I, for one, welcome ML being back in the "AI" fold. The sooner "AI" expands back into the full spectrum of fields of study that it used to mean, the better. Imagine what ML could do with the massive amount of GPUs now deployed everywhere, I've been putting the bug in folks' ears that where they're wasting money on chatbots trying to do math, they could be training some awesome ML models that would have far more precision for those tasks and not have to pay a subsidy every month into perpetuity.
I've been doing a deep dive into the state of the art for OCR models recently, and it boggles my mind how many people are resorting to vision LLMs when dedicated OCR models eat them for lunch with two orders of magnitude fewer parameters.
Also a case study on how few resources are being dedicated to the compute stack for traditional ML workloads these days-- as far as I can tell there's no reasonable way to run a decent OCR model on an accelerator card without doing a ton of your own plumbing. This is the kind of thing that should be available out of the box like running CrapModelOfTheDay4-GarbageTier on an AMD NPU is.
I suggested the far less hysterical title ‘Applying AI to the "Dark Art" of RF Chip Design’. Even the summary makes it clear that there’s no true artificial intelligence here:
Princeton researchers use reinforcement learning and inverse design to rapidly create RFICs from scratch.
Diffusion models rapidly generate novel or human-interpretable RF layouts, achieving record performance and drastically reducing design time.
Hate to break it to the marketing hype machine, but this is regular old machine learning, just with the new spice of transformers incorporated into the dish.
This is reminding me of Adrian Thompson's work on evolving a tone discriminator that was so tuned the the silicon it was evolved on that no one could figure out how it worked other than "quantum effects of the underlying substrate." If this keeps up, we could be in for a Dark Age of understanding.
That's a super weird paper. The writing is very matter-of-fact, but the result presented is outlandish.
Has that work ever been replicated? It seems to me, a programmer without an electrical or electronics engineering background, to be a simple experiment, and thus likely to have been tried again, but I've never seen anyone claim a similar result, or claim practical uses of a similar result.
I don't know if it's been replicated. All I know is that the result was quite outlandish, but (in my opinion) believable, as it involves evolving a solution.
Given that the most optimistic outcome is something tied to a specific model of FPGAs, and given that there is always a risk of the next batch being up to the original public specification but a different model in terms of parasitic electromagnetic coupling — I would expect a reluctance to depend on such tricks for anything practical…
People who could easily reproduce it are limited to those having an FPGA board lying around without use — maybe those people do not in fact find this result outlandish or even surprising?
On the one hand, for the communication part «electromagnetic coupling» is a bit more specific. On the other hand, of course this depends on the exact behaviour of transistors when used «wrong», and transistors depend on quantum effects to work, so strictly speaking any off-label thing with transistors depends on the details of quantum effects in the substrate.
we could be in for a Dark Age of understanding
… I am not convinced that we are not there anyway.
For decades now, we have way too many huge software systems of which people say that those systems do what those systems do, and tracing behaviours from engineering first principles is not always tractable without natural-science-style experiments.
And some physical things are already designed by gradient-descent…
This isn't "AI", this is plain ML. IEEE should know better, but I suppose that gets the clicks these days.
I, for one, welcome ML being back in the "AI" fold. The sooner "AI" expands back into the full spectrum of fields of study that it used to mean, the better. Imagine what ML could do with the massive amount of GPUs now deployed everywhere, I've been putting the bug in folks' ears that where they're wasting money on chatbots trying to do math, they could be training some awesome ML models that would have far more precision for those tasks and not have to pay a subsidy every month into perpetuity.
I've been doing a deep dive into the state of the art for OCR models recently, and it boggles my mind how many people are resorting to vision LLMs when dedicated OCR models eat them for lunch with two orders of magnitude fewer parameters.
Also a case study on how few resources are being dedicated to the compute stack for traditional ML workloads these days-- as far as I can tell there's no reasonable way to run a decent OCR model on an accelerator card without doing a ton of your own plumbing. This is the kind of thing that should be available out of the box like running CrapModelOfTheDay4-GarbageTier on an AMD NPU is.
IEEE has been publishing these paper-like things since forever. First with genetic algorithms, then NNs.
I suggested the far less hysterical title ‘Applying AI to the "Dark Art" of RF Chip Design’. Even the summary makes it clear that there’s no true artificial intelligence here:
Hate to break it to the marketing hype machine, but this is regular old machine learning, just with the new spice of transformers incorporated into the dish.
This is reminding me of Adrian Thompson's work on evolving a tone discriminator that was so tuned the the silicon it was evolved on that no one could figure out how it worked other than "quantum effects of the underlying substrate." If this keeps up, we could be in for a Dark Age of understanding.
That's a super weird paper. The writing is very matter-of-fact, but the result presented is outlandish.
Has that work ever been replicated? It seems to me, a programmer without an electrical or electronics engineering background, to be a simple experiment, and thus likely to have been tried again, but I've never seen anyone claim a similar result, or claim practical uses of a similar result.
I don't know if it's been replicated. All I know is that the result was quite outlandish, but (in my opinion) believable, as it involves evolving a solution.
Given that the most optimistic outcome is something tied to a specific model of FPGAs, and given that there is always a risk of the next batch being up to the original public specification but a different model in terms of parasitic electromagnetic coupling — I would expect a reluctance to depend on such tricks for anything practical…
People who could easily reproduce it are limited to those having an FPGA board lying around without use — maybe those people do not in fact find this result outlandish or even surprising?
On the one hand, for the communication part «electromagnetic coupling» is a bit more specific. On the other hand, of course this depends on the exact behaviour of transistors when used «wrong», and transistors depend on quantum effects to work, so strictly speaking any off-label thing with transistors depends on the details of quantum effects in the substrate.
… I am not convinced that we are not there anyway.
For decades now, we have way too many huge software systems of which people say that those systems do what those systems do, and tracing behaviours from engineering first principles is not always tractable without natural-science-style experiments.
And some physical things are already designed by gradient-descent…