The other side is Cognitive Radio [1] which also evolve the OTA protocols for cooperative diversity from IEEE 802.22 onwards. Now I can see AI, via a local SLM/NPU plus agentic GNURadio loops for new radio use cases. This is going to be much more wide spread in the upcoming 3GPP 6G releases in 2030.
Was going to say much the same. I recall one story about a genetic algorithm to make an oscillator with the fewest possible components, and it successfully did so by surprising the humans with a single wire, i.e. an antenna picking up nearby stray RF.
That is my favorite part of GA. Gradient free optimization but it turns out making a good fitness function is hard and like 70% of the time it just exploits some assumptions or gap you have in your theories. Really reveals the problem in different ways that traditional ML.
As someone who does a lot of genetic programming (like, old-school, without AI/LLMs, etc), I can confirm that the fitness function is very difficult to get right, especially if you are trying to evolve programs that have "adversarial fitness" -- you'd need to maintain a hall-of-fame, and that just makes the runs take _much_ longer, because, chances are, your fitness function is the bottle-neck.
So, it is very hands-off, but also very expensive, and it is never clear if optimizing the fitness function is worth it, because the fitness function itself may be insufficiently or incorrectly specified.
However, I do think that people should try, even with just a whiteboard or a notebook, to design a fitness-function, for their problem, as if they were going to try to evolve it, because (1) it forces them to explicate their correctness constraints, and (2) they may discover that the program that they are trying to write _is equivalent_ to the fitness function.
I'll give you an example for point 2. Many years ago, I had to parse a gnarly language, and I chose to do it via Chomsky Grammars (that automatically build a tree based on the grammar-spec). Chomsky Grammars are cool, in that they are basically just a state-machine, but they are incredibly difficult to debug: when they work, they might work incorrectly (malformed tree), and when they fail, they give no reason for failure (because even with a trace, you are trying to figure out which backtrack should not have happened). So, out of desperation, I started to consider using genetic programming to just evolve a correct Chomsky Grammar. It became clear that there are only 2 possible fitness functions (1) a function that tests a hand-picked input against a hand-crafted tree-output (which is vulnerable to over-fitting), and (2) a function that is not (well, is much less) vulnerable to over-fitting, but is effectively a pre-existing, correct grammar that can produce those trees.
If you are in situation 2, then the genetic programming is not necessary, unless you are trying to create an optimized (or obfuscated) parser, and even then the optimization may be overfit to the test-inputs (even if they are generated test-inputs from the grammar itself). If you are in situation 1, then you are better off re-evaluating your approach (I abandoned the Chomsky Grammar notation, and invented one that is much easier to understand and debug, without losing any of the expressiveness -- it also happens to be slower, but fast-and-broken is worthless compared to not-so-fast-and-works-fine).
One place where genetic programming has been consistently awesome, is in parameter-search style problems (e.g. your genome is a long list of floats, representing weights and/or anti-weights, and you need to find out which weights give you more fitness (or less error)). I hear good things about variable-neighborhood-search, but have yet to try it.
That sounds apocryphal but there was a noted paper describing a frequency discriminator implemented using a genetic algorithm and it ended up tied to the exact piece of silicon used to evolve it, with logic cells not connected to anything still changing the output.
That second paper is absolutely amazing, I’ve always heard this story and never bothered to find the source.
The section with oscilloscope traces showing the progression of the “designs” over time was extremely interesting - I’d love to see what the 10x10 grid of functions looked like at each snapshot.
GA’s optimize only combinatorial problems though — where you have discrete set of choices (~genes) for each variable, and therefore do not have a gradient
I wonder if our common expectation that true theories somehow had to be beautiful and elegant is going to survive the coming century. What if "real" nature phenomenon were actually best described by horrible mess of impossible equations, that only machines could actually manipulate and reason about ?
I think you're going too far with this. Most people understand scientific theories to be an approximation. F=ma is approximately true, in the sense that it's only accurate within the newtonian regime and each of those terms includes so many asterisks that you will only ever measure it approximately.
The latter is the jokes about the physicists "assuming a perfectly spherical cow."
In fact that's kinda the whole point of the "unreasonable effectiveness of mathematics" essay. It is unreasonable that mathematical approximations are so good at describing our world.
I often think this about medicine and the human body. We want to believe that our bodies are some miraculous well oiled machine. But it often seems that it’s a barely held together bag of mess.
Biology is incredibly robust!! I'd say barely held together bag of mess describes something like an internal combustion engine. A primate, on the other hand, is a self-replicating machine capable of self-repair and just about universal fuel sourcing. It has a robust defense network capable of identifying and eradicating a staggering number of foreign replicators. It has holographic design storage, with each cell containing the plan for the whole organisim. It has general cognition based on a world model, and does all this on almost no energy.
I think your reply and the parent can both be true, you're just using slightly the same words to describe different things.
The parent is talking more about elegant simplicity vs. sprawling, seemingly haphazard complexity, and you're talking more about durability to failure points and 'completeness'.
Likewise, in code, a lot of the most durable, battle tested software looks extremely inelegant and duct taped, as 90% of the code is dedicated to handling one-off patches and weird edge cases.
My suspicion is that we had a sense that generality and compactness was really neat, so we liked easily-remembered laws like F=ma. Applies everywhere, is clean.
When you attempt to hyper-optimize, even with humans in the loop, you end up a mess. You're lucky if you can find clean guiding principles anywhere. If you can hyper-optimize hyper quickly, you end up with an extra layer of mess.
Math is a language to explain systems. Teaching someone that force varies linearly to mass is a helpful first pass. It isn’t exactly linear but is not exponential at all.
Gaining expertise is always the hard part and our new LLM overlords are making that much harder. So the simple “pure” functions as a teaching aid have never been more important.
End users have never cared about how the sausage is made though.
