>Much of economics is built on models that were designed to provide a formal, logical basis for understanding the economic world, which is often quite unintuitive.
I disagree. It's purposefully unintuitive.
>(For example, many intelligent people uneducated in economics exhibit intuitions opposite of basic economic ideas like opportunity cost or comparative advantage.)
Most people don't believe in comparative advantage. They believe in something that economists can explain away as comparative advantage.
All unconsumed fixed size investments will result in something that is mathematically the same as comparative advantage. This is the intuitive view that people have. You go to university and get a 5 year degree. Now your cost basis for work that suits your expertise is much lower than for work that is out of expertise. A worker buys an expensive machine, now the cost basis for hiring the guy with the machine is lower than buying your own machine.
This also explains why specialization emerges: All specialization is basically a form of an investment that has some residual left over results that can be monetized in the future. If there was no residual it would be as if you forgot your education and at that point the investment is fully consumed and you turn back into a non-specialized worker.
All of this is incredibly intuitive, but economists instead insist on an invisible "factor" [0] to drive efficient production.
[0] The "factor" concept implies comparative advantage exists first rather than emerges as a result of past decisions.
>My broader point is just that there's nothing inherent to the structure of LLMs that stops them from updating their weights and continuously learning from environmental feedback in the way humans do, and there's already solid templates for how they could push even further in that direction.
"LLM" is a branded model as a product. Of course it could be anything, as long as it fulfills the product category.
But we live in reality, we can only look at what models are out there and we see that they don't do any of those things and yet we're supposed to act as if these models already do.
What can a transformer not do that people say they can do?
The parent comment said, paraphrasing, "learn from interaction with the world", and I'm responding, they absolutely can already do this by taking their logs of interaction with humans and updating their weights through backprop.
The reason you don't see that done "live" is primarily an economics problem rather than a limitation of the model structure.
Alright let's assume your premise is true, that transformers can learn from interaction with the world by updating their weights - then why isn't this done?
Because backprop fundamentally wants the entire data set in every pass. It doesn't behave well and is destructive when you update after the pre-training phase. RLHF/LORA are attempts to work around that and effective at what they do, but it is not learning in the sense you are talking about and also do not fully address the catastrophic forgetting problem. This architecture as is - is not compatible with continual learning.
It's not "not quite true", it's literally true because alternative architectures like RNNs and Mamba fully update their own internal states, whereas transformers only append to the context.
RNNs and Mamaba do not update their weights, but you could hypothetically scale the internal state to be as big as Fable's and GPT 6's parameters.
At least one mechanism to update transformers' internal states already exists, there is nothing stopping anyone from performing backprop after every session.
It just has big technical and economic challenges. But I expect advances there. There have actually already been big advances, though done in bulk fashion (RLHF).
If you predict an event that is 100 years away and 30 years pass, does the event that is still 70 years into the future vanish simply because the present arrived as predicted by the forecast?
Mind you, the forecast told us that we're still fine in 2026.
Places that used to be 32C in summer are consistently hitting 36C+. It absolutely sucks to be outside in those temps. You are running from aircon to aircon. Doom scenarios of unbearable heat is definitely true.
Germany had two sizable heatwaves this summer, of which one was so intense that for the first time in about "ever" ACs tripled/quadrupled in price (note this includes those on online stores where the shipping times would allow for trucking them in from all across central Europe)... I'm not referring to installation labor/technicians even. But like window units and similar ones that don't rely on HVAC technicians to take out the box and turn on.
That doesn't relieve people of the responsibility of having enough scientific literacy to understand that 30 years is only an instant on geological time scales. And the fact that we see meaningful change at all on this time scale is very concerning.
Ah yes, I believe one of those models was called the Strawman Model, generated by Elliot F. Strawman during his tenure at MIT. He was forced into retirement due to his deep embarrassment over his model’s inaccuracies.
Anyway, Illinois is about to break the all time record for tornadoes by a single state, previously held by Texas, during the strongest El Niño ever measured. I’ll be proud tucking my kids in and telling them as long as they pray hard enough to the Gods of technological progress, we’ll be guaranteed to stumble upon a carbon sequestration technology akin to the Haber process and they won’t have to be forced into a Sophie’s choice when there’s only enough clean water for 2 of their 3 children.
Worth noting Texas is having one of its quietest tornado years in the ~20 years of comparable data, at the same time. Which is the problem: the reliable detection era is about 30 years, and the preliminary dataset behind these headlines starts in 2004. The IPCC has low confidence in observed trends in severe convective storms for exactly this reason. And your El Niño point argues against you, since El Niño is associated with suppressed spring tornado activity in the southern Plains, which is what we're seeing in Texas.
