foota

I was thinking more along the lines of https://xkcd.com/378/ :)

Finally, I can write my code as it was intended.

In theory if people are giving all of the money to Nvidia, then it's because they find it valuable (in this case companies purchasing GPUs) so even if the GPUs weren't so expensive in theory there would still be economic value derived from the transactions.

Wouldn't it be better to enforce pass via the harness and then measure cost/quality since that's more what people are likely to do?

I guess my suggestion is my being more Ukraine like

I wonder if it wouldn't be better to have a Lowe level of constant military spending and have the ability and reputation to be able to quickly spin up and scale military production, so that you don't need to maintain as much military spending 24/7

Hah, I had basically the idea your company is based on :) It's a neat thing to try out, it makes me hopeful for the future.

The auto version is extra horrible because what it reads to you is just a summary of what the model actually said. I don't ask it anything when driving because it's too frustrating.

How are failure and leadership handled in choreographic programming? Is it up to the runtime to decide? E.g., if I start a new Bob (Bob prime) in a Bob and Alice exchange but the old Bob still exists, what happens? I guess the transport, failure handling, and leadership are all outside of the scope of these libraries? (someone linked a similar looking chorus library).

"Wind and solar power do not provide dispatchable power and they therefore do not replace conventional power plants. Germany isn't keeping their coal-fired power plants and they're not building new gas-fired power plants despite the energy transition but because of it."

4 hour storage battery operated renewables are now cost competitive with gas plants.

Honestly? I don't think so. If anything I think they would obscure this using averages (e.g., maybe they solved some hundred problems and the easy ones took a few minutes and the hard ones took a couple of days), but if they have a model that's as good at math as they are as coding I don't see why it'd be impossible.

I think novel mathematics is intrinsically harder than coding (much in the way that computer science research is harder than coding), but the models are scary good at coding.

My theory here is that providers cover the non-constant costs of output tokens as context length caries using the cache input fees.

Yes

From their github: "The average result used the equivalent compute of roughly three hours of ChatGPT Pro thinking." That's pretty crazy.

Not to mention: "retain, cache, or store Output beyond what is reasonably necessary to display such Output to your authorized end users in the ordinary and real-time course of use" which would seem to preclude storing it in a long lived session.

"A bank run happens when everyone tries to withdraw all their money at the same time and the bank runs out of cash. Not really possible in the stock market where companies literally can create/destroy shares and there is a whole secondary pricing layer to it." Sure, it's not possible for stocks to become insolvent in the same way, but if a large shareholder sells it could trigger a panic and greatly decrease the price.

I don't think so. Large surface area helps with convective cooling I think by increasing the surface area that participates in heat exchange with the air (or other thermally conducting material), radiative cooling wouldn't benefit from this because you can't concentrate light beyond the source that it's emitted from (etendue).

Though I do wonder if it would be possible to have some kind of internal heat pump driven by electrical power to juice up the temperature of the radiators to increase the power being radiated away? E.g., run a heat pump to increase the temperature of a working fluid and then run high temperature radiators? I think it would work and I don't immediately see that it would violate the laws of thermodynamics? (this is ignoring all practically, I'm sure the engineering would be devilishly hard, although if you're already shooting for the moon you might as well throw in some artificial gravity to boot, it's not like the robots get motion sickness)

Getting this all up into orbit it obviously the hard part, but if you're already building so much solar capacity the cooling actually doesn't seem unreasonable?

It's a little crazy they haven't tried to sell it off yet. Maybe they don't want to trigger a bank rush or they're limited in their ability to do so?

I have a bit of an obsession with trying to find hot springs. Washington state, for all of its mountains, is relatively devoid of hot springs. I feel like there must be some that remain unknown. I'd like to someday use a thermal camera on a drone of some kind to try and find one :) Maybe go along fault lines that have other hot springs?

Economic background: profit maximizing businesses would like to extract the most value out of every transaction. Not everyone derives the same value from the same thing. E.g., someone might be willing to pay $5 for a burger, whereas others might only pay $4. Price too high and you lose the people that will only pay (or can only afford) less. Price too low and you don't charge people as much as they would have been willing to pay and lose out on profit. The economic POV is that voluntary transactions happen because both sides benefit (or at least don't lose out) or they wouldn't happen. Therefore every voluntary transaction produces some economic benefit in the form of producer and consumer surplus. The amount of these is determined by the gap between the price and the consumer's WTP (willingness to pay, i.e., the highest price they would be willing to pay) and the price and the producer's WTA (willingness to accept, i.e., the lowest price they would accept at). Together these form the economic surplus of a transaction.