LLMs can explain complex things to humans with tons of specific context that you don’t find in textbooks or even a google search.
It’s probably never been easier to grasp a large codebase than it is today for example. You can probe and ask specific questions without going through a maze of imports and relationships and config files yourself.
Learning things will always be up to the person, it’s still a choice and dedication to a craft can still be taught.
But those proofs are showing that the fundamental axioms (which are generally simple and elegant) are still enough to build a complex result.
I think of elegance as not having to add epicycles, not that everything in the system has to be simple.
Also, without a working theory the, the space of possible solutions is near infinite. LLMs manage to pluck out the space of comprehensible English strings from n-dimensional hell. Even if this is done with a black box of billions of parameters, it’s still elegance in the sense that such a space even exists and was found
The "common expectation" I think, misses the point. The idea isn't that fundamental theories are simple or elegant (quantum physics equations are pretty darn ugly), it's that, given the choice between a more complicated and a more simple theory, generally the simplest one is the most accurate choice.
Are you thinking of any specific examples? I don't disagree that complex things generally end up having complex explanations, but I'm admittedly drawing a blank trying to come up with things where the most complex explanation ended up being the correct one.
There are many marvels of evolution's ability to come up with robust complex distributed systems which work way better than anything we build. The one I've learned about most recently is the https://en.wikipedia.org/wiki/Immunological_synapse in which different kinds of white blood cells gather around a bit of evidence that one of them found and decide whether to shoot the messenger (clonal anergy), or raise a clone army to defeat the invader (T-cell activation).
Imagine that it's maybe the 1800's and you're asking why somebody who has already survived smallpox is not susceptible to becoming infected again. If you offered an explanation involving tiny detectives wandering around and collecting evidence which they present to each other and decide whether to multiply... one in which the tolerance comes from the detectives from the previous fight still hanging around in your lymph nodes ready to spring into action if they run across the right kind of evidence. Well that would probably be a more complicated explanation that anybody at the time would offer, and it would also be correct.
Incomplete prior knowledge doesn't mean it's simpler, just that it's inaccurate. Would the phenomenon you're describing really accurately be explained by something _simpler_?
You should look into Solomonoff induction. Nature and physics, absolutely, tautologically, have to follow the "shortest explanation is more likely principle".
Solomonoff induction doesn’t concern itself with what is truth and reality. It just says which theory to prefer and how to determine so objectively when multiple are equally precise in making predictions of observations. It’s a formal description of Occam’s razor.
OPs argument is that reality is expressed by very complex equations and interactions; by definition this is outside of Solomonoff induction because it’s easy to imagine this accurate model by definition is the shortest algorithmic explanation, it’s just orders of magnitude more complex than our current approximations.
You should include the error correction code length in the description length. This means Newtonian mechanics was a much longer theory to describe Mercury's orbit than general relativity. It was only the shorter theory before they had the data showing a discrepancy. Which is the correct approach to describing your reality, because until you see a discrepancy, the extensional properties all follow the shorter rules.
I guess the argument from OP would look like: "Yes, now imagine we poke and extend our universe as far as we can. How much bigger do you think our final 'shortest description' would be? I imagine it may be orders of magnitude more complex."
Well, I can imagine a squared circle... doesn't mean the math checks out. I would reply that you do not have to imagine, you can go about looking at different mathematically possible universes in Tegmark IV and find the expected number of bits for the one you actually exist in. Which is ~0 bits more complex than the shortest description based on the data you currently have.
Also, note that Newtonian mechanics is not actually a very short theory for building a universe, because you have to instantiate every object in the universe. You actually get a lot more of the structure for free with general relativity (re: Wigner's classification of the particles). An observer in a presumed-Newtonian universe calling it a simple theory would be like saying, "I compressed Wikipedia to one byte, just by putting it all in the decompiler!"
>I wonder if our common expectation that true theories somehow had to be beautiful and elegant is going to survive the coming century.
That's the layman's idea of physics theories. They are beautiful and elegant only on the surface, that's why they're technically models and approximations of the real world. The standard model renormalization techniques are a mess of patches and ad-hoc heuristics, pretty far from the "this lagrangian literally contains all physics". Generally you just _ignore_ higher order terms and just call it a day. The famous E=mc^2 it's just the first term of a Taylor expansion. The beautiful form of physics it's what you would call "good enough" and often just a pedagogical tool.
It would be really cool. We already know everything at the lowest levels is a probability cloud. There’s beauty and contentment in not really being able to nail anything down for eternity…
That's a result of the Copenhagen Interpretation. There are other interpretations of the math which don't rely on reality fundamentally being a probability cloud/wave/field.
How would nature be best described as a horrible mess of impossible equations? They would be best described as elegant and beautiful no?
I think your point is more that we might be able to initially describe complex phenomena as messy, horrible complex equations, that doesn’t mean we shouldn’t work to simplify them and make them more understandable to us.
One great application of AI design is patent poisoning. Use AI to churn out masses of variant designs, make them publicly visible on a web site, and if future patents come out use any collisions to invalidate them or at least restrict their scope (generalization of a patent is limited by prior art.)
I’m reminded of lawyer Damien Riehl’s (performative) reaction to the Sam Smith infringement decision, back in 2019/2020. He and programmer Noah Rubin algorithmically generated every possible melody (within a certain combinatorial space, in MIDI format as I recall), and purported to release them under CC-0 license [0]. He went on to attract some attention and explain his argument at a regional TEDx event [1].
I seem to recall legal commentators reacting with an eyeroll—apparently judges split much finer hairs than these for a living—but it was a cute stunt.
For sure, but I suspect the law might look similarly dimly on the argument that “machines systematically generated all possibilities in the problem space” === “somebody already had this idea.” I’d imagine maybe by reading specific human intention into “prior art” and “existing work” and those sorts of terms.
