I know it’s a joke, but it does make me wonder if LLMs would even be good at assessing if an idea is “novel”.
I only have a rudimentary understanding of how neural networks work, but I wonder if rather than “understanding” what “novel” really means to humans, an LLM would be most likely to agree that something was novel based on having seen that specifically referred to as novel in its training data.
So that if you give it an example of something that already exists, but which was very recently invented at the point in time when the LLM was trained, and you ask “is this a novel idea?” that because it had several sources in its training data describing that idea as novel, it would say “yes that’s a novel idea”. Whereas what we really meant was to ask it if someone else had already thought of this thing prior to us right now in this later moment.
And then on the other hand, even if something was “novel” at the point in time when the LLM was trained, perhaps we would fare better to ask it “has anyone thought of this?” rather than asking if the idea is “novel”? And that even though it considers the idea novel in a way it would also be able to say that yes this has already been thought of.
I took Claude Code with the prompt copied from GP, and threw it at a repo that's just a collection of ad-hoc narrow-purpose single-page HTML tools, all of them vibecoded on the fly. It found four things "it would show the patent attorney" and three more under liberal interpretation of "novel".
Notable inventions include:
1. Taking well-known math for trilateration, but putting it inside a phone app and making it ergonomic to use. I vibed it so I can survey a plot quickly using a hand laser ranger and a foldable phone for data entry; apparently, the novelty is in combining the live display of how the measurements resolve to a 2D structure, with error residual, and a list of additional measurements to make, priority-sorted by how much error can be removed by making it. Apparently it's "closing the loop from adjustment back to "what should I physically measure next."" (and yes, it in fact does that, and works well).
2. A shader doing "resolution-independent procedural texturing of a hyperbolic ground plane". Basically the result of me asking "I want something like Hyperbolica to play with on my phone, now now now", followed by "cool but it got ugly tessellation artifacts far from origin". I'm not qualified to judge whether this is in any way not obvious, but Claude is framing this as "specific technical solution to a specific floating-point problem, which is exactly the flavor of thing that survives §101 scrutiny best".
3. There's one I'm afraid to even describe because it has to do with UX of VLMs and that could make it novel enough somebody will patent it. But to hell with it: basically, take an image annotator (draw colored rectangles on an image), allow user to label the rectangles in the interface, and then save, alongside the image, a file mapping colors to labels. This was my idea to avoid having to write stuff like "that red thing is XYZ" and "the purple rectangles are VWX". Pretty obvious UX streamlining that's one of many that apparently no major AI provider thought of yet...
--
1. is a shared invention - me pushing for ergonomics, LLM figuring out the "game loop". 2. was Claude Fable all on its own - I just told it it's ugly and looks like the kind of nonsense you get at the poles of a globe made of triangle strips, and it came up with the shader. 3. was my idea from the start.
Question to patent people: are things like this really patentable? I hope not...
Businesses know no shame, if applying for a trivial patent is useful, they will of course do so.
The real problem is patents which are trivial and/broad being granted.
Not the first time, not the last time in US patent history ...
“Claude, please review my repo for any work that could be novel enough to receive a patent. Be liberal in your definition of ‘novel’”
I know it’s a joke, but it does make me wonder if LLMs would even be good at assessing if an idea is “novel”.
I only have a rudimentary understanding of how neural networks work, but I wonder if rather than “understanding” what “novel” really means to humans, an LLM would be most likely to agree that something was novel based on having seen that specifically referred to as novel in its training data.
So that if you give it an example of something that already exists, but which was very recently invented at the point in time when the LLM was trained, and you ask “is this a novel idea?” that because it had several sources in its training data describing that idea as novel, it would say “yes that’s a novel idea”. Whereas what we really meant was to ask it if someone else had already thought of this thing prior to us right now in this later moment.
And then on the other hand, even if something was “novel” at the point in time when the LLM was trained, perhaps we would fare better to ask it “has anyone thought of this?” rather than asking if the idea is “novel”? And that even though it considers the idea novel in a way it would also be able to say that yes this has already been thought of.
I took Claude Code with the prompt copied from GP, and threw it at a repo that's just a collection of ad-hoc narrow-purpose single-page HTML tools, all of them vibecoded on the fly. It found four things "it would show the patent attorney" and three more under liberal interpretation of "novel".
Notable inventions include:
1. Taking well-known math for trilateration, but putting it inside a phone app and making it ergonomic to use. I vibed it so I can survey a plot quickly using a hand laser ranger and a foldable phone for data entry; apparently, the novelty is in combining the live display of how the measurements resolve to a 2D structure, with error residual, and a list of additional measurements to make, priority-sorted by how much error can be removed by making it. Apparently it's "closing the loop from adjustment back to "what should I physically measure next."" (and yes, it in fact does that, and works well).
2. A shader doing "resolution-independent procedural texturing of a hyperbolic ground plane". Basically the result of me asking "I want something like Hyperbolica to play with on my phone, now now now", followed by "cool but it got ugly tessellation artifacts far from origin". I'm not qualified to judge whether this is in any way not obvious, but Claude is framing this as "specific technical solution to a specific floating-point problem, which is exactly the flavor of thing that survives §101 scrutiny best".
3. There's one I'm afraid to even describe because it has to do with UX of VLMs and that could make it novel enough somebody will patent it. But to hell with it: basically, take an image annotator (draw colored rectangles on an image), allow user to label the rectangles in the interface, and then save, alongside the image, a file mapping colors to labels. This was my idea to avoid having to write stuff like "that red thing is XYZ" and "the purple rectangles are VWX". Pretty obvious UX streamlining that's one of many that apparently no major AI provider thought of yet...
--
1. is a shared invention - me pushing for ergonomics, LLM figuring out the "game loop". 2. was Claude Fable all on its own - I just told it it's ugly and looks like the kind of nonsense you get at the poles of a globe made of triangle strips, and it came up with the shader. 3. was my idea from the start.
Question to patent people: are things like this really patentable? I hope not...