Version467

It's actually running the code. It doesn't run all code it generates. But if you specifically ask it to, then it does. It also has access to a bunch of data visualization libraries if you want it to calculate and plot stuff.

I wouldn’t have any confidence in any predictions I make 100 years into the future even if we didn’t have the current AI developments.

With that said, I’m pretty confident that the movie industry will exist in 10 years (maybe heavily tranformed, but still existing and still pretty big). If it’s still a big part of current popculture by then (vs obviously on its way out) then I’d expect a collapse of it to require a change that is not a result of AI proliferation, but something else entirely.

Nothing about this idea makes any sense at all. Classic Musk moment.

This is basically the same thing as websites running crypto miners in the background, just for cars.

I wonder how he comes up with this stuff.

Hollywood has incredible financial and political power. And even if fully AI generated movies reach the same quality (both visually and story wise) as current ones, there’s a lot of value in the shared experience of watching the same movies as other people, that a complete collapse of the industry seems highly unlikely to me.

You can generate images through the api, but edits and variations have not made it to the dall-e 3 api.

We don't actually know how big the Dataset is, right? It could be the same dataset used for Llama 2, but trained for more Epochs.

Just to clarify, when you say never. Do you actually mean never (or some practical equivalent like ~100 years), or do you mean not right now, but possibly in 5-10 years?

I'm just asking to try to build some intuition on what people who actually train soa models think were capabilities are heading.

Either way, congrats on the launch :)

not your point at all, but you say almost all of the chips are Intel chips. I was under the impression that literally all of them were Intel chips. Have they started manufacturing chips for others?

As far as I can tell, the only cryptocurrency that actually delivers on its name (i.e. being used as a currency) is Monero. Sure, it's all drugs and stolen credit cards, but it does undeniably solve a real world problem for its users instead of just being used as a vehicle for speculative investment.

With that said, I think if anyone comes up with a "killer-app" for crypto, then it'll be on the Ethereum chain. They seem to be the only ones who consistently work towards adding capabilities to the core technology.

Edit: I realize I haven't commented on the article at all. This sentence stood out to me:

> Today, we have all the tools we'll need, and indeed most of the tools we'll ever have, to build applications that are simultaneously cypherpunk and user-friendly. And so we should go out and do it.

Clearly, this is an important step. But the two examples he provides as a beacon of what's possible (Daimo and Farcaster) don't inspire a lot of enthusiasm. Daimo is just a decentralized version of Venmo and Farcaster is a protocol to build social networks on the blockchain, which is yet another tool and not an application.

I do still like reading Vitaliks thoughts. He's a pretty good writer, and it's evident that he spends a lot of time actually thinking about the topics he writes about.

On the one hand, you're right, the price seems absurd. On the other hand, I've only ever heard this argument from people who don't really want such a device anyway.

I don't own this phone, but I paid $500 for a Boox Note Air 2, which is an eink tablet that I use for notetaking and reading and replaced my ipad with it. The limited usability of these devices is a selling point, not a downside. But just because I want a device that does less doesn't mean that I want a device that feels cheap. I still want something sleek, well built and snappy. And eink screens aren't particularly cheap either. So $500 it is.

You're right, it doesn't. But most of the books claims don't depend on a specific climate sensitivity. Climate Change is just one of the chapters. It also covers air pollution, deforestation, agriculture, biodiversity loss, ocean plastics and overfishing.

It combats the notion that we aren't achieving anything regarding long term sustainability because we're unable to coordinate on a global scale, by showcasing that we are not only moving in the right direction in many areas, we're also accelerating.

If Hansen is right and we live in a world with very high climate sensitivity then it's not looking good, but I'd still prefer to live in a world were we make noticeable progress and fall short, than in one were we watch the world turn into an uninhabitable ball of death while doing nothing.

It gets easier the more you do it. As you become more comfortable with the software as well as develop an intuition for what makes a good card, the effort shrinks dramatically. Not just because I'm faster at creating cards, I also create fewer cards overall.

Even if they were, writing cards yourself forces you to engage with the material in a way that's highly conducive to learning it. It makes knowledge gaps more visible and forces you to understand it if you want to have any chance breaking the material down into pieces that make good cards.

The positive effects of SRS rely just as much on the repetition as they do on the creation of the cards. I know that not everyone agrees with this, it's certainly been true for me.

Premade decks might work for language learning, but even then I'd be wary.

Because I perceive the cost of memorizing it to be almost zero.

There's of course a need for nuance here. I do write scripts to automate stuff and I do have aliases configured for certain commands, directories, etc. I don't try to remember all the things, just so that I never have to look something up again.

If I wasn't a heavy anki user already, then I certainly wouldn't set it up just to remember some commandline stuff. But that's not the situation I'm in. I use it for lots of things already and I have an established habit to review the cards every day. Adding a card to remember specific flags has incredibly low overhead for me.

A different, but similar example is using Anki for learning library/framework syntax. I made heavy use of it when I first learned Pytorch. There are so many different commands to wrangle your data into the shape you need it to be. I frequently mixed them up and got frustrated because it doesn't necessarily result in runtime errors and then it's hard to debug (because I didn't know what I was doing). So everyday I chucked a couple of the new commands I encountered into Anki and by the end of the week I was comfortable doing all kinds transformations. Would I have learned to do that without Anki? Probably. Did it give me an additional opportunity to consolidate that knowledge for very little cost? Also yes.