Which is not to say let’s not do it anyway and see!
"Humans couldn't even imagine" seems like overselling it, but I'm sure that machine learning algorithms can brute force their way to chip designs no one has tried before and that some of those might be useful to us. That seems like a pretty reasonable thing for a computer to do.
It's marketing bullshit. For one, it's like proving a negative; you can't prove to me that humans couldn't have imagined it. Second, humans have already imagined quite a lot of crazy stuff...
I’m a bit frustrated. AI can do a looot of things; but I think as we continue to muddy the waters between LLMs and more traditional machine learning like Monte Carlo, Genetic Algoriths, Expert Systems and other Statistics magic tricks, we’re too aggressively conflating established and morally neutral activities in ML with the concerns that people have about LLMs and Stable Diffusion.
It is a problem because people will talk about what AI can do implying that an LLM can do that thing, making it seem like a pure LLM can do almost anything. On the other hand people will say AI will never be able to do X because an LLM can’t do that thing well natively. AI has become too vague of a term to be useful.
We're learning that people are way too lax with where they apply the term "intelligent". LLMs aren't remotely intelligent, but people are trying to ride the hype train and call them intelligence.
Maybe I'm just significantly and unrepresentatively unlucky, but Claude is significantly more intelligent than the average human around me on most any metric I can think of.
I wish I could wave a magic wand and just make the word "AI" go away. It has no actual meaning. It could mean anything from your opponent in Mario Kart to Stable Diffusion.
"AI" == "what (through tech) can replace a professional"
It may seem similarly vague, but it does in fact open interesting, productive, and necessary questions. A "computer" was a professional crunching numbers - "replaced", "easily" because of the deterministic procedural nature of said work, but what about the technical effort to arrive there, and what about the less "mechanical" jobs? When do "processes" become "intelligence"?
Some of us had studied AI originally to study the mind - "how do we formalize thought". It's the interdisciplinary, transversal nature of the area.
Also maybe compare that with that large and important intersection between CS and Economics - the "science of optimization" and its implementation in efficient IT systems. The effort in terms of that different discipline may not be evident, yet lots of engineering is "optimizing" and the generalization of those solutions we call Economics (see the book Algorithms to live by).
So: the term "Artificial Intelligence" may not be important as CS solutions to practical problems are built (you just focus on the better solution), but there is relevance to the "side disciplince" of AI, and from that perspective that is the cone, the scope anyway. "How would an intelligent solver approach the problem".
Game AI uses behaviour trees, usually coded by hand. Decision trees are used for classification and are normally learned from data. The latter are a traditional AI technique from the early days of the modern machine learning era, in the 1990's.
I disagree. AI is doing exactly what it was predicted it would in science fiction.
The computer can now literally talk to you in natural language and then perfectly produce sophisticated actions in response to completely arbitrary and unstructured input. It trivially passes the Turing test. By any definition prior to the year 2023 we are living with Artificial General Intelligence and it’s here now.
Yeah. We'll be arguing "is it really AGI" for many more years. Meanwhile, everyone interesting is going to have moved on from that question, choosing to spend time on "who cares if it's AGI, can it do $foo", for whatever value of $foo is interesting to them. Whether the machine is folding clothes or folding proteins, AGI isn't well defined other than "I'll know it when I see it", so whether or not it's AGI, the question is what job is the machine capable of and is it cheaper than a human? A humanoid robot that can work a warehouse is not putting anyone out of a job if it costs a billion dollars, and neither is a digital AI employee that costs the company a billion dollars either.
So where are the androids? If it's AGI, why is it used as a tool, waiting to be prompted or executed by humans? Where is Skynet? Military applications still rely on human operators.
Yes, but unlike a lot of science fiction, robots, LLMs and other AI remain tools for human use. Augmented Intelligence would have been the more accurate word for real world AI.
I doubt LLMs will give us full embodied intelligence that science fiction androids have. Maybe some other approach. But I suspect for the forseable future LLMs, robots and other AI methods will remain tools, not independent agents like Star Trek Data or Skynet.
"AI" is a term cursed by cool sci-fi implications. It makes it a kick ass marketing term because most people are going to have some familiarity with sci-fi AI and "X media predicted Y technology" is a pretty widespread belief for a lot of values of X (star trek, Hitchhikers Guide to the Galaxy, Arthur C. Clarke) and Y (internet, cell phones, VR). If you want to tell someone we're making big strides in something, linking it into some popsci understanding of sci-fi being the great predictor of human achievement is low effort and high impact for quite a few people.
People aren't trying to communicate accurately if their first priority is getting you excited about the thing!
I have been practicing saying ML for traditional machine learning and LLMs for LLMs for just this reason. Trying not to say AI anymore. Too ambiguous. Sometimes I'm talking about game AI even, I'll try to use shorthand for whatever algorithm I think the AI is using (often I'll talk about its flowchart, though not always sure it's literally using that under the hood).
What is ChatGPT then? Sure it's an LLM, but I can give the app pictures and audio, and it can generate pictures for me. Do we distinguish between the bits of the architecture to accomplish those features separately from the LLM part of the product?
Just as more successful machine learning fields distanced themselves from the term during the AI winter, I suppose we will (and perhaps are?) be seeing them adopt it again, now that we are in an "AI summer".
As always, it's a matter of funding. Both inside academia and outside of it. I remember when nanotechnology was all the rage. Everyone flocked to writing grant proposals about their "nano" technology that was thousands or millons of nanometers, aka micrometers or even millimeters. Stupid but if it works it's not stupid. The old joke is what do you call AI that works? Machine Learning.
The real question is how much compute do you need. With LLMs getting popular, so is compute. That's the real win for non-LLM technologies. The sheer availability of GPU capacity. Yes, it's expensive, but time in a GB300 supercomputer isn't even possible if they don't exist.
