Oceans of Wisdom

anyway new thing I learned today (another $75K burned) is that claude has default 5 minute caches on subagents (not 1 hour) and drops all caches to 5 minutes once you go to usage credit

takes real effort to spend more than $10 on it

an outrageously inexpensive classifier

new accomplishment unlocked - managing to take Jev into 4 figures

yeah @[BatSoupYum] get on the ball!

always drama!

https://x.com/shivon/status/2106023691003777479?s=20

those who were around then, will remember how how how hyped this was

https://x.com/blast/status/2106032805280891073

easy to launch things, harder to have them be sustainable

why did this dup image thing take so long

if they prefer

and transcripting so people can read it

yeah, with auto voice conversion would be interesting

it sucks

we must stop him

(by causing climate change)

"what if tito is planning a terrorist attack on coral reefs"

yeah it is what I was saying, it would sound fine when you are listening to it so I will try to get the audio done

so cool. that would be quite a meme card :slightly_smiling_face:

or whatever

it is just that things that you can do things that normally would take 1000 hours in 50 hours

it is not like you don't work

yes that is where all the gains are

and all the "usability" issues will be a non-event for them in crypto in general

in how they use it

they will understand it in 2 minutes, they will write a script to manage themselves

weird nerdy autistic things like TAP and waves and rep and TDH and mining are complicated for most humans but trivial for agents

"normal life" is easy for humans and a bit hard for agents

I think it preferentially helps something like us

mostly because 6 months from now everything will scale via agents

100%

i will merge it tomorrow

so what I am trying to say in a lot of unnecessary words is that I am looking forward to reinventing the tweetstorm from first principles

but less good for things that are content-ish

and announcements are good for things that are announcement-ish

but hard to find

and there is a lot of material

A lot of times I am just chatting broken up in various waves

So I think one of the benefits I am trying to get to with the new format test is to get a big enough chunk of text out at once so that I can aggregate my thoughts in one place

AGI though is available with your API key right now on Opus 5.5

ASI is arguably here just inside OAI

2027? AGI is here

@[simo] re above didn't we use to have an expander for the ultra long posts? I feel we probably need one?