You're free to dislike the book, but I can't really follow your criticism. The book is clearly not "based on vibes". It's chock full of data to support the authors claims.

Now you might argue that the data either doesn't support the claims made, or you might doubt its legitimacy, but instead of providing a reason you just told me that it should be obvious why the book is nonsense. That doesn't really give me a lot to engage with and I'm not going to try to make your arguments for you.

For me the Book had its intended effect. It argues that the climate movement, as it gained mainstream popularity, also lost its ability to convey nuance in favor of projecting a strong "we are heading towards certain doom" message. This helped grow the movement, but is also responsible for a lot of young people feeling completely hopeless. That's certainly something I could heavily relate with before reading the book. The book then goes on to highlight areas where the world has made much bigger progress than I would have thought. For me this turned a feeling of intense hopelessness into one of motivation.

So was the Book meant to make me feel good? Sure. In that sense I guess you can call it a feel-good book. Were the claims unsubstantiated and "without any evidence to support" them? Certainly not.

Not necessarily. Hypothetical example, but let's say I use some obscure tar command twice a month. That's too far apart for me to remember it, but often enough that it would annoy me to look it up every time. Chuck it into anki and I have a short time where I repeat it often enough so that it sticks and after that the time between repetitions quickly becomes far longer than the 14 day interval where I actually use the command.

And now I never have to look it up again :)

So I have never actually used Anki for languages specifically, but I use it heavily for almost everything I deem important enough that I want to remember.

There are two very common reasons why people lose interest/motivation after trying out Anki. The first one is creating too many cards when you first start out. People get excited by SRS and then start creating cards for a lot of things, especially for stuff they don't actually care to remember. This is understandable, since you can't do SRS when you have nothing to repeat. But that gets overwhelming quickly, because it's hard to be motivated if you don't actually care about the material.

The solution for this is to only create flashcards when an opportunity arises organically. This has the additional benefit of your review sessions being extremely short in the beginning, which makes it easier to establish the habit of reviewing.

The second problem that often leads to people quickly giving up is not knowing what makes a good card. It's actually not as easy as one might think. Especially not when you want to get more out of SRS than just rote memorization of trivia. There's an excellent article from Andy Matuschak on this topic that explains it way better than I ever could https://andymatuschak.org/prompts/

If you specifically want to do language learning, then you might be tempted to download premade decks. This might work a little bit better than for everything that's not languages, but in my experience making a card is at least as important as the repetition itself. It forces you to distill the knowledge down into good cards, which is only possible if you engage with the material and also helps you find gaps in your knowledge. So I'd recommend against it.

Regarding your comment on adjusting the repetition algorithm... don't. It's highly unlikely that you'd be able to improve on the defaults if you don't have a good feel for how slow/fast you forget things. Even then, it's pretty difficult to make good adjustments. It's much more likely that the flow you're missing is just from not doing it long enough.

Hope that helps :)

There’s a new book from Hannah Ritchie called „Not the end of the world“, that argues for optimism, without denying the reality of the situation. It’s a fantastic book and I highly recommend it.

Not really. They already chose to show the benchmark where it does best and even then it’s still quite a bit worse (though definitely impressive for its size). If you take a look at other benchmarks, for example MMLU@5-shot then this does 46.3, while gpt-3.5 does 70.

But there might be some use cases where this one is close enough in performance and the difference in cost and speed make it a better choice.

Absolutely. We got a usb c iPhone, apple will have to allow sideloading in the very near future, there’s some repairability stuff in the works, etc. All in all, as an EU citizen I very much feel like the regulation has been a big benefit for consumers.

I actually wouldn’t be too sure about that. Some dark patterns have become so ubiquitous that a new developer might just think that that’s the way to implement certain features.

Sure, but everyone knew they were doomed from the beginning, no? I imagine that many of the very thin wrapper products exist solely to make some profit while ChatGPT (and others) don’t offer that functionality out of the box, and not to evolve into a sustainable business over the next decade.

Dumb question but I thought tauri was the lightweight alternative to electron. Did I remember that incorrectly?

e/acc is short for effective accelerationism and not effective altruism. Those groups are interested in very different things.

You don't have to guess. It was announced at meta connect that that is the plan.

Afaik Ship of Theseus is replacing parts incrementally until eventually no part is from the original ship. This is more like taking the ship, building an exact copy of it and then throwing away the original ship. Not sure that thought experiment is applicable.

I've never understood this. Even if we would have perfect mind-uploading capabilities, this doesn't help the billionaire who is afraid of death, right? It'd just be a copy of them. An immortal one, sure, but the original person would die just the same, no?

I can only comment on german. For general stuff, like articles, blog posts, etc. it performans really well. For local stuff, like finding restaurants, opening hours, etc. it's okay-ish, but switching to google with the bang commands is easy enough to not annoy me. It is the weakest part of the product though.

Not out of business, but if the 5x rumored compute is true than they’d need to do something phenomenally bad to end up with a model worse than gpt-4.

You seem to be very sure of this. I'm curious where that confidence is coming from.