Alexnet succeeded for many reasons but a big reason is that computers got good enough to apply those algorithms and techniques in practice. Outside of LLMs, what new AI/ML systems await us in the future? The LLM bubble popping, if it ever does, is going to leave us with supercomputer capacity going unused and available for cheap, meaning experiments that were once infeasibly expensive become practical. I can't afford $10 million to run a weather simulation, but at $1,000 for the same amount of compute, a lot more experimentation becomes practical.
Recently I heard some people conflate procedural generation and generative AI and I had to explain why there isn't some legal or ethical issue with what breaks down to essentially scattering some points.
It's really getting annoying having to have these conversations.
AI is not a real thing or a natural kind but a perspective. Whether something qualifies as "AI" or not cannot be decided by the objective features of the thing. Ergo, it can be defined at the author's pleasure.
> conflating established and morally neutral activities in ML
LLMs are no more or less morally neutral than other ML techniques.
things like the ones I had enumerated in my first comment. I'm old enough to remember FUD about all of those and many others, and perceptive enough to identify many ongoing public opinion campaigns as such.
I was going to post this as well, its delightful to see that other people enjoy it since it was really mind-blowing when I read it.
It's interesting since I saw another comment near yours that raised the question of robustness of the lab-grown design, which I thought was kind of the most fascinating part of the damninteresting article was the revelation that the evolved programs were inseparable from the single physical FPGA used in the training. Since this RFIC training model employs a simulator, do you suppose that the quirks of the physical hardware on which the simulator runs are sufficiently isolated from the training such that a pair of designs would behave similarly when the simulator was run on distinct hardware? And I guess the even more obvious question is whether a design evolved on a simulator would have any hope of behaving as expected in physical hardware?
My hunch about the latter is no, although it still seems like an interesting study, and I often find myself thinking that really understanding what was going on with the FPGAs might be a prerequisite for really understanding how to master reinforcement learning.
Anyway I'm glad you posted this and if you have any other favorites related to this domain send them my way!
the biggest question for me is how robust are these designs.
in the journal articles they did show measurements of real devices which agreed fine with predictions, but i didn't find them addressing it explicitly in the text. also, some systems they presented contained subblocks that were conventionally designed that could be carrying some of the weight.
or maybe i'm just sour that they're coming for my job? or maybe that's what they want us to think?
i think what wins in practice is simple ideas that can work in spite of all manufacturing and environment variations, and model limitations -- think stuff like feedback and symmetry. and what they show here is the opposite of that.
i've done blind optimization of circuit parameters some times only to end up realizing some pretty simple such ideas that i'd missed (like "you need symmetry here" or "you just need more bandwidth here") and made complete sense when you thought about them. so i wonder if we can't tweak a few pixels in their structures and reveal something simpler.
> but i didn't find them addressing it explicitly in the text
Yes, this is exactly what bothers me about this article and about a few similar articles published in the past, that they do not contain any evidence that their claims about the usefulness of AI in design are true.
In TFA it says that the role of AI is replacing the electromagnetic simulator in the optimization process, by guessing the behavior of the structure, which is many orders of magnitude faster than a simulation.
This sounds plausible, but in order to believe this I would want to see the differences between AI guesses and real measurements, in the case of structures with geometries that are very different from those used in the training of the AI.
Also I would want to see exactly with which simulators they have compared the speed of the AI model.
There are various simulation approaches for electromagnetic fields and electronic circuits, that can trade-off accuracy for speed, so I am not convinced that AI inference takes necessarily much less time than some faster low-accuracy methods of simulation, which would still be more accurate and more reliable than AI guesses.
Since you beat me to it, I'll add something that relates relates you were saying on "realizing some pretty simple... ideas".
I think a big plus of computer aided design like this is "innovization"[1]. Somewhat awkward term. But, a system like this leading one to deeper understanding of a particular process is the general idea. It's a fun feeling in practice.
> the biggest question for me is how robust are these designs.
Maybe it doesn't matter?
I mean, of course it matters. But most of this sort of design space is effectively NP-complete, where the creation starts with a blank schematic page and has an impossibly large search space, but where the checking of the design is much simpler.
> also, obligatory mention: "genetic antennas"
Exactly. How does this work? When confronted with the question, of course, everybody gets all excited about the constrained randomness of the GA, but if you think about it, what really makes it work is that there is a comparatively cheap test for fitness for purpose.
One takeaway from the article is that they had to get rid of the tried and tested fundamental building blocks of chip design to generate this advancement. I wonder if the same applies for mundane coding. Are the incredible innovations in AI coding actually hampered by rust and python? Should we let AI tools just code in the lowest level possible?
No. I'm really not fan of Python but you're not going to parse JSON at assembler level. You need to choose right level of indirection for each task. And LLMs are good at more than Python and Rust. I'm using LLMs to write programs in my own language that is compiled by a very fringe language and it happily does everything from scripting on ESP32 to audio plugins and 3D gfx on desktop.
> But freed from the constraints of human-designed templates and the need for humans to even understand the rationale of electromagnetic structures, ... ICs ... can take on truly wild-looking yet efficient designs.
I feel like technology is going to become alien at some point. We're all going to be using magical runes instead of chips.
It's not really that magical. As TFA points out, RFIC design, way beyond normal RF engineering, is close to black magic that relies a lot on the knowledge and experience of the designer, assisted by what would have been supercomputer-level-a-few-decades-ago modelling and design tools. What AI can do is a breadth-first exploration of all possible outcomes and then pick the best-performing one rather than the human-level "this seems like a good path to go down, let's explore it further".
> That’s not even to speak of all the movie plots that would have been ruined.