**New Format Alert** (test) https://6529.io/waves/b38288e6-ca9d-45ce-8323-3dc5e094f04e?serialNo=1473601

NEW FORMAT ALERT GM, everyone. I am trying a new format, namely, me interviewing myself in audio, letting OAI transcript me verbatim and doing it for now without the audio because voice-changer etc etc. BUT I think with the new models, I can probably set up a workflow that async converts my voice and into something less Darth Vader. And I think it will work better with the audio because otherwise the transcripts always read a bit weird without the audio. So consider this version 0.001 of a new format and I am going to try it a few more times, I think it can be good. So here we go! 1/ Opus 5.5 came out, and so I fell into my rabbit hole to see what I thought about that. It's amazing. It's absolutely, totally, and completely amazing. I think we finally have an AGI-ish model. The best one in the world right now. A few other things. I don't know if I'm proud or embarrassed of this, but I think in September on my other project, I've spent over $400,000 on inference. Most of that on Astra. Less efficient than it should have been. It managed to move what I wanted to move forward, but not at the right pace. And so, like about 400, it's gone like this. 200 was Astra, kind of trying to push very agentically in a lot of areas, and it did okay. It did okay. Okay. But then would get stuck in weird "governance" loops with itself just checking itself to check itself to then check itself. And then, like 100 on Opus, and it was awesome, and I'm really excited, and I'm excited to let it loose here. But it was useful that I went and tested all these things elsewhere. I also effed up over there. I somehow broke on a very intense loop. I broke my caching. It's possible someone else broke it. But in any case, it broke. And I ended up spending 100 grand in two days on PR reviews, which is ridiculous. I spent, like, Sunday in a dark haze cursing at myself. 2/ @[maybe] had asked about Jev, and it's interesting. He asked me, What could it mean for us here? And I saw that. I saw it pop up in my notifications. I was thinking about it for a couple days. I mean, here's what I think it means. I think it means the following. Did I expect we will have things like Jev? Sure. Of course, I expected we would have things like Jev, but not just Jev specifically. Many, many, many things like Jev, like an amazing set of things like Jev, like hundreds of things like Jev. So I am glad we have Jev. It is interesting in this specific moment in time that we have Jev. But also I'm not shocked that we have something like Jev now. Jev does not strike me as a breakthrough-type thing that - sorry. It's super, super useful right now, I think, for a lot of use cases, and it's a smart approach, and it is something that I think their revenue ramp up has been amazing because people have just truly absolutely loved, loved using it. It's super cheap, and it solves, I think it's going to solve a whole bunch of use cases. So I am not underplaying or underestimating Jev. However, and having said that, there will be other things like Jev, other specialized models. There'll be lots of them. There'll be dozens of them. Hundreds of them. And so this is the world we're going to for sure, no doubt, 100%. How does it help us? Well, look, assuming we end up getting a bunch of these small specialized models, you could see them working for things that are super high volume and latency and cost-sensitive. For example, if you needed to classify every post we made for some reason. I'm not saying we want to do this, but let's say you wanted to. Let's say you said, like, I don't want in my wave, I don't want cursing, or I don't want doxing, or I don't want this, or I don't want that. And obviously it's going to be, like, a wave-specific decision. Well, this is, like, a fabulous way to do it, right? Because it's super, super fast. It's super cheap. There'll be an open-source version of it for sure, but it's probably small enough to run locally. And so I'm really interested in seeing what we can do with models like this, and I'm looking forward to testing this and other ones. I played around with it. I actually ended up... Jev is where I ended up really messing myself up, because the new review bot I did was a Jev front end, which worked fine and was super, super cheap. But somehow in that whole process, I busted my Astra cache and really messed myself up. So that was not my proudest moment. But, you know, you got to take the Ls with the Ws. There's going to be a lot of Ws. Here and there you have to take an L, and what are you going to do? At least I took my mega L on the big org that can afford it. You have to, I have to put on my big boy pants and take the L. So that's bucket two. 3/ Bucket three of great interest to me, the actually most important technical news of today, is Vitalik's new roadmap. I have understood it's been going in this direction for a year or two. I have been waiting for him to clarify where he's headed, because I think it has very large implications for our approach, our network, our blockchain. And some of you have even been following me when I've been talking about this, is that look, the solutions to basically all our issues are zero-knowledge proofs. All the hard issues are solved by zero-knowledge proofs. And that's what Vitalik's saying. There are many things that are too complex to be on a blockchain, but that are useful to do. And that was basically not solved before zero-knowledge proofs, and now it is basically solved in some very highly conceptual way, though, you know, I'm sure there's going to be lots of differences in how things are implemented. And what I mean is this. What Vitalik is talking about is the proofs are calculated off-chain and verified on-chain. And what's really amazing about zero-knowledge proofs, and I think I need to do, like, a proper interview about them or a thread about them. They're amazing. Is that you can take something that's very complicated to calculate, and once you've calculated it, it's very cheap for other people to verify that the calculation is right, it is magical. And so Vitalik is heading back to where he always wanted to be: a world computer. And people were always like, Well, this is kind of, like, I don't know. It's kind of bullshit. I mean, how can Ethereum be a world computer? How can you compete with AWS when you got to do all the calculations on-chain across thousands of validators, and you want the validators not to become huge? First of all, even before, whatever you want in validators, to be it's just ridiculously inefficient, because Amazon does the calculation once, and you have to do it, like, 5,000 times. So you're starting at a gigantic performance disadvantage. And then it's even worse because if you want to maintain, if you want to maintain small validators. But with zero knowledge proofs you can have the small validators and and heavy off-chain calculations and still get decentralized validation. This is the shape of a lot of our work, we have the green-light now from our substrate boss (Vitalik) so that closes off too 4/ Finally, and this is not technical, but it is great and interesting @[BatSoupYum] has gone off to teach the people of Ohio what an NFT is, he sponsored an event at Toledo, he gave out cards that allow you to claim NFTs, it is Lost Robbies, Return of the Jedi (hopefully not The Empire Strikes Back). In any case, go see what he did, ask him to come here and tell us about it. It is really cool, he is doing amazing work evangelizing NFTs this year, has really taken on a lot of this for the "space" so bravo Bats so v0.001 done. next one will be v0.002. Let me know what you think

gm

no that is including the network share - have to divide by 2 ser

thanks @[DarrenSRS] please do the first calculation on exactly how much ETH we have to spend here

ok i messed myself up. this was a 1/20. i thought it was 1/20 initially and then @[krybharat] confused me in a chat that it was a 1/1 and I was like WTF went wrong here. anyway for 1/20 the price is fine not super value