Profit maximizing businesses want to capture as much of the economic surplus of transactions as possible by optimizing the price they charge. When businesses offer a single price their ability to do so is limited because some people with a lower WTP that is still above the producers WTA don't elect to purchase and on the flip side some people who have a higher WTP would be willing to pay more and don't.

To increase their profits therefore businesses can attempt to do what's referred to as "price discrimination" which is when they offer different prices to people based on the person's perceived WTP (there are different means of doing so, such as geographically based pricing, etc.,) and when they offer exactly the customer's WTP to every unique customer it's called perfect price discrimination, because they're capturing the entire value of all transactions.

In competitive markets, businesses ability to price discriminate is reduced, but not entirely eliminated.

Now... this surveillance pricing is basically a form of price discrimination. However, the interesting part is that while price discrimination in net is beneficial for businesses, it actually can also benefit lower income/lower WTP consumers by allowing them to buy at a lower price (since they wouldn't have bought at a higher price -- both the consumer and the business benefit here) but hurts customers with a higher income/WTP since the business can charge them more.

This is interesting because this is a somewhat rare regressive (hurts lower income people more than higher income people) anti-business policy. Generally, I think most anti-business policies are also progressive (well, except for the idiotic ones like broad tariffs) but in this case banning the ability of price discrimination through personalized pricing hurts businesses and lower income people while benefiting higher income people (the surveillance aspect of it could be thought of as an externality, which hurts everyone).

If you're highly anti-surveillance you might argue that it's net positive for everyone because lower income people wouldn't be surveilled in the same way (well, at least it wouldn't be applied, I don't think it would actually change the surveillance side of things) but that requires a normative position on whether surveillance is bad.

In theory this policy could probably be made non-redistributive (benefitting higher and lower income people equally) by adding a grocery tax that would be used to offset the impact to lower income people, but in practice it seems like it would be difficult to administer (especially in Seattle, which doesn't collect city taxes from people directly today, not to mention the opportunities for arbitrage).

The idea that I'm proposing basically isn't an embedding (which is context independent) but rather combining the embedding model with the context of the LLM. It sounds like the embedding model here is normally a "vision transformer" that maps image chunks into tokens with positional embeddings for both the position in the context as well as position within the image. Maybe it could be given the ability to consume the context and decide to emit multiple tokens for a single chunk? So for example, let's say that I ask a question "How many blades of grass are in this image" (a hard question for traditional image embeddings in models since the embedding won't contain the information). If the proposed architecture was both aware of the context and able to "decide" to emit multiple tokens for an image, then for the above it could emit tokens that represent the answer to the question posed, instead of just being the embedding of the image. Or you could ask something like "How many green pixels are there" and again I think it would work better under this architecture than it would otherwise.

I'm not sure how practical it is to train that architecture though or whether there would be performance issues.

Isn't think only true if getting enough vitamin D is as important to you as cancer?

This /looks/ at least more official. Most unofficial Google projects have a disclaimer in the repo.

I think what I'm saying is that I don't understand why the embedding exists. I assume it's some kind of training and inference cost issue? But why can't the Gemma architecture linked above just learn to represent pixels in the LLM model's embedding space directly, rather than having the embedding from 48 x 48 pixel chunks? Or rather, give the embedding model some context to produce the embedding? (Which, as you note, wouldn't really be an embedding anymore, but seems like it would better understand fine detail)

I'm not an ML expert, but I was thinking of a sort of "guided" embedding. E.g., give the image model some prompt for what it's trying to do? I don't understand why multimodal models generate an embedding that doesn't understand what the model is trying to "figure out".

I think this is similar to how Gemma 4 12B is implemented, but even then I don't think the single layer image embedding is "aware" of the context.

I feel like multimodal models that can read images should work differently than they do. My understanding is that multimodal models basically first generate an image embedding and then the model is trained to interpret that embedding, but in the same way that text is lossy, it seems like the embedding would be as well. Why don't multimodal models learn to interpret images themselves without an embedding? Or e.g., by passing some "prompt" to the embedding model?

Jokes aside, I think the idea is that the law is simple to code, the proof that it holds is where the agent is responsible. This probably becomes less true though as you try to express more complicated laws.

A generation is a time that people were born between and useful because it embodies much of the cultural context that people share.

Funny enough I was thinking about something very similar to this based on the Jev model posted yesterday.