I clicked on all the links. Pretty much all of those movies could still work with wired technology. Even the one called cellular, in which a woman is trapped in an attic with a broken landline phone and manages to connect wires and dial a random number.
Yes I'm nitpicking. I guess I'm glad we have Wi-Fi and all, but don't try to sell me on it as a crucial plot device
Yep, can't wait until everything is free and costs nothing to generate content. Free hosting and electricity will be super sweet too. Won't need admins or even the Internet. Everything I want will just be free for me because I don't think anything has value.
The methods outlined in this article aren't new. Scientists were using "genetic algorithms" to design antennas that weren't understood by anyone, but worked well, decades ago.
Chips? I've tried to task Opus, Gemini and Codex with a simple PCB. All of them placed holes correctly but can't understand that the traces should not cross physically.
>To arrive at the “sweet spot” where all these different parameters are balanced into optimal harmony, designers will typically lay out several different versions of the circuit, using intuitions and methods they have picked up in their years of training.
I kind of thought the real success is when the designer comes up with key things that are well beyond their training or any training that could have been done up until that time. Based on their years of experience living in an environment where training is table stakes but that's not the thing that's relied upon the most in the end.
With LLMs it seems like odds are that a concept which is statistically insignificant in the training set may surface in place of a truly novel solution, effectively displacing the real breakthroughs that actually go beyond trainable performance.
In a way that decision-makers can not tell the difference, and that could be the worst part.
In case anyone feels déjà vu, Popular Mechanics wrote about this professor's lab in Jan 2025, with almost the same title: "AI Designed Computer Chips That the Human Mind Can't Understand".
I feel a bit of unease when I read this title, not because of the threat of AI, but because the prevailing aphorism that "RF is black magic" is a slap in the face to the millions of physicists and RF engineers who DO understand every bit of this. It's a fun harmless anti-intellectual saw that I don't believe is harmless at all. We need more RF engineers and telling people it's all "black magic" and "wizardry" (and worst of all, saying "even RF engineers don't understand RF") makes it seem like it's not worth studying.
> telling people it's all "black magic" and "wizardry" (and worst of all, saying "even RF engineers don't understand RF") makes it seem like it's not worth studying.
I think the opposite is true. It being advertised as difficult to understand is one of the reasons I personally decided to study RF Engineering. The prospect of learning something so challenging pulled me in. The Smith Chart helped.
I am confused, every day I read on HN that AI's can just interpolate the data they have seen in training, and that they are structurally incapable of coming up with something new, creative and not in the training distribution.
This is analogies to finding a new prime number by brute force using existing maths, rather than inventing new maths to get there.
The AI in this case didn't create a novel technology- it merely used the existing technology without basing the new design on a previous one. The whole "human couldn't come up with it" is because the possible design space is so large, there's no reason a human would start where the AI did.
The thing the AI did better than humans was brute forcing a solution faster. Still a very handy thing to have, but it isn't "creating" in the sense that it invented new materials or fabrication processes or anything novel.
Have you read the article? The creative element came from the researchers:
> In our new approach, the architecture begins essentially from nothing and is progressively assembled through successive iterations. The system explores the design space by generating myriad candidate circuit combinations and mapping the resulting performance trade-offs as it navigates this landscape. Because the process is not biased by prior human design choices, it can produce completely novel circuit topologies that look markedly different from those created by human designers.
> I read on HN that AI's can just interpolate the data they have seen in training
No. That can be said about LLMs, but not about all forms of AI. The technique used is not a LLM.
Sadly we've bastardized the term AI that, if it ever meant anything, it's meaningless now. The currently most voted thread in this post discuses the topic.
In my experience, if you tell them to research the web to see if their idea has been pursued before, you can get them to keep proposing new things until something is sufficiently new, even if it's a new interpolation between existing concepts, that it's effectively an original idea.
This is wrong - the training data is necessary but insufficient. There are a lot of other parts of the architectures used that add a lot of value - otherwise Markov chains would be all you need. There are layers upon layers with non linear activation functions, learned residuals, etc. They still absolutely must interpolate but the space they interpolate through is much more complex than the training data, and they can definitely create things not in their training data. What they can not do is wander outside their non linear parameter space’s convex hull. But this is a really permissive constraint on what they can do “creatively.” People generally under estimate the advantage the architectures confer on that constraint. This is why there was a step function change in expressive power as the architectures (attention, self attention, transformers, diffusions, others) evolved given the same training data. Generally though I challenge you to define “creative” in a way that is precise enough to measure and isn’t self referential or refer to concepts ill defined.
The key tho is can they solve problems not easily solved before with prior techniques. Further can they identify problems not readily presented. Then identify novel solutions. Etc. The answer is emphatically yes they can. These features don’t have to literally exist in their training data, but the supporting highly convoluted network of associations of all their training data does have to in some complex space allow for it to produce these answers. It’s not the same as they’re stochastic parrots at all.
Are they creative? No, because they don’t have awareness. My personal imprecise definition of creative requires both self and awareness as well as free will. There is no driving awareness in all AI architectures, it all derives from extrinsic impetus. Creativity is derived, IMO, from a layer of our minds that is not readily assessed or measured and is only indirectly expressed through language, art, and music. Hence it is not directly trainable and therefore a learning model can’t learn it by reinforcement. It can learn the proxies, but the proxies are not, as we all deeply know, the same as our experienced awareness. We are not our words, our art, our music. We try hard to bridge it, but it’s impossible and you and I know this to be true from experience. In fact we can not even examine our own awareness because it’s not directly observable or possible for us to directly reason about. This is core to a lot of philosophy, especially mid and far eastern philosophy of the mind, the self, the five aggregates of Buddhism, etc. Psychology points at it, and modern psychology avoids it because it’s practically difficult for outcome oriented treatments.
>Generally though I challenge you to define “creative” in a way that is precise enough to measure and isn’t self referential or refer to concepts ill defined.