From what I read Texas will have larger and more intense heat domes. On the plus side, this high pressure shifts tornado alley towards the east away from Texas. On minus side, Texas will be the new Arizona.
I was responding to the claim that many models had predicted catastrophe. An Inconvenient Truth is a movie made by a politician. It, of course, chose to depict the more extreme ends of probabilities. And despite that, I think it stands up pretty well, all things considered.
TALA is terrible at respecting the "direction: <up, down, left, right>" directive. It should choose an aspect ratio based on the direction, but instead it is trying to make the diagram as square as possible.
Georgism makes it possible for the government to profit off the infrastructure it builds, which makes it possible to fund public transportation projects without the tax payer having to pay income tax.
Instead the people who benefit from the infrastructure have the bill for the public transportation system rolled into their land value tax.
Right now in the current system you pay both income taxes and then you pay the landlord again because to him as the monopoly owner of land, the public infrastructure is pure producer surplus.
The demographic that would be favoured under that system is people who work for a living and the people being disadvantaged are the ones sitting on land or extracting the producer surplus.
If the AI companies ever get the idea that it would make sense for the AI model to modify its KV cache, it would end up adding a write step to the KV cache per token (moderately expensive) and increase the cost of inference by the width of the update (arbitrarily expensive).
This in itself could double or triple the demand for both memory bandwidth and compute. I honestly don't see a future where AI demand disappears...
I would argue that the primary use for this type of prototyping is to test analog chip designs, because analog components don't scale down that well. Digital information only needs to distinguish between 0 and 1 but with analog electronics the maximum current or voltage you support grows as you use a bigger area or thicker layers.
I personally don't believe that even a $100 service delivering you 1 micron chips could compete against a $10 FPGA unless you are intentionally doing something the FPGA was not designed for like analog electronics or low static power.
I don't have the impression that a $10 FPGA can implement the equivalent of an 80386. I can believe it of a $1000 FPGA. I don't know what the right intermediate point would be. But big FPGA's are very expensive, mostly because the customers for such things are well funded.
I partially agree with you, in the sense that any useful custom integrated circuit must include some analog part, otherwise it would not be competitive with standard programmable logic.
Nonetheless, having actually worked as an analog IC designer, I can assure you that today, and especially when using a CMOS IC process, where the analog devices that are available, like MOSFETs, are inferior to the bipolar junction transistors and junction FETs, which were available in the fabrication processes traditionally used for analog ICs, it is completely impossible to make a competitive analog circuit that is purely analog.
Any CMOS analog circuit needs a digital part, even if it is something as simple as an amplifier, and more so for things like a power supply or a motor controller, or something that acquires data from analog sensors.
At the minimum, a digital automaton together with a lot of configuration switches is needed to perform the auto-calibration of the analog parts when the IC is powered on. Otherwise, things like differential amplifiers would be unbalanced, bias current sources and amplifier gains would have values very different from those needed for the circuits to work as designed, etc.
Digital parts may also implement in a simpler or more reliable way various kinds of control feedback loops and various protections to undesirable conditions, e.g. overcurrents, overvoltages etc.
So any useful custom CMOS circuit must be a mixed digital-analog design, i.e. it must include both a digital part, with flip-flops, digital gates and switches, and an analog part, with amplifiers, comparators, analog multiplexers, oscillators, voltage references, etc.
An one-micron CMOS process would be especially useful if its component list would include some NMOS transistors with a higher breakdown voltage, of at least 12 V, but preferably even of 20 V, or ideally of 25 V (to have a safety factor if using an 18 V power supply, as required by the gate drivers of some power MOSFETs).
An one-micron CMOS IC would likely use 3.3 V for the main power supply, and it should be able to use 5 V at least for I/O buffers. But if some transistors with a higher breakdown voltage would also be available, they could be used for open-drain output buffers that would be tied to an external higher voltage and they could be used to command directly some external power MOSFETs, without an intermediate gate driver. A standard 5 V I/O buffer could command directly an external GaN HEMT (gallium nitride switching transistor), but presumably the output current of the buffer would be low, so the switching of the external transistor would be slow in comparison with using an additional gate driver IC.
It's Poe's Law in action. We can't tell if you're unhinged enough to post that or you're imitating people who would post that to make fun of them.
reply