IBM Patented Euler's 200 Year Old Math Technique for 'AI Interpretability'

https://leetarxiv.substack.com/p/ibm-patented-eulers-fractions

by busymom0 • 8 months ago

172 67 8 months ago

leetarxiv.substack.com

mv4• 8 months ago

I remember reading that IBM holds a staggering number of patents - over 150,000 in the US alone.

When I worked there many decades ago, I had pressure and saw people get rewards for patenting everything you could. Seems like things haven't changed.

stuffn• 8 months ago

This is the primary reason IP patents should be thrown out. They're far too easy to get, exist primarily for the purpose of rent seeking, and generally provide nothing useful to the public knowledge.

anthk2• 8 months ago

Good in case of Europe has a shortage on toilet paper.

Also, this might be tangentially related to Fractan. Any Mathematician there? if there's some relation, the patent might be void/null.

https://en.wikipedia.org/wiki/FRACTRAN

noir_lord• 8 months ago

While we might get to laugh at those silly people with their software patents (and describing them as toilet paper is quite funny) it still has real world effects for us.

For many years I had to install a linux distro and then mess around with freetype because there was a valid patent on subpixel hinting that was valid in the US and irrelevant in Europe but because the distro's I used at the time where US based I got caught in the cross fire.

Given the huge imbalance between US tech and Europe Tech (and rest of the world really) in software they sneeze, we get flu.

That only stopped been a problem when the patents expired[1].

[1] https://freetype.org/patents.html

anthk2• 8 months ago

That's why you should always use either European or Canadian mirrors. On FreeType, maybe Clear Type related; but I remember some slight hinting methods not related to cleartype which looked far superior on flat LCD screens IMHO.

Another one was Lame and MP3. And I remember the codecs issue with the Penguin Liberation Front for Mandrake, LIVNA for another distro and Debian Multimedia, among others.

pkaye• 8 months ago

Isn't this just a patent application? Also what are the actual patent claims? I was not able to find a link to the application.

kristjansson• 8 months ago

https://patents.google.com/patent/US20230401438A1

ipfreely• 8 months ago

Misleading. This has not been patented. The author links to a patent publication, which happens to the vast majority of patent filings in the U.S. There was a rejection of this patent application mailed by the USPTO in August 2025.

analog31• 8 months ago

Rejection is when the fun begins. Note that I'm not a lawyer, but I do have 20+ patents, none in the software realm. Rejection means you get to revise your application and re-submit with a new deadline and new fees. Your chances of getting it accepted on the second or third try can vary.

Applications are seldom withdrawn or ultimately rejected altogether, but may be granted with severely narrowed claims. Also, you can revise the claims without changing the text, within limits, which results in granting of a patent with a very broad text followed by narrow claims

phendrenad2• 8 months ago

Applications are seldom withdrawn or ultimately rejected altogether

For a rather unpopular definition of "seldom", sure.

rdtsc• 8 months ago

The patent system for software is garbage, but it also doesn't seem to jive with the title. They don't claim to invent continued fractions, what they patent is applying them to some domain.

https://patents.justia.com/patent/20230401438#history

A method, a neural network, and a computer program product are provided that provide training of neural networks with continued fractions architectures.

It's a bit like the propelling device https://patents.google.com/patent/US1331952A/en -- it doesn't invent or patent shoes, or invent or patent springs, but it patents attaching springs to shoes.

It would seem using continued fractions with elliptic curves what the author wants to do, wouldn't be covered.

However, I still think it can still be challenged if someone can show that continued fractions have been used in with NN before. Or even better, maybe pytorch or other open source projects can explicitly reject crap that's patented. If you put your shit in a junk patent, take it out of the project and enjoy yourself, don't spread it around. So, if the authors of the patent are the ones pushing for the inclusion, then someone should challenge that and have it removed unless the patent is withdrawn.

ajb• 8 months ago

The problem here is whether just applying some well studied technique to a new area like AI is really inventive enough to be patentable. There were loads of patents which basically added "in a computer" to some existing technique. Just because NN's are novel doesn't mean people should be able to get an exclusionary property right over the use of well known techniques in NN's because they put in half a day's work to be the first to apply that technique to NN's. This isn't the Guinness book of records.

IceHegel• 8 months ago

Whether or not the patent was actually granted in this case, I have not been able to think of a compelling reason to have patents for software. In fact, I think most intellectual property laws need to be seriously rethought.