While I have no hope for a rigorous definition (I don't think it's possible), there are two very distinct kinds of creativity:
1. Result is sufficiently novel for the system itself, i.e. it never seen it previously. This kind is too trivial to even talk about.
2. Result is novel for the side observer. This kind of creativity is meaningless because it depends on at least one unknown (side observer).
https://en.wikipedia.org/wiki/Evolved_antenna
[1] https://en.wikipedia.org/wiki/Cognitive_radio
So, it is very hands-off, but also very expensive, and it is never clear if optimizing the fitness function is worth it, because the fitness function itself may be insufficiently or incorrectly specified.
However, I do think that people should try, even with just a whiteboard or a notebook, to design a fitness-function, for their problem, as if they were going to try to evolve it, because (1) it forces them to explicate their correctness constraints, and (2) they may discover that the program that they are trying to write _is equivalent_ to the fitness function.
I'll give you an example for point 2. Many years ago, I had to parse a gnarly language, and I chose to do it via Chomsky Grammars (that automatically build a tree based on the grammar-spec). Chomsky Grammars are cool, in that they are basically just a state-machine, but they are incredibly difficult to debug: when they work, they might work incorrectly (malformed tree), and when they fail, they give no reason for failure (because even with a trace, you are trying to figure out which backtrack should not have happened). So, out of desperation, I started to consider using genetic programming to just evolve a correct Chomsky Grammar. It became clear that there are only 2 possible fitness functions (1) a function that tests a hand-picked input against a hand-crafted tree-output (which is vulnerable to over-fitting), and (2) a function that is not (well, is much less) vulnerable to over-fitting, but is effectively a pre-existing, correct grammar that can produce those trees.
If you are in situation 2, then the genetic programming is not necessary, unless you are trying to create an optimized (or obfuscated) parser, and even then the optimization may be overfit to the test-inputs (even if they are generated test-inputs from the grammar itself). If you are in situation 1, then you are better off re-evaluating your approach (I abandoned the Chomsky Grammar notation, and invented one that is much easier to understand and debug, without losing any of the expressiveness -- it also happens to be slower, but fast-and-broken is worthless compared to not-so-fast-and-works-fine).
One place where genetic programming has been consistently awesome, is in parameter-search style problems (e.g. your genome is a long list of floats, representing weights and/or anti-weights, and you need to find out which weights give you more fitness (or less error)). I hear good things about variable-neighborhood-search, but have yet to try it.
https://osmarks.net/assets/misc/evolved-circuit.pdf
Starting with: https://sci-hub.ru/storage/moscow/4324/11d145b2c2c3ab320f70b...
The section with oscilloscope traces showing the progression of the “designs” over time was extremely interesting - I’d love to see what the 10x10 grid of functions looked like at each snapshot.
Thank you!
That would be really sad..
I think you're going too far with this. Most people understand scientific theories to be an approximation. F=ma is approximately true, in the sense that it's only accurate within the newtonian regime and each of those terms includes so many asterisks that you will only ever measure it approximately.
The latter is the jokes about the physicists "assuming a perfectly spherical cow."
In fact that's kinda the whole point of the "unreasonable effectiveness of mathematics" essay. It is unreasonable that mathematical approximations are so good at describing our world.
Not to detract from your point at all, but I only ever heard this joke about mathematicians!
Some are useful.
Biology is incredibly well oiled!
The parent is talking more about elegant simplicity vs. sprawling, seemingly haphazard complexity, and you're talking more about durability to failure points and 'completeness'.
Likewise, in code, a lot of the most durable, battle tested software looks extremely inelegant and duct taped, as 90% of the code is dedicated to handling one-off patches and weird edge cases.
When you attempt to hyper-optimize, even with humans in the loop, you end up a mess. You're lucky if you can find clean guiding principles anywhere. If you can hyper-optimize hyper quickly, you end up with an extra layer of mess.
Gaining expertise is always the hard part and our new LLM overlords are making that much harder. So the simple “pure” functions as a teaching aid have never been more important.
End users have never cared about how the sausage is made though.
LLMs can explain complex things to humans with tons of specific context that you don’t find in textbooks or even a google search.
It’s probably never been easier to grasp a large codebase than it is today for example. You can probe and ask specific questions without going through a maze of imports and relationships and config files yourself.
Learning things will always be up to the person, it’s still a choice and dedication to a craft can still be taught.
Occam's Razor is a useful heuristic, but it biases us towards simpler explanations.
I think of elegance as not having to add epicycles, not that everything in the system has to be simple.
Also, without a working theory the, the space of possible solutions is near infinite. LLMs manage to pluck out the space of comprehensible English strings from n-dimensional hell. Even if this is done with a black box of billions of parameters, it’s still elegance in the sense that such a space even exists and was found
Imagine that it's maybe the 1800's and you're asking why somebody who has already survived smallpox is not susceptible to becoming infected again. If you offered an explanation involving tiny detectives wandering around and collecting evidence which they present to each other and decide whether to multiply... one in which the tolerance comes from the detectives from the previous fight still hanging around in your lymph nodes ready to spring into action if they run across the right kind of evidence. Well that would probably be a more complicated explanation that anybody at the time would offer, and it would also be correct.
https://en.wikipedia.org/wiki/Solomonoff%27s_theory_of_induc...
OPs argument is that reality is expressed by very complex equations and interactions; by definition this is outside of Solomonoff induction because it’s easy to imagine this accurate model by definition is the shortest algorithmic explanation, it’s just orders of magnitude more complex than our current approximations.
I guess the argument from OP would look like: "Yes, now imagine we poke and extend our universe as far as we can. How much bigger do you think our final 'shortest description' would be? I imagine it may be orders of magnitude more complex."
