#191 - Sora leak, Pixtral Large, OpenAI email archives
01:42:11 Share

Step into the title dream, where pixel come alive, AI paint species in real time washed the future the code in.

Hello, and welcome to the latest episode of last week, eye, where couldn't hear us chat about what's going on with the eye. And as usual in the sub sob will be summarizing and discussing some of last week's most interesting AI. I'm one of your regular host and the current my back on is that I studied the A I in grad school and I now work as gentle A I started up and once again jeremy is out uh, due to thankful for giving and presumedly also more travels and more work on his side so we wants to go have a guest cohoes and IT is once again, uh, govern newsome and yeah.

I made IT. I've made IT yes, i'm the government of california and my me is my .

name is Gavin percent but I see morning I missed i've been thinking too much about california clearly .

it's alright. I haven. I get confused a lot. It's of those things. But yes, hi everybody, money again.

So i've been here before, but I am the coast of a pocket called A I for humans. What you can find on your audio or on youtube. And my background is media.

I was a show runner for the tonight show for a while and then fell into the AI world and involved in a bunch of weird started up stuff, but mostly like have been kind of following the space and specifically from the media angle. And there's a lot of faster media things that are happening in the air world, you know not just beneath leics. There's a lot more.

There is is quite a lot yeah and I do enjoy following IT. I'm glad i'm for one who .

added to spcas leave .

that and I love yeah really well, that is a fun, a little brain fart. But yes, it's it's exciting to have you on this week, as you'll see, is very heavy on tools and applications as well as new. There is and know why, but there's just a tone of stories on that front and on your pocket of that is sort of your real house is sort of A I that people can use in their daily lives or for fun. And yeah, we really talking a lot about that in this episode. So yeah, we really specialized .

and trying to make IT a little bit more accessible for I wouldn't say like the full norm audience because the omi audience is like maybe a few steps time, but for the super curious kind of like non non technical techy person. But we do find techy people like us as well, mostly because capable I say dumb things.

So that's part, that's part of what happens. Yeah and you have a pretty fun show. I feel like you have some demo. Sometimes you have your A I cohoes, which I found delightful. So yeah, is is a fun time and before we get into news, as usual, I wanna give a shout out to some no comments and reviews, a couple new reviews here must listen for your news sharing about helpful things and skills to learn as an A I engineer. So that's always nice to see and we do have A A detailed one that I really liked.

Uh it's titled the index ensor A I podcast turns out from a tex CEO and I start up I wonder if I have met you in the bay area uh but I have read your uh review uh here's there's actually one request which is to keep covering the state of meta and german I and how they compare, which I think is a good one. We I haven't touched on german ize quality. And in a while, uh just to touch on IT a little bit, I will say that I try germany every now again.

And I always come away disappointed and kind of confused as to why it's lagging, at least for my applications that are more coating heavy, a little typically more like long context documents and sea. I haven't found IT to be as good. I don't know what your experiences honestly.

I feel the exact same. And and I the part of me thinks that somebody at google should be held directly responsible for what's going on with gemini. And I don't want to crap on them because I I know that I logan cul Patrick or carpet ric is very active and trying to like promoted in they come out with stuff that's kind of edge on on the on the boards.

But like what I mean, that is like the performing well. But every time I use IT, i'm pretty disappointed. I use german.

I live recently they just release in IOS APP. And it's just the divide between where open uh, opening eyes of advances voice is and that is so wide. And I don't understand how a company that as biggest google could be fumbling IT this much. But I know I feel bad for saying that some ways, but IT does feel like every time I open IT, i'm not getting the results out of IT that I want yeah.

it's kind of weird to me and and maybe surprising because they do have a great team of deep mind, have a lot of hardware at sea, but know the gama isn't bad per say it's a powerful chatbot. May be equivalent to meta lama seventy b or four hundred that I have been tried but just compared to wealth on pic and open the eye, it's kind of worse every time, which is, I do want to shout .

out one thing. Speaking of google, I know that this is like a random thing, but there's a guy, josh water, who is the VP of labs dog, google at google, who posted a thing where basically, and this is, by the way, this was super fun, and I wish they would do more of this. But it's thing called je chess.

Have you heard about this at all? You know, je chess basically is a small dum little thing they created, which allows you to create chess pieces. I play a chess game out of, out of general prompt chest.

