sort of "vitality index" and not what you'd expect based on how it's introduced. At first glance I actually mis-interpreted it the same way the report does, so it's hard to view this is as clear evidence of AI misuse. Yet the GPTZero folks do make very strong claims based on nothing more than a URL scraper script.
I mean, if you’re able to take the citation, find the cited work, and definitively state ‘looks like they got the title wrong’ or ‘they attributed the paper to the wrong authors’, that doesn’t sound like what people usually mean when they say a ‘hallucinated’ citation. Work that is lazily or poorly cited but nonetheless attempts to cite real work is not the problem. Work which gives itself false authority by claiming to cite works that simply do not exist is the main concern surely?
>Work which gives itself false authority by claiming to cite works that simply do not exist is the main concern surely?
You'd think so, but apparently it isn't for these folks. On the other hand, saying "we've found 50 hallucinations in scientific papers" generates a lot more clicks than "we've found 50 common citation mistakes that people make all the time"
Let me second this: a baseline analysis should include papers that were published or reviewed at least 3-4 years ago.
When I was in grad school, I kept a fairly large .bib file that almost certainly had a mistake or two in it. I don’t think any of them ever made it to print, but it’s hard to be 100% sure.
For most journals, they actually partially check your citations as part of the final editing. The citation record is important for journals, and linking with DOIs is fairly common.
the papers themselves are publicly available online too. Most of the ones I spot-checked give the extremely strong impression of AI generation.
not just some hallucinated citations, and not just the writing. in many cases the actual purported research "ideas" seem to be plausible nonsense.
To get a feel for it, you can take some of the topics they write about and ask your favorite LLM to generate a paper. Maybe even throw "Deep Research" mode at it. Perhaps tell it to put it in ICLR latex format. It will look a lot like these.
People will commonly hold LLMs as unusable because they make mistakes. So do people. Books have errors. Papers have errors. People have flawed knowledge, often degraded through a conceptual game of telephone.
Exactly as you said, do precisely this to pre-LLM works. There will be an enormous number of errors with utter certainty.
People keep imperfect notes. People are lazy. People sometimes even fabricate. None of this needed LLMs to happen.
A pre LLM paper with fabricated citations would demonstrate will to cheat by the author.
A post LLM paper with fabricated citations: same thing and if the authors attempt to defend themselves with something like, we trusted the AI, they are sloppy, probably cheaters and not very good at it.
Further, if I use AI-written citations to back some claim or fact, what are the actual claims or facts based on? These started happening in law because someone writes the text and then wishes there was a source that was relevant and actually supportive of their claim. But if someone puts in the labor to check your real/extant sources, there's nothing backing it (e.g. MAHA report).
Interesting that you hallucinated the word "fabricated" here where I broadly talked about errors. Humans, right? Can't trust them.
Firstly, just about every paper ever written in the history of papers has errors in it. Some small, some big. Most accidental, but some intentional. Sometimes people are sloppy keeping notes, transcribe a row, get a name wrong, do an offset by 1. Sometimes they just entirely make up data or findings. This is not remotely new. It has happened as long as we've had papers. Find an old, pre-LLM paper and go through the citations -- especially for a tosser target like this where there are tens of thousands of low effort papers submitted -- and you're going to find a lot of sloppy citations that are hard to rationalize.
Secondly, the "hallucination" is that this particular snake-oil firm couldn't find given papers in many cases (they aren't foolish enough to think that means they were fabricated. But again, they're looking to sell a tool to rubes, so the conclusion is good enough), and in others that some of the author names are wrong. Eh.
LLM are a force multiplier of this kind of errors though. It's not easy to hallucinate papers out of whole cloth, but LLMs can easily and confidently do it, quote paragraphs that don't exist, and do it tirelessly and at a pace unmatched by humans.
Humans can do all of the above but it costs them more, and they do it more slowly. LLMs generate spam at a much faster rate.
>It's not easy to hallucinate papers out of whole cloth, but LLMs can easily and confidently do it, quote paragraphs that don't exist, and do it tirelessly and at a pace unmatched by humans.
But no one is claiming these papers were hallucinated whole, so I don't see how that's relevant. This study -- notably to sell an "AI detector", which is largely a laughable snake-oil field -- looked purely at the accuracy of citations[1] among a very large set of citations. Errors in papers are not remotely uncommon, and finding some errors is...exactly what one would expect. As the GP said, do the same study on pre-LLM papers and you'll find an enormous number of incorrect if not fabricated citations. Peer review has always been an illusion of auditing.
1 - Which is such a weird thing to sell an "AI detection" tool. Clearly it was mostly manual given that they somehow only managed to check a tiny subset of the papers, so in all likelihood was some guy going through citations and checking them on Google Search.
I think we should see a chart as % of “fabricated” references from past 20 years. We should see a huge increase after 2020-2021. Anyone has this chart data?
Quoting myself from just last night because this comes up every time and doesn't always need a new write-up.
> You also don't need gunpowder to kill someone with projectiles, but gunpowder changed things in important ways. All I ever see are the most specious knee-jerk defenses of AI that immediately fall apart.
Under what circumstances would a human mistakenly cite a paper which does not exist? I’m having difficulty imagining how someone could mistakenly do that.
The issue here is that many of the ‘hallucinations’ this article cites aren’t ’papers which do not exist’. They are incorrect author attributions, publication dates, or titles.