If the objective is to maximize investment by protecting successful results, I don't think our system is doing a very good job.

keeda• 8 months ago

Before this devolves into the morass of outraged comments that usually comprises any discussion of patents, here is the golden rule I coined for such threads:

RTFC: Read the Fucking Claims

The claims are the only part of the patent that really matter because those are the only enforceable language. Plus, this is a patent application so the claims have not even been examined yet.

Without commenting on the merits of this patent itself, remember that new applications of existing techniques are still novel, and hence patentable. In fact, if you think about it, almost all inventions and innovations are just applying novel combinations of well-known techniques to new use-cases. (Anything that doesn't fit that definition and introduces genuinely new methods is usually in the "groundbreaking" category.)

So I'd guess "applying old technique to new problem" is probably the case here. I'm no subject matter expert, and there may be prior art that invalidates this, or it may not meet the non-obviousness bar... but when the "200 year old math" came about neural networks were not really a thing.

linuxhansl• 8 months ago

Software patents are mostly garbage and unneeded. And I say that having some in my name (I actually tried to get my name off them, but our lawyers said it's not possible).

Show me one useful software patent that (a) is not "obvious to one skilled in the art", and (b) benefits society by being granted a monopoly. Just one!

Software rarely requires expensive research that would be worth protecting. Rather than enabling a fair market, this takes fairness out of the market.

Software patents are like getting a patent on "Murder story with final revelation of who did it." Maybe add one or two features, like a "detective with hat", etc. In one fell swoop you would be able to own most murder mysteries.

Software (like books, stories, art, etc) is better handled by Copyright law. May the one who actually has a better product win!

Sorry for the rant.

JohnFen• 8 months ago

I agree with your rant. I have a similar one. I personally think that software should mostly just not be patentable at all, since (among others) these things are not patentable according to US patent law: scientific discoveries, mathematical methods, aesthetic creations, and rules or methods for performing mental acts.

bandofthehawk2• 8 months ago

In addition, software is also copyrightable, which makes much more sense than patents for protecting unauthorized use. IMO, patents for software should be mostly eliminated, and even copyright terms should be much shorter.

fsckboy2• 8 months ago

Show me one useful software patent that ... (b) benefits society by being granted a monopoly. Just one!

that's a bullshit criterion, and nothing about what you said applies to software in particular.

it's generally agreed that monopolies are bad for society, but patents only grant a temporary monopoly to reward innovation, in exchange for which society gets disclosure (at the time patent systems were promulgated, many ideas died with their inventors as trade secrets) and encourages more R&D by creating a system for payoff.

bandofthehawk2• 8 months ago

The only "system for payoff" I've seen with software patents is patent trolls. Are there cases of software inventors being rewarded for their software more fairly because they had a patent?

Dylan168079• 8 months ago

It's not bullshit. You said it yourself, the benefits are supposed to be disclosure and encouraging R&D.

But disclosure is rarely an issue with software, and patents are bad at properly disclosing software details in the first place.

And in software there's already a huge motivation to do R&D, while patents are more likely to block useful work than in most fields. Even if I think of highly optimization-motivated fields like video encoding, patents slow down innovation more than they accelerate it.

So can you name some software patents where those motivating factors actually worked? It's a fair question.

yyurgenson• 8 months ago

RSA

Dylan168079• 8 months ago

What was the societal benefit of putting 20 years of monopoly on that algorithm? I don't think potential profit was a big motivator in that research work.

And that patent got invalidated in most of the world anyway.

josefritzishere• 8 months ago

Gold star comment.

mring33621• 8 months ago

I agree that software patents are generally garbage and hurt innovation.

Unfortunately, there are probably many people here on HN that make a living off software patents.

smj-edison4• 8 months ago

What would you think about still having software patents, but having them expire in 2-5 years? I feel like the concept of a patent is still a good one, but the time it takes to go from zero to a product is drastically lower with computer programming. The patent length should reflect that.

OgAstorga• 8 months ago

How would you feel about patenting language? I.e. If you speak with certain words or certain patterns then you have to pay a royalty (only for 2 to 5 years).

greg_w2• 8 months ago

One cannot (in the US) get a patent for software itself. This was settled a while ago. There needs to be more in the claims. In fact, the patent discussed here does not claim continued fractions and nobody would be in danger using them even if the patent issued as is (which is not certain, because the patent claims rather trivial modification of a classic neural network architecture, which should be brought up by the examiner as obvious).

