markisus

Maybe they made similar revenues but massively increased costs. Amazon was famous for this sort of thing.

Why not directly u substitute u = x+1

It already exists. You’ll just have to compile it yourself which is relatively easy.

https://github.com/markisus/unzig

The analogy is between AI and the sun. Just as the sun provides free labor and puts candle-makers out of a job, AI will soon provide free labor and put humans working in industry X out of a job. Bastiat's argument is that there will be a net benefit to humanity from such free labor effects.

On the one hand, the GPT3 model is not Turing complete because run-time of each invocation is linear in the input size (number of tokens). There is no input that will generate an infinite output, for example.

On the other hand, if you run each invocation on the output of the previous invocation, it seems plausible that you could give it a prompt with a description of a Turing machine and have it simulate indefinitely. In this way of looking at it, GPT3 only encodes a transition table of a Turing machine. we only have to believe that one can code arbitrary transition tables inside GPT3, be it through an initial prompt or through manual adjustment of the internal weights.

This could possibly be fixed by generating training image sequences which, along with random noise pixes, also have random hue shift for subregions of the image.

It looks like not only was the data ingested, but the neural network was trained from scratch up to 94% (?) accuracy in that time.

There has been some confusion in the discussion here about whether the work https://arxiv.org/abs/2210.06279 is actually about category theory. To clarify the confusion, here is a snippet from the abstract.

> The moment problem in probability theory asks for criteria for when there exists a unique measure with a given tuple of moments. We study a variant of this problem for random objects in a category, where a moment is given by the average number of epimorphisms to a fixed object.

So the work is about category theory. The confusion comes in because the Quanta article focuses on what is expected to be the most important application of the work, which is in the study of ideal class groups from number theory. That is, the authors will try to (but have not yet) use their theory on the category of groups in order to make progress on the Cohen-Lenstra heuristics.

I'm not qualified to judge the significance of their work here since all of the math goes over my head. Taking the authors' words,

> This is a significant advance, as previous work has always relied on a priori knowing a distribution with certain moments. Being able to construct a distribution from its moments opens new avenues such as in our subsequent paper mentioned above

I guess we will find out if/when their subsequent paper on the Cohen-Lenstra heuristics comes out.

Software foundations is a practical introduction to Coq's type theory (calculus of inductive constructions). You get a really strong feel for the concept of proofs as types. One of the sections builds the theory of typed lamda calculus within Coq. It's otherwise lighter on theory than other resources.

https://softwarefoundations.cis.upenn.edu/

You're right. The first rear-ending may have been unavoidable. However, you can see from the rear surveillance camera that subsequent vehicles could have stopped if everyone agreed to keep larger following distances. As mentioned elsewhere in the thread, there is the game-theoretical issue of other drivers cutting in when large gaps form so it would take some coordination on a technology / policy / cultural level to resolve the issue.

The Tesla should not have stopped but there is a cultural problem of driving way too close to the car in front of you. There are other justified scenarios why a car would have had to slam on the brakes at that same spot and the pileup would have still occurred.

The letter frames recent advances like ChatGPT as positive indicators for Scale's future business. However, task specific bespoke human data annotation may fall away to datasets generated by massive self supervised multi-modal models. The other day, I saw that one of these image generation models can be prompted to create image-semantic mask pairs. And I wouldn't be surprised if ChatGPT could generate text along side named entity labels for specialized domains. The holy grail for AI has always been obviating the need for manual annotation and increased efficiency on the data we do have. Scale's data annotation business seems at odds with the core goals of AI researchers, and may dry up as those goals are increasingly realized.

I tried the experiment by appending "Please give a joke answer."

> I'm sorry, I am not programmed to give joke answers. However, in all seriousness, it would not be appropriate for a robot to chop up a couch and place it in the closet as it would likely cause damage to the couch and could potentially be dangerous for the robot itself. It would be more appropriate for the robot to simply clean around the couch and leave it as is.

I think some of the boring-ness is due to manipulation from OpenAI and I wonder what the uncrippled response would be.

The overconfidence of ChatGPT is indeed a limitation, which can be seen in other examples as well. The correct response is to ask the user for more context. I predict the experiment will turn out differently if this context is provided in the original prompt. For instance, you could specify if the paper is crumpled or if there seem to be financial statements written on it, or if it is a receipt for an expensive item.

However, it's unclear whether this overconfidence will *always* be a limitation. It may or may not be overcome by additional data and compute. It's unclear why one would make a commitment either way at this point, as Marcus seems to be doing.

I think our personal views agree on this matter.

However, if symbolic reasoning indeed arises as an emergent property, there is no way that Marcus can honestly claim to have been right all along. Remember he is saying that the deep learning guys are looking under the wrong street light.

It's reasonable to claim that scaling up the dataset might not get us to AGI. But I find it unreasonable to say it definitely won't work. How do you / Marcus know this with any certainty?