Well, I can imagine a squared circle... doesn't mean the math checks out. I would reply that you do not have to imagine, you can go about looking at different mathematically possible universes in Tegmark IV and find the expected number of bits for the one you actually exist in. Which is ~0 bits more complex than the shortest description based on the data you currently have.
Also, note that Newtonian mechanics is not actually a very short theory for building a universe, because you have to instantiate every object in the universe. You actually get a lot more of the structure for free with general relativity (re: Wigner's classification of the particles). An observer in a presumed-Newtonian universe calling it a simple theory would be like saying, "I compressed Wikipedia to one byte, just by putting it all in the decompiler!"
That's the layman's idea of physics theories. They are beautiful and elegant only on the surface, that's why they're technically models and approximations of the real world. The standard model renormalization techniques are a mess of patches and ad-hoc heuristics, pretty far from the "this lagrangian literally contains all physics". Generally you just _ignore_ higher order terms and just call it a day. The famous E=mc^2 it's just the first term of a Taylor expansion. The beautiful form of physics it's what you would call "good enough" and often just a pedagogical tool.
Is this actually true? My understanding was that E=mc^2 is exact for a particle at rest.
I think your point is more that we might be able to initially describe complex phenomena as messy, horrible complex equations, that doesn’t mean we shouldn’t work to simplify them and make them more understandable to us.
I seem to recall legal commentators reacting with an eyeroll—apparently judges split much finer hairs than these for a living—but it was a cute stunt.
[0] https://allthemusic.info/
[1] https://m.youtube.com/watch?v=sJtm0MoOgiU and https://www.the-independent.com/tech/music-copyright-algorit...
Which is not to say let’s not do it anyway and see!
Though I also imagine that that is the point.
Maybe I'm just significantly and unrepresentatively unlucky, but Claude is significantly more intelligent than the average human around me on most any metric I can think of.
by any meaningful measure of intelligence. the latest models are much smarter than the bulk of the population.
how would you define intelligence?
It may seem similarly vague, but it does in fact open interesting, productive, and necessary questions. A "computer" was a professional crunching numbers - "replaced", "easily" because of the deterministic procedural nature of said work, but what about the technical effort to arrive there, and what about the less "mechanical" jobs? When do "processes" become "intelligence"?
Some of us had studied AI originally to study the mind - "how do we formalize thought". It's the interdisciplinary, transversal nature of the area.
Also maybe compare that with that large and important intersection between CS and Economics - the "science of optimization" and its implementation in efficient IT systems. The effort in terms of that different discipline may not be evident, yet lots of engineering is "optimizing" and the generalization of those solutions we call Economics (see the book Algorithms to live by).
So: the term "Artificial Intelligence" may not be important as CS solutions to practical problems are built (you just focus on the better solution), but there is relevance to the "side disciplince" of AI, and from that perspective that is the cone, the scope anyway. "How would an intelligent solver approach the problem".
But as you point out, we used to have human calculators. So is a simple desk calculator a form of "AI"? If so, what type of software isn't AI?
The computer can now literally talk to you in natural language and then perfectly produce sophisticated actions in response to completely arbitrary and unstructured input. It trivially passes the Turing test. By any definition prior to the year 2023 we are living with Artificial General Intelligence and it’s here now.
People aren't trying to communicate accurately if their first priority is getting you excited about the thing!
The real question is how much compute do you need. With LLMs getting popular, so is compute. That's the real win for non-LLM technologies. The sheer availability of GPU capacity. Yes, it's expensive, but time in a GB300 supercomputer isn't even possible if they don't exist.
Alexnet succeeded for many reasons but a big reason is that computers got good enough to apply those algorithms and techniques in practice. Outside of LLMs, what new AI/ML systems await us in the future? The LLM bubble popping, if it ever does, is going to leave us with supercomputer capacity going unused and available for cheap, meaning experiments that were once infeasibly expensive become practical. I can't afford $10 million to run a weather simulation, but at $1,000 for the same amount of compute, a lot more experimentation becomes practical.
It's really getting annoying having to have these conversations.
AI is not a real thing or a natural kind but a perspective. Whether something qualifies as "AI" or not cannot be decided by the objective features of the thing. Ergo, it can be defined at the author's pleasure.
> conflating established and morally neutral activities in ML
LLMs are no more or less morally neutral than other ML techniques.
Also, CFCs.
One of my favorite little morsels of internet goodness.
It's interesting since I saw another comment near yours that raised the question of robustness of the lab-grown design, which I thought was kind of the most fascinating part of the damninteresting article was the revelation that the evolved programs were inseparable from the single physical FPGA used in the training. Since this RFIC training model employs a simulator, do you suppose that the quirks of the physical hardware on which the simulator runs are sufficiently isolated from the training such that a pair of designs would behave similarly when the simulator was run on distinct hardware? And I guess the even more obvious question is whether a design evolved on a simulator would have any hope of behaving as expected in physical hardware?
My hunch about the latter is no, although it still seems like an interesting study, and I often find myself thinking that really understanding what was going on with the FPGAs might be a prerequisite for really understanding how to master reinforcement learning.
Anyway I'm glad you posted this and if you have any other favorites related to this domain send them my way!
in the journal articles they did show measurements of real devices which agreed fine with predictions, but i didn't find them addressing it explicitly in the text. also, some systems they presented contained subblocks that were conventionally designed that could be carrying some of the weight.
or maybe i'm just sour that they're coming for my job? or maybe that's what they want us to think?
i think what wins in practice is simple ideas that can work in spite of all manufacturing and environment variations, and model limitations -- think stuff like feedback and symmetry. and what they show here is the opposite of that. i've done blind optimization of circuit parameters some times only to end up realizing some pretty simple such ideas that i'd missed (like "you need symmetry here" or "you just need more bandwidth here") and made complete sense when you thought about them. so i wonder if we can't tweak a few pixels in their structures and reveal something simpler.
also, obligatory mention: "genetic antennas"
Yes, this is exactly what bothers me about this article and about a few similar articles published in the past, that they do not contain any evidence that their claims about the usefulness of AI in design are true.
