>> and that is that we understand (or we think we understand) how our brain/mind works. But the truth is that we don't know. And there's even not a single clue that we actually know too much, and not a clue that our brain/mind and cells work 'as the machines we build'.
>> I highly recommend people in the AI research space should read philosophy and modern linguistics.
I highly recommend the philosophers read some neuroscience. The whole "model weights" thing in AI is modeled after the synaptic connections and between actual neurons. There is already quite a bit known about how the brain works at a low level. There is also a lot that is still unknown. There are also differences between discrete neuron firing and weights as signals, but there is enough similarity to make artificial neural nets useful and do things similar to what real one do.
Hard disagree. We only discovered the role that glial cells play in processing around 2014. We're still uncertain how patterns of activation consolidate through long term potentiation, let alone how signaling encodes information. We understand quite a bit about the role of the hippocampus and subiculum in encoding memories; but we don't understand the structural layout of engram complexes - which were themselves mapped for the first time only in 2022!
Taking effective results in machine learning, and somehow assuming that they apply to cognition - simply because neural nets were inspired by our limited knowledge of neural signaling and structure - is like trying to apply aircraft engineering to studying ornithology. For a better articulation of this point (from the reverse direction) check out the paper 'Could a Neuroscientist Understand a Microprocessor?' from 2017 - https://journals.plos.org/ploscompbiol/article?id=10.1371/jo...
>> We understand quite a bit about the role of the hippocampus and subiculum in encoding memories...
Hard disagree ;-) You're talking about high level architecture of the brain. I don't think (not my area I may be wrong) we know how memories are encoded in a real brain. Is it weights or something else? If it's weights that's supporting my point (but we don't know what the weights represent in a brain, where in LLMs many weights are just token encodings). If brains store memories in something other than weights I'd really like to know as it's something I haven't read about yet.
> There are also differences between discrete neuron firing and weights as signals, but there is enough similarity to make artificial neural nets useful and do things similar to what real one do.
There is barely a surface-level similarity. The best example I can come up with is this…
Imagine the most intricate and beautiful tall building that you can think of. Think like an older skyscraper in Chicago or a palace. There are water features and moving parts everywhere but also tiny little handmade carvings and materials throughout.
Now imagine we have no reference designs and no blueprints - we hire an architect to attempt to study the building by looking at it from a distance and understand everything they possibly can about it. She can go into the building to check but every time she does, it stops functioning normally.
That architect is a neuroscientist.
Then the ML researcher is like a graphic designer who sees the work that the architect is doing and makes a napkin sketch of the building the architect has been studying, to use for a project later. Sure the designer has some of her representations. But the difference in complexity between the designer’s napkin sketch and the architect’s analysis is massive. Several orders of magnitude.
Then another many orders of magnitude is the level of detail that the architect can understand about this strange building without being able to fully interact with it, versus the actual complexity of the building.
So yeah, an AI is modeled after neurons in the sense that they represent a couple of surface level features of neurons. But the difference in complexity is about as much as a napkin drawing of a grand building represents the actual structure and details of the building, no matter the level of skill that the graphic designer has
> I highly recommend the philosophers read some neuroscience
Philosophers do read neuroscience and technical reports on AI.
> The whole "model weights" thing in AI is modeled after the synaptic connections and between actual neurons
In reality, it is a really poor and basic model of what is actually happening in a real brain
Brains and modern AI systems (LLMs for example) are structurally different.
(Don't get confused by topology. A structure is more than topology: it is also what the structure is made of, thus the properties of the material contribute and define what can emerge atop the structure)
Yes, but it looks like there is no hardware floating point. The description of the CORDIC module indicates fixed-point calculations, which is consistent with the lack of any reference to floating point.
I am happy the have CAN-FD and Motor PWM module, but nowhere did I see conversion times listed for the ADC. For motor control I demand 1uS conversion time or less, and in the last year I've switched from fixed point to floating point after holding off on that switch for ~15 years.
The datasheet apparently doesn't say, but judging by their other products' listed 12 bit SAR ADC sampling rates (and assuming this one is similar to what appears to be their standard ADC ) the conversion time will be on the order of 10uS.
It's not that important if you use current sensors on the motor phases. But then you're looking at HALL sensors or a shunt with a very high gain amplifier with good common mode rejection - looking for mV signals on top of a +12V or +48V square wave at PWM frequency.
By using low-side shunts under each half-bridge you don't need the common mode rejection, but you can only measure phase current while the low side FET for that phase is on. That means limiting the PWM duty cycle to ensure that FET is on long enough to measure current, so we trade available voltage range for sample time.
I've also written code to measure all phase voltages with a single low-side current shunt under the whole 3-phase bridge. That requires careful phase shifting of the PWM signals and very fast conversion time, but you don't have to compromise available voltage range 0-100 percent duty cycle is possible.
Typically we run the control loop at PWM frequency, but the measurements need to be faster than that.
Field-oriented Control schemes modulate phase currents at high frequency; the feedback loop must be much faster than the motor phases. Until fairly recently, this stuff was the exclusive province of dedicated ICs (Trinamic et al.) and FPGA. Today, FoC can be done in (mostly) software with MCUs.
Fast feedback loops are also necessary in SMPS, another area where precision, low latency MCU peripherals and software are actively displacing traditional approaches.
But even if you update your PWM signal on every PWM cycle, you won't go much beyond 30kHz. At some point you're running into high switching losses on your MOSFETs.
The closed loop experiences a phase margin loss that is exponential with the frequency. At lower frecuencies it is negligible, but if you get close to the frequency of the delay the phase margin reduction becomes dramatic and the control goes from stable to unstable very fast.