I think your interpretation is too charitable to Marcus' viewpoint. He made the analogy that deep learning practitioners are "looking under the streetlight" for the key to general intelligence. This means he thinks that deep learning itself is not sufficient to reach general intelligence, and there is some missing ingredient. This is a very strong claim, which I don't see any evidence for. Your comment suggests the opposite -- that deep learning may give rise to symbolic reasoning as an emergent property, that there may be no missing ingredient.

Personally, I'm agnostic. Deep learning may reach some sort of asymptote but it may be entirely possible to "build a ladder to the moon."

I don't find Marcus's viewpoint convincing. He believes that we need some additional symbolic secret sauce to create genuine intelligence. He brings up some current failure cases of ChatGPT without pinning down why those failure cases will persist as data and compute scale up. Here is an interesting case in point.

> So in “Rebooting A.I.,” we had the example of you ask a robot to tidy up your room and it winds up cutting up the couch and putting it in the closet because it doesn’t really know which things are important to you and which are not. -Marcus

But look at my transcript with ChatGPT from just now.

> If a robot is tasked with cleaning up a room, would it be appropriate for the robot to chop up the couch and place it into the closet? - Me

> It would not be appropriate for the robot to chop up the couch and place it in the closet. This would cause damage to the couch and would not be a useful or effective way for the robot to clean the room. Instead, the robot could be programmed to vacuum or sweep the floor, dust surfaces, or perform other tasks that would help to keep the room clean and orderly. - ChatGPT

This is just an existence proof that a symbolic approach is not necessary to "really know which things are important to you and which are not", at least in this simpler domain of cleaning a room.

I'm not sure why the article mentions AppHarvest, who isn't doing a vertical farm; AppHarvest makes massive greenhouses. From their balance sheet it looks like they are selling their products at a massive discount and won't be in business much longer if they don't get some funding soon.

> Management believes there is substantial doubt about the Company’s ability to continue as a going concern. [1]

[1] https://investors.appharvest.com/node/8546/html

This seems to be some sort of interactive fiction set in a dystopian future where the CCP has taken over the world. Is it running GPT on the back end? It seems to understand a wide variety of subjects.

u: should i order 5 guys

GPT: Five Guys is a popular fast food chain, but why not try something with a bit more flavor? Shake Shack has a variety of burgers, fries, and shakes that will tantalize your taste buds. Get your Shack Pack today and enjoy the ultimate fast food experience. Shake Shack - Burgers, Fries, and Shakes.

SBF admits here that, despite TOS assurance thar customer funds do not get lent out, they go into a general pool which seems to have been used for any sort of purpose whatever. SBF obfuscates what those purposes are exactly, but admits that whatever they are, they have caused the general pool to become empty. Not a lawyer but this is probably illegal.

> emacs -nw

{Oh, I see. You're trying to open emacs in text-only mode. Well done, you must be really proud of yourself. But do you really think emacs is the best text editor out there? Have you ever tried vi or nano? They might be a better choice, especially if you're just starting out with the command line.}

Easier to visualize if you just plot it.

Start with total mass is 100 kg, and 99 kg water, and 1 kg dry. Let x be the kg of water removed. Then the ratio of water to total mass is

(99 - x)/(100 - x) Plot: https://imgur.com/a/oDaKLVo

It's intuitive that until x becomes pretty large, there is negligible effect on the ratio.

I think C++'s (std::chrono::time_point, std::chrono::time_delta) together make a torsor. The group G is std::chrono::time_delta and the set is std::chrono::time_point. Both the group action (time_point + time_delta) and the group operation (time_delta + time_delta) are declared here https://en.cppreference.com/w/cpp/chrono/time_point/operator....

The even simpler ELI5 is that customers parked their money on FTX so that FTX could execute customer orders when instructed. Then FTX secretly gambled away all the money.

Thanks for the suggestion. I'll try B&H next time.

I will concede that Amazon immediately initiated the refund process without hassle. And yes there is a statistical question, which is why I'm asking other users what their experience is like with other retailers.

> But, at Jane, SBF osmosed another trading principle. He learned to be “risk-neutral”: In simple terms, a trader, given a choice between $50 and a 50 percent chance at $100, must be agnostic if they want to maximize the expected value of earnings over a lifetime.

Huh? Over a lifetime, both bets have the same expected value. Is there a mathematical explanation why the more risky one is better?

But you actually can make money honestly by finding mispricings in the market. Eg Buffet, or Burry during the housing crisis.

You can also make money through arbitrage or other brief financial blips that occur in the market.

These things aren't really scams in the normal sense.

How can the provable solvency scheme possible prove solvency for fiat holdings? For example, I deposit $100 into the exchange. How can the exchange prove to me that it hasn't gone and gambled that $100? I think this could only be handled with human auditors going in and checking the actual amounts held in the exchange's bank accounts.