ACCount39
A "reorder the vector components" transform preserves the similarity perfectly. So does a transform that maintains orthogonal equivalence.
That's exactly what Platonic Representation Hypothesis would predict.
There also have been plenty of studies showing that various neural networks, including LLMs, develop implicit world models that reflect the geometry/structure of a task they're trained on. A trained "chess engine LLM" would have a "chessboard-shaped" world model inside it.
So, if you can figure out what the geometry of a task is, you can search the weights for correlates of that geometry. Then you can try to improve those correlates so that they track the task better.
Seems like bleeding edge LLMs might already be superhuman at "figure out geometry from looking at weights".
You're making a mistake there - focusing on the tip of the iceberg without considering what lies beneath it.
The bulk of what the human brain does isn't "decide whether skub is good or bad". The foundation that the human cognition is built upon is vast, and mostly invisible from the inside. You don't get to see the feature matching operators of your visual cortex, or the patterns that decide whether a sound is "soft" or "sharp". And a lot of those foundational features are assembled by biological development programs spun up all the way from DNA.
On that level, when it comes to the bulk of neural circuits and the processes implemented by them, the representations within a "Pro Skub" and an "Anti Skub" guy might easily be 90%+ compatible. The two people who had vastly different life experiences, worldviews and opinions would be more alike to two different LLMs fine tuned from the same base model than they would be to two different architectures trained on completely different datasets.
When it comes to representational similarity analysis, we're trying to isolate the representations from the "coordinate systems" they sit in. If you take an image, and make a second image by rotating it 90 degrees and mirroring it, then, for the purposes of platonic representation hypothesis, they're the same image. Once you get past the "coordinate transform", the similarity between the first representation space and the 90-deg+mirror replica of it is total.
A mapping function between LE and BE can be constructed easily, and it's going to be lossless both ways. Which means that those two representations are fully compatible.
Now, INT16 and FP16? Those are far less compatible representations. There will be an overlap subspace where lossless conversion between the two is trivial, but this compatibility window will be relatively narrow.
Observe an unknown word used a few times in different contexts, and you'll get a very good bead on what it means.
It's the "natural" way of learning a language. No one gets every word explained to them in detail - most words you know in your own first language are the ones you picked up on the go.
I think this "structural redundancy" might be a big part of what enables LLMs to learn human languages "from nothing".
Ironic that it holds the first place by understandable lines. The vocabulary isn't the complex thing - it's the content itself that eludes understanding.
Surreal and dreamlike - almost as if inviting you to connect the dots, but not giving you the labels, not giving you enough dots or an obvious enough outline to connect them confidently.
The best time to build up nuclear power was 4 decades ago - when the supply of fossil fuel was obviously fraught, and things like solar and wind power were nonviable.
Whether it makes any sense to push for nuclear power now is unclear. Nuclear was quite expensive even back then - but decades of NIMBYism, "environmental" activism and overregulation have damaged the industrial capability and eroded supply chains to the point that getting a buildout going at scale and on a sane budget and a sensible timetable would be an uphill battle. At the same time, renewables aren't anywhere near as opposed, the economic outlook for solar, wind and grid storage is better than it ever was, and the situation is likely to improve even further.
Unlike fossil fuel power plants, NPPs don't depend on a steady supply of a high volume of costly consumables just to keep running. Reactor fuel loads last months to years, naturally buffering any price shocks or supply chain issues. And you can maintain an "energy security" stockpile of nuclear fuel that would add you even more years of runaway.
To have your nuclear power disrupted, you have to face an adversary that's both capable and willing to hit targets on your soil, and on top of that, is willing to target NPP infrastructure specifically.
A random rogue state lobbing mid-range drones over a strait wouldn't cut it. Neither would a war halfway around the world.
Intelligence is a tool - one that can be put to use against any problem. "Get diverse people to come to the table for compromises" is no exception.
It's invisible, insidious censorship. They don't hit your hands with a ruler and say "stop, no buying that, bad". They make it so that what they don't like doesn't get to be on the store shelves in the first place. Hard to buy what isn't being sold, isn't it?