So I created, uh, uh, hot dogs said versus a suu, he said, and each little pon is a hot dog and it's a very fun like little toy that I was like, now this is cool. This felt like something, of course, that doesn't matter in them in the largest cheam of things. You know, no one is going to use je just like improve their business.

But like that felt like old. It's like a new book. I am like he was like, maybe that's what google is gonna good at is this we're little spin offs, unfortunately, that those are not to be businesses which might be a big problem .

for google going forward. Yeah I I think to say one more thing on this topic um IT is probably the case that germany is good enough that the integration of IT into various google products will be very useful. Lc, some mary out to complete the uh summarizing of your like voice notes, meeting notes eeta google does benefit a lot from having their own L M to use box google .

dogs at a big part of that right in google dox. You can use IT to write or complete .

sense and things like that. Yeah, exactly. So using g by that I could see being very useful. But as an alternative to ChatGPT or claude IT is just not as good.

IT doesn't work, I mean, I sure, and that he doesn't work. But I ve just never go back to IT.

I never go back to IT now. No, exactly. The other thing, if you has, which is pretty beneficial, is a little kind of guide on how to compare different image generators.

I don't know govern if you have listened in every time we cover a new image generator on the podcast lately. We are like, I can tell if this is Better or not because they all look really, really good now. And so this review goes into a kind of view. Metro c now is not so much image quality but prompt hearings and their ability to do exactly what you want, which is true. And on that front, it's it's kind of still hard to tell, but there's definite a difference between different ones.

Know the other thing I would say about image generation that makes a difference to me, weirdly, is ux now. And I know that sounds crazy. Most people like I doesn what the U.

X. Is because you are. But truthfully, that's how people are going to use IT. And I think that idiot gram surprisingly is is, I think, now becoming my favorite image generator.

My journey, i've been paying for an a problem gna cancel my subscription soon because mijares Y U X is Better. Now that is a desktop up. But I still feels like i'm not getting the results I won out of IT.

Where is the video? gram? Um it's pretty good. And influx, obviously people are so many people are building.

U X is for flax, i'm sure will talk about some of the stuff today and dream machines, new open thing but like flux is amazing and prompt rehearing is really good. But like I had yet to have like the perfect U. X experience with IT.

But I think that's why U. X. We lay is like maybe the future of this stuff is the same thing without allam. s. Is like as things get Better and faster and more same across the board, it's gona be how do people implement IT more than IT is like what the actual thing is doing that makes sense?

Yeah, exactly. It's going to be the case that you know less about technology and more about how useful IT is to you in practice.

And animals like the steel beams of a building are still steel. It's really what the building around IT looks like and really want to be in the right that the vibe that i'm get eventually, it's like the pipes work. The pipes work and deliver the water and what I drinking out of that, right?

So yes, thank you for that detailed review over a quest and note that they are useful to us and do for free to send our comments. Long things. It's always nice to read.

And just one last thing before we get to discussing the news, we do have our sponsor to promote as we have recently. IT is the generator, which is the A A I lab focus on entrepreneur A I from boston college. Boston is a major school for entrepreneur ship in the U.

S. Ranked very highly. And last fall, professors from box on partner that with students to launch this generator lab that has various focuses like A I entrepreneurship and business and A I F X and society, AI arts and performance and more stuff like that.

So they do work across fees kind of categories. They peer train the faculty in A I concepts and A I tools. And they love a podcast actually. So that's why they reached out and had a promote them as a sponsor.

Most recently, they had a generate A I and writing symposium they hosted and they had different representatives from major universities like covered m of eastern and tough. So kind of that gives you a taste of what they're up to. It's pretty interesting initiative, and i'll be curious to see what comes out of IT like that.

already. And finally, getting to the news, starting with tools and apps. And the first story is fresh of press. We according a little later, unusual is the wednesday.

And I think just yesterday, this was announced or or kind of covered that OpenAI sora video generator appears to have leaked. So the way IT has leaked, unfortunately, I was excited first because I was like other weight open. But the way has happened is that there was an interface on hugging face that you can use to generate ten second videos.