Patents are propelling the society when they work as intended. They made XIX century and at least good chunk of XX century. Without patents, people fall back to copying each other, because it is much easier to copy than to innovate.

Retric2• 8 months ago

Patents go way back possibly as far as 500 BC, with other examples dating to 1331 etc, yet the introduction didn’t kick off any great wave of progress.

Instead a wide range of factors like better plants feeding both population growth and an ever larger percentage of society could do something other than grow food where the real root causes here. Devoting land and labor to cotton for example requires a surplus of food.

Dylan168079• 8 months ago

One cannot (in the US) get a patent for software itself. This was settled a while ago. There needs to be more in the claims.

The "more" is basically busywork for lawyers. We effectively have patents for software itself.

versteegen• 8 months ago

Without patents, people fall back to copying each other, because it is much easier to copy than to innovate.

This is probably the only argument I've seen in favour of software patents in which I see any merit!

dpark• 8 months ago

A huge chunk of software patents are also “well known thing, but with software”. Looking at you, Sonos.

Even if these things were novel at one point, doing them again in software is not novel.

thwarted• 8 months ago

IBM Slapped the Buzzwords 'AI Interpretability' on Generalized Continued Fractions and their Series Transformations and was awarded a Patent

A good portion of the perception and use around patents is the word "awarded", which implies that there was some evaluation, either objective or subjective, by an expert in the topic and that the invention met some high level of… something that made it worthy of being "awarded" for the "work" that it took to invent. Most parents are not evaluated like that today, and software patents specifically probably never have been.

trhway2• 8 months ago

> IBM Slapped the Buzzwords 'AI Interpretability' on ...

we've been through several such cycles - "...on computer", "... on Internet", and now " ... with AI" .

(of course i have several, including couple with AI, and i'm against patents, especially software ones)

pinkmuffinere• 8 months ago

Software patents are mostly garbage and unneeded. And I say that having some in my name (I actually tried to get my name off them, but our lawyers said it's not possible)

I want to know about these patents that you want your name off of! What were they for? Why do you want your name removed? This is the opening to a great story, please tell!

aeonik2• 8 months ago

I'd be super cool with software patents if they actually open source the software and had actual usable modules included with the patent.

otherwise I agree, copyright is the way to go.

CamperBob2• 8 months ago

I wouldn't. Many if not most patents are awarded to the first person to encounter a problem and apply a (likely obvious) solution, not the first person to solve a longstanding problem at great expense after many others have tried and failed.

And even if that weren't the case, nobody needs to be given exclusive rights to math. In principle the law says as much, but in fact that's never been an obstacle.

wahern• 8 months ago

And I say that having some in my name (I actually tried to get my name off them, but our lawyers said it's not possible).

I was originally named on a patent filing (along with the CTO), but left the company and apparently it was too much effort to communicate with me so they just swapped my name for another engineer. But it was literally my idea (in as much as any such software patent is one person's idea). I was literally given a problem to solve and came back with two alternative solutions to implement, without conferring with anyone else at the company or elsewhere. The only input from anyone else in the company was selecting one of the two options, which I then fully implemented entirely myself.

Not only have I always disliked software patents on principle, I was also cheated out on having a patent to put on my resume. (Heck, it was hardly my best idea and for all I know the company patented a bunch of my other work without citing me.) Which is pretty much their only value--as social currency for hiring or highly specious pretend asset security for VC investments.

Technically, removing my name from the filing invalidated the patent, but it's all just a ridiculous shell game.

zkmon• 8 months ago

I don't think they are patenting any classical math as the blog claims. It's an architecture using CF that is patented, I believe.

Ofcourse, patents are trash.

shmerl• 8 months ago

Reminds this: https://theonion.com/microsoft-patents-ones-zeroes-181956466...

radiator• 8 months ago

Is there not a legal way to fight this back? With this patent application, someone is trying to establish an advantage *against everyone else*. So legally, anyone might have self-interest to sue the applicant, if the application is fraudulent (because for example it is a stolen 200-year-old mathematical method). I mean, such patent applications should not be a freeroll, they should also entail some risk for the applicant.

Beijinger• 8 months ago

Relevant: https://www.reddit.com/r/explainitpeter/comments/1ovy31o/exp...

juancn• 8 months ago

They probably did it for the bonus.

Most large companies pay you a bonus when a patent is filed, when it's granted and when it's licensed.

It's one of the few ways to change your income on such companies as a technical person.

IBM Patented Euler's 200 Year Old Math Technique for 'AI Interpretability'

Related Stories