In TFA it says that the role of AI is replacing the electromagnetic simulator in the optimization process, by guessing the behavior of the structure, which is many orders of magnitude faster than a simulation.
This sounds plausible, but in order to believe this I would want to see the differences between AI guesses and real measurements, in the case of structures with geometries that are very different from those used in the training of the AI.
Also I would want to see exactly with which simulators they have compared the speed of the AI model.
There are various simulation approaches for electromagnetic fields and electronic circuits, that can trade-off accuracy for speed, so I am not convinced that AI inference takes necessarily much less time than some faster low-accuracy methods of simulation, which would still be more accurate and more reliable than AI guesses.
Since you beat me to it, I'll add something that relates relates you were saying on "realizing some pretty simple... ideas".
I think a big plus of computer aided design like this is "innovization"[1]. Somewhat awkward term. But, a system like this leading one to deeper understanding of a particular process is the general idea. It's a fun feeling in practice.
[1]: https://dl.acm.org/doi/10.1145/1143997.1144266
Maybe it doesn't matter?
I mean, of course it matters. But most of this sort of design space is effectively NP-complete, where the creation starts with a blank schematic page and has an impossibly large search space, but where the checking of the design is much simpler.
> also, obligatory mention: "genetic antennas"
Exactly. How does this work? When confronted with the question, of course, everybody gets all excited about the constrained randomness of the GA, but if you think about it, what really makes it work is that there is a comparatively cheap test for fitness for purpose.
I feel like technology is going to become alien at some point. We're all going to be using magical runes instead of chips.
That said we’ve had some success internally having Claude do parameter sweeps
I clicked on all the links. Pretty much all of those movies could still work with wired technology. Even the one called cellular, in which a woman is trapped in an attic with a broken landline phone and manages to connect wires and dial a random number.
Yes I'm nitpicking. I guess I'm glad we have Wi-Fi and all, but don't try to sell me on it as a crucial plot device
I kind of thought the real success is when the designer comes up with key things that are well beyond their training or any training that could have been done up until that time. Based on their years of experience living in an environment where training is table stakes but that's not the thing that's relied upon the most in the end.
With LLMs it seems like odds are that a concept which is statistically insignificant in the training set may surface in place of a truly novel solution, effectively displacing the real breakthroughs that actually go beyond trainable performance.
In a way that decision-makers can not tell the difference, and that could be the worst part.
I feel a bit of unease when I read this title, not because of the threat of AI, but because the prevailing aphorism that "RF is black magic" is a slap in the face to the millions of physicists and RF engineers who DO understand every bit of this. It's a fun harmless anti-intellectual saw that I don't believe is harmless at all. We need more RF engineers and telling people it's all "black magic" and "wizardry" (and worst of all, saying "even RF engineers don't understand RF") makes it seem like it's not worth studying.
I think the opposite is true. It being advertised as difficult to understand is one of the reasons I personally decided to study RF Engineering. The prospect of learning something so challenging pulled me in. The Smith Chart helped.
The problem isn’t the design: its manufacturing restraints.
This is nothing new or impressive.
The AI in this case didn't create a novel technology- it merely used the existing technology without basing the new design on a previous one. The whole "human couldn't come up with it" is because the possible design space is so large, there's no reason a human would start where the AI did.
The thing the AI did better than humans was brute forcing a solution faster. Still a very handy thing to have, but it isn't "creating" in the sense that it invented new materials or fabrication processes or anything novel.
> In our new approach, the architecture begins essentially from nothing and is progressively assembled through successive iterations. The system explores the design space by generating myriad candidate circuit combinations and mapping the resulting performance trade-offs as it navigates this landscape. Because the process is not biased by prior human design choices, it can produce completely novel circuit topologies that look markedly different from those created by human designers.
No. That can be said about LLMs, but not about all forms of AI. The technique used is not a LLM.
Sadly we've bastardized the term AI that, if it ever meant anything, it's meaningless now. The currently most voted thread in this post discuses the topic.
The key tho is can they solve problems not easily solved before with prior techniques. Further can they identify problems not readily presented. Then identify novel solutions. Etc. The answer is emphatically yes they can. These features don’t have to literally exist in their training data, but the supporting highly convoluted network of associations of all their training data does have to in some complex space allow for it to produce these answers. It’s not the same as they’re stochastic parrots at all.
Are they creative? No, because they don’t have awareness. My personal imprecise definition of creative requires both self and awareness as well as free will. There is no driving awareness in all AI architectures, it all derives from extrinsic impetus. Creativity is derived, IMO, from a layer of our minds that is not readily assessed or measured and is only indirectly expressed through language, art, and music. Hence it is not directly trainable and therefore a learning model can’t learn it by reinforcement. It can learn the proxies, but the proxies are not, as we all deeply know, the same as our experienced awareness. We are not our words, our art, our music. We try hard to bridge it, but it’s impossible and you and I know this to be true from experience. In fact we can not even examine our own awareness because it’s not directly observable or possible for us to directly reason about. This is core to a lot of philosophy, especially mid and far eastern philosophy of the mind, the self, the five aggregates of Buddhism, etc. Psychology points at it, and modern psychology avoids it because it’s practically difficult for outcome oriented treatments.
While I have no hope for a rigorous definition (I don't think it's possible), there are two very distinct kinds of creativity:
1. Result is sufficiently novel for the system itself, i.e. it never seen it previously. This kind is too trivial to even talk about.
2. Result is novel for the side observer. This kind of creativity is meaningless because it depends on at least one unknown (side observer).