And IT seems to be a case that there was someone who had access to the API, who could submit request to generate videos, and they were unhappy with some of the things that opening he is doing. So there was this group that calls themselves sora P. R. Puppets that has claimed that opening up pressures testers to promote saw a positively and criticize IT as P R focused and not really are just friendly so seems to be the case but someone in that kind of vein that has done this and as a result uh access to IT was shut down pretty rapidly um but yet again and exciting a story of drama arising from opening eye yeah I mean.

there's a couple things to talk about here. One is what the outputs look like and then two is kind of like my larger conversation, I think, around what sort means to OpenAI t hink l et's s tart w ith t he w ithout o utputs o f l ike i s l ike t hey l ook p retty g ood l ike I t hink e verybody I h ad h eard u h, from the back and that like the sora was like not as good as like what we had been seen with run wage and three or with luma or with minimum, specifically the chinese models.

I think there's a lot to like here about the outputs that you see. And you know there's a lot of really cool things that came out of IT. There is some like, you know, funny ss to IT.

And just so everybody knows from what i've heard, this is a new sora model is called sora turbo right, which allows people to mix up faster because the the world of the story also is that when sora first came out in our set access to IT, a one mini video took like ten hours to to to generate. I don't know that's exactly the case, but like some crazy at a time to make IT. So this is a turbo model.

And I think all like i'm pretty impressed by that I would love. Obviously, this is you know I ve never had accession. We've interviewed somebody on our show, one of the artist s had early access.

They had an interesting experience with IT. And i've actually talk to the people at opening. I had dinner the other night where I was across the table from somebody um a woman who works with craters opening eyes so like I know that they do care deeply about creators there.

So I think that's one thing that's a little tRicky here. Um I think the the bigger question to me is what is sort IT means to open an eye because here's what i'll say is somebody entertain business. AI video is a pretty big headache when IT comes to uh, both legal ramifications and sorts of like partnerships that might might come out of this.

Now opening eyes, as most people know, as already facing lawsuits by the neuro times and all these other people for the training of their data, that is, text, I think sora opens up a whole, another large set of problems for them. And my god tells me the way they have been pitching, sorry, is not as a consumer product, persae as much as IT is a professional product. So they don't have to deal with some of those headaches.

So what's interesting here is this idea that you might have just been like the people that control sorrow within, open a eye. And by the way, open a eyes are very large company now it's much bigger than I was our first premier. And there might be like some people who are who that kind of iron fisting, what people can say, what they can say about IT that never goes well for companies here.

I mean, like when you you never goes well for a company to say like you can only say good things like you know, if you watch the history of what markets Brownly has two companies across youtube, you know that that's the case. I think what's going to be interesting here is I would just like them to come out with IT and say here sa and and I think you let us play with IT. Obviously, I think the thing that I they might be worried a little bit about IT is that IT may not compare as favorably to the very top level of the of the chinese models particularly, but we don't know that yet.

You know, the results some of the videos I saw were amazing. Like what IT feels like in a lot of ways is like sora is much Better at the kind of like stock footage prompts that than than even to or anything else is like clearly it's got of like a lot of stock footage background. I love the artist videos that have come out through the opening eye channel like the official one.

There's an amazing artist name. Nice on that makes these incredible surrealistic AI videos about her. You know, this kind of like korean grammar kind of vibe, that stuff is so cool. But I don't think anybody really knows until they released IT. I I am trying to figure out and wonder, what do you think is that this onna make IT less likely to come out? Or is this more likely now that they're gonna drop IT?

Yeah that's a good question. And I realize I don't know actually mention what sorry is. So hopefully all listener is already know sorry, is the video generator from opening at the first, uh, showcase? Think at the begin of this year yeah and at the time that blue everyone's mind because I was by far Better than anything you're seen in video generation.

As you ve said since then, you seen a lot more come out at the time. There was really no publicly accessible way to generate videos. Now you have things like luma, things like runway of that.

There are pretty good job. Uh, and especially you do IT like in a usable way where it's fast enough that you can actually use IT for stuff. And in that time, opening eye has been pretty quiet on sora.

They have no gone to some people and how you would have done these like private partnerships, but they haven't showcase many new results and they haven't kind of talked about that much in general. And I do think there's a reason for that. A one of the reasons, as you said, is this whole a legal angle of an value are in trouble for using some data, sorry, is definitely trained .

on corp data has to be right? Yeah exactly.

So I doubt want that. And to me, every angle on that front is that I just don't see much of a benefit for them to release IT. exactly. Yes, because we have GPT that's printing money, that's their business. yeah.