If the sensor has a limited bandwidth, you add the conversion delay and then the computation delay on top of that you end up with a max workable loop bandwidth in the low tens of kHz and anything higher will have overshoots, oscillations, etc.
You see this in low cost products like MKS SERVO42x, where they're doing FoC with a GD32 MCU. It works; the motor runs cool, smooth and quiet, but the system is limited to 3000 RPM, and struggles with rapid acceleration because the control loop is too slow.
I have tried one. It has no torque. For what looks like an awesome product, it does not have the power to drive a peristaltic pump. I used the same motor on a TMC stepper controller and it's completely silent and works. It's open loop, so comparing apples to oranges but I am not sure what the MKS servo driver on a motor could actually do, aside from spin unloaded.
These can deliver 2.5-3A/phase, which should ample for a pump. Respectfully, I wonder what motor was involved and whether the current was configured: they come out of the box with conservative configuration so people don't burn up motors.
This is exactly correct. Low 10s of kHz is quite functional for machines moving in human space / speed.
If one is trying to do some assembly line (max # of operations per second), the power requirements alone get hard to manage. And then you're managing back EMF, eddy currents, heck, air resistance!
My rule: have dedicated low-level hw provide smooth PID response, mostly on the P term; and have a higher-level control produce the setpoint. Faster response means less need to rely on I or D terms as much (just because delta-T is so relatively small).
If you don't encode locally as the video is created, you either need to store RAW frames which takes enormous amount of storage, or you use a different format and suffer quality loss by transcending.
> you use a different format and suffer quality loss by transcending.
Compressing to AV1/h264/265 etc is really only done for the final version, but that doesn't mean that videos are stored in RAW format during editing, where it is very common to store frames locally in Apple ProRes, Avid DNxHD, or some other compressed format that's targeted towards professional editing.
Contrary to AV1 or whatever similar format which offer compression ratios of 1000x and more, these formats have a compression ratio of around 10x. They are very simple, and the quality loss is low enough that it doesn't matter. They also tend to store images with 30 bits per pixel instead of the 24 bpp that's normally used for streaming.
You’re not wrong but I do think it’s worth clarifying that any professional production with a budget, even a modest one, is generally being shot with a raw codec -> edited/colored with proxies -> rendered with the original raw codec where they compress for the final cut.
ProRes and the like are used for proxies or quick and dirty productions that are mostly shooting their look in camera because of a fast turnaround time. This is usually event work on a budget or something for social media.
Again, you can grant this and a huge number of agriculture subsidies still aren't justified.
People have an instinctive defensiveness over farms/farmers, but anyone who has studied farm subsidies in any depth knows there's no way to rationally justify huge swathes of them. I don't know anyone with the requisite knowledge who wouldn't agree with that including farmers and lobbyists (because they generally only like a subset of the subsidies themselves).
I’m from a historically agricultural state, and live in the farming area. Government interventions are regularly mocked
- always have been.
Demand for food plummets when it is no longer fresh. Throwing food away is politically toxic. This creates major problems.
As people get richer they don’t want more food, they want better food. Fresher and more meat based. Which is fine. But means the food when you talk about food “that which prevents us from starving to death” you are quite divorced from actual demand.
People don’t price food based on its anti-starvation capabilities.
Either they follow traditional diets, or they buy for convenience (highly processed), or they are health nuts who live off rice, beans, and kale.
Nobody is trying to maximize calories. Very few people are trying to match their food intake to their amount physical exertion.
All these ontological and teleological models are divorced from how food is actually valued: market “taste” is insanely important under normal circumstances.
Our agriculture sector won’t succeed if it’s based around preventing famine.
>> a structure where 50 to 90 percent of the purchase price is financed by debt, and that debt is loaded onto the balance sheet of the acquired company, not the firm making the acquisition.
This just seems wrong. The buyer takes out a loan, how does that become the responsibility of the company they purchased? I thought loans used to buy a business treated the business as collateral, like a home mortgage. What lender would participate in this? and why?
> The buyer takes out a loan, how does that become the responsibility of the company they purchased?
Because the company they purchased is now a part of them.
As for why a lender would agree to it, it's because these transactions are not as simplistic or universally disastrous as they are usually described. A lender will obviously only make that loan if it has a reasonable expectation of being paid back, and most of them are. They may get additional collateral like parent/affiliate guarantees and the loans will have covenants relating to financial performance etc.
One more reason I'm sticking with my current system for as long as possible. Though if my neighbor pays to bring the pipe the rest of the way to my house on the way to theirs I'll re-evaluate.
"The useful part is not that the secret is hard to compute from too few shares. It is that too few shares contain no information about the secret. With one share missing, every possible secret is still possible."
Reminds me of factoring numbers with the Quadratic Sieve or its variants. You find a system of congruences mod n that eventually allow you to compute prime factors, but until you have enough of them that isn't possible. I've often wondered... Each congruence must contain some information right? What space are we reducing degrees of freedom in?
Same thing here, each piece restricts the space of polynomials, but does not restrict it enough to tell where the key crosses the axis.
EMDR therapy was invented by Francine Shapiro. She found walking to be effective on herself, and later found that most any form of "bilateral stimulaition" at the right pace will cause certain kinds of brain activity.
>> I highly recommend people in the AI research space should read philosophy and modern linguistics.
I highly recommend the philosophers read some neuroscience. The whole "model weights" thing in AI is modeled after the synaptic connections and between actual neurons. There is already quite a bit known about how the brain works at a low level. There is also a lot that is still unknown. There are also differences between discrete neuron firing and weights as signals, but there is enough similarity to make artificial neural nets useful and do things similar to what real one do.
reply