As for why they do it - it's unclear, but there's speculation that payment processor leadership might be in part ideologically captured by the usual "think of the children" moralizers and puritans. The same types that used to spurn moral panics, rally against D&D and later GTA, and are the first in line to propose sacrificing freedoms for imaginary safety - now allowed to wield covert influence over most of the payments that happen online.
This by itself is a reason to push for more crypto acceptance, in my eyes.
If not for the success of Raspberry Pi, Broadcom would have zero SBC presence today. Like many companies chasing the high volumes and high margins, they don't see the market value in a segment like this. Thus things like poor docs and weird blobs. Raspberry Pi is open to the degree it is despite Broadcom, not because of it.
Ironically, that Broadcom tie could have caused RPi to eventually lose ground to other SBCs - because the likes of Rockchip just deliver more bang for buck in SoC dept. Broadcom doesn't seem to be investing in staying competitive in the segment. So RPi's edge is just the ecosystem being more important than raw price/performance.
For now though, that's pretty much on hold, because SBC price tags are increasingly DRAM dominated instead.
So as long as you can test things on cell culture level? You can do a lot with a high throughput setup.
The issue, of course, is just how much time, effort and skill is required to do it.
It might become a viable commercial process outside third world countries if RAM prices spike even more - or if someone finds a way to teach an off the shelf $1000 robot arm advanced autonomous microsoldering.
Then there's the problem of forcing the device itself to recognize the new RAM. It's a device-specific trick. RAM sticks have a little tag chip that tells a PC "how much RAM this stick has", but RAM chips themselves don't - so the device has to infer its RAM capacity from elsewhere. Like strap resistors on the board. Which you have to find and adjust for the new RAM capacity. That, or fix the RAM values in firmware, which is also device-specific.
I've been eyeing LPCAMM2 interposers for non-RAM uses - as an off the shelf connector for stacked PCB designs. The conclusion was that it wasn't produced or used at sufficient scale for it to work.
A server that both holds KV cache in its own memory, and uses its own compute to perform attention operations. That way, you don't have to transfer cache entries server to server for every attention sweep - only the "requests" and "responses". Attention bandwidth stays server-local.
The "no root access, no bootloader unlock, no installing your own Apple-unapproved software" bullshit is mandated by Apple, with no opt-out.
This shouldn't exist in the first place. There would be no need for it if Apple didn't insist on "we own all the iPhones, our users just get to tap at them".
If you could just unlock the bootloader, re-enable SWD and do what you want with an iPhone that you actually own, there would be no need for this restrictive program at all.
Causality is uncertain, but the "Jevons paradox" interpretation is certainly not contradicted by the available data. Anecdotes of per-employee AI spend rising, and multi-agent freerunning AI tasks becoming more common are also plentiful.
Personally, I think that the "Jevons paradox" interpretation is basically correct. Humanity doesn't have a limited demand on intelligence - but humans are expensive, so it's supply-constrained. The cheaper intelligence gets, the more of it ends up used. As AI gets cheaper and more capable of substituting for human intelligence, usage will rise.
But that's a bit of a problem too?
Because, unlike humans, LLMs have no trouble whatsoever recalling all the mnemonics. Hell, they can often recall all the opcodes and encodings too and disassemble from hex "manually".
So if the primary advantage of this syntax is better human-readability and human-understandability? The need for it in a future where low level assembly routines are written by LLMs in a perf evaluation loop isn't that great.
I think it's a very promising approach - it makes assembly a lot more intuitive, especially to people with background in higher level languages. It makes listings far more immediately readable, even without memorizing every instruction mnemonic. But it just came in too late. By the time Blackfin happened, coding in assembly was firmly an afterthought.
I'd like it to succeed, but I have my doubts that it will.
It's dead obvious that companies like Apple and Google and Sony and John Deere lock their hardware down to protect their revenue streams - not to help the users in any way. And yet, they're peddling the lie of "security" - and have been getting away with it for decades.
Now that AI is there to give even an average user a leg up when it comes to things like "I don't want to see ads in my TV's menus, can you do something about it", the companies will see more and more reasons to implement more "security" to stop it. And modern "security" measures are hard to defeat with technical measures - so this has to be curbed with legal action instead.
The reason why I'm not "team sensors" is that we're nowhere near to being able to as much as squeeze out even one mid-range camera sensor for all its worth. Things like AI assisted scene reconstruction show just how much data is there in even a small series of photos - when the AI is there to back the raw measurements with sensible assumptions.
The difference between "a plausible reconstruction" and "an accurate measurement" is academical rather than practical. You never get perfectly accurate measurements (because real world sucks), so a well designed "plausible reconstruction" will often have better error averages AND bounds than a naive measurements-first pipeline - simply by taking better advantage of spatial and temporal redundancies. Adding more assumptions to the pipe from there gives you even better performance.
I expect approximately none of the near-term gains in robotics performance to come from the sensors themselves. There's still room for improvements there, mind. There is some cool sensor tech that could result in a better performance ceiling. But we're nowhere near that "sensor ceiling".
Today's robotics sensors are so far ahead of what robotics AI can reliably utilize that trying to push for "we need better sensors" is a mistake in nearly every case.
The power of modern AI robotics is that, much like the biological mess it's trying to replicate, it doesn't care that much about absolute sensing.
Imperfect sensors fused by a high performance control model with a good foundation of world dynamics and self dynamics? Gets you impressively close to "perfect sensors" in practice. Being able to take a big bundle of noisy nonlinear ambiguous measurements, and implicitly reference them off each other and over time to get to a usable control loop is powerful.
It's the kind of power human nervous system wields to force the horrid mess that is the human body to perform.
For "detect when you've accidentally picked up two items from a bin instead of one", a wrist camera is a far more natural sensor than a bundle of actuators - especially for lightweight items. Wrist cameras are awesome. But even a bundle of actuators, if fused well enough, can often do it.
The problem is that designing algorithms to sensor fuse to this level by hand is pretty much a no-go. The kind of sensor wrangling hell I would only wish upon an enemy. Which gets us right back to general purpose robotics being an AI problem at its core.
In practice, if you want really good performance across an extreme dynamic range, it's usually easier to stack multiple specialized sensors within a single sensor unit. A filter-free reduced resolution sensor for high QE in low light conditions (or at high framerates), and a daylight-optimized sensor that gets high resolution, full Bayer, IR-cut and a polarized glare filter to get you as much data as it can in high light environments. This gives you straight up superhuman range with no adaptation time.
Do the raw low light images look nice? No. But human brain itself takes a lot of postprocessing for "looks nice" to happen. Salvaging good images from series of noisy measurements is an AI problem. We've had some pretty capable sensors for a while now, but it took AI-powered computational photography to truly make low light photography happen on a smartphone.
I'm aware of the "detection of individual photons at above-chance rates" experiments with human vision. But modern backside illuminated CMOS, if plugged into the same setup, would also be able to detect photons at above-chance rates. Not to mention the actual photon event cameras. Very niche items, but they can detect photons pretty reliably - not merely "more likely to be right than wrong".
There are also thermal imagers and image intensifier stages, if you want to take the "niche extreme sensors" route. Not commonly done. But goes to show just how much ass engineered sensors can kick.
The limits aren't in the sensors - it's in making sense of the data they give you.
Backdrivable actuators know how much current are they sinking - and thus, how much force are they facing. So even the simplest two finger gripper can easily "know" that one object is squishy and the other is not. The response curves are different. Same can apply to arm actuators for weight estimation.
Now, what this doesn't actually get you is an estimate of how hot/cold an object is, and how moist an object is.
Where cameras come in is specifically the "guide a grasp" part. Force sensing alone can tell a robot "the grasp is failing". A camera pointed at the grasped object can tell "why".
But, just like being able to build a load cell and a thermometer doesn't translate to "skin-like sensing", we can't go from "chemical sensors" directly to smell.
For one, things like sensor diversity, density, longevity and manufacturability matter a whole lot. Can you give a robot a weather station type "pressure + temperature + humidity + PM etc" air quality sensor? Yes. Can you give that robot a sensor with enough molecular coverage and fidelity to replace the human sense of smell? Almost certainly not. Can you even get close on a sensible budget? Not with the tech we have.
Getting a system to perform molecular sensing with discrimination against many molecule classes - the kind of operation human smell performs - is very much not trivial.
Now, whether that's worth pursuing at all - I'm still leaning towards a "no". It's a very niche capability.