I mean, like that's their business and and source expensive, I bet, to generate, right? So like what do they do? They really need to another giant like computer right now. Truth, they do.

you know. yeah. And everything is like, unlike image generation and ChatGPT, video generation is one of these things that to most people, just be a novelty unless you are creative, professional and you can use IT for something like um maybe as a youtube as can make someone who works. You are more professional product and you want to to get cheap B O of a specific .

real you tell me that making baby joe rogan talk to .

baby turns Howard is a professional job. But no, that's what me. People can use IT and paid for these things.

But that's a relatively small. No, very small. You're .

absolutely very small distance. exactly. So yeah open the eye. I think, uh the upside of productive zing IT is a small yes and they don't need to like show off anymore that much but they are a leader yeah I so yeah but in some ways .

that almost the opposite, like they came out and IT was kind of underwhelming. IT might like corner them a little bit right. Like I don't me like right now, as we all know, I A is a big hype game. And part of opening eyes hype is that they are the cutting edge. They have some version of a mote, even though it's real or not, we don't know, obviously, hopefully talk a little bit about deep seek and all the old one support.

You may be already talked about that, but like the idea that like if opening I comes out and it's fine and the compares to like you know okay with some of the more cutting edge models that start not great for their company and if they don't need IT, maybe this kind of thing they do, right? And so maybe this leaking is weird in that way because they weren't planning on this. So that could be in .

a thing as well too, right? And then going back to a question of what does speak mean of a more likely or less likely to make IT more public, it's hard to say. But my guess would be is that this would you prompt to be even more cautious of sorrow than they already are? Um you know this approach of handing out access to certain people has now backfired. And my guess is, well, keep doing IT and and be even more strict or even more careful.

The one I say about this is that you read the artist statement on the hugging face page. It's interesting because you know it's a conflicted thing with these artists who want to use eye tools, but then they're often trained on copyright at work or work.

And I think even in that statement, you can read into some of that, right? Like, I think that there's this gonna be this ongoing conflict in artist minds about, like, do we trust these corporations? And like what the messaging of this is, which is like, why? Like, open source tools are really interesting.

And I know where a long way away from IT, there's there's a new open source video more out there, I think called L, T X, which is interesting, but it's not amazing. I do think eventually, like open sourcing video and audio tools is going to lead to probably interesting stuff. But right now, we're so kind of under the film of these giant .

corporations yeah definitely on that statement on hugging face is pretty fun to just look at that yeah you should read sure yeah it's fine.

But just of IT is that these appear to be some uh, that have received access and instead of being testers and creative partners, they say that they're being lured into art washing to tell the world for for artists and kind of like you know saying that is good and not criticizing IT, which we call corporate art washing and and that's the motivation here oh, already moving on of the next story is about mistle and some updates from them. So first. They have now extra large and update to wear a muli model, a model that was first released last year.

So it's a Better one, able to sort process images and and deal of image inputs. And then secondly, I think the bigger deal is there has been a large upgrade to a leh chat, which is their platform for using their alliance, very much like of a charge B T web interface or a cod web interface. And so there's a lot of updates to this web interface.

We have added image generation, web search and interactive canvas. Uh, they have document and image nosis and they say even task agents. So they seemingly in one fell swap added kind of all of the stuff of that ChatGPT and claude have been adding gradually over the past year stuff like canvas, which of course now is both on, uh, tragic t and claude, this thing that is a nice to interface for working of documents and editing element so on.

You get all excited, the output instead of being IT in mine. And then of course, web search is for your question, IT with such urban productions and IT really does seem like we're trying to frame IT as being competitive with charge P, T, which I IT wasn't feeling earlier to me. I felt earlier there are more, uh, competing for the A P, I. Access for people to pay for the L M. But of this movie seem like they might want those kinds of users, which is interesting.

I think that this is the great alcohol, the great reframing of A I companies. When IT comes to this, right, what I think is going on in by business person per say. I've been around business quite a bit and think about product a quite a bit through my career.

I think everybody is trying to kind of jacky now for where they're gonna sit in the kind of stack of where the money is made, right? And I think that's an important thing. Like if if anthropic and open eye become like the kind of A P I tubes, like in some ways, like in met as obviously I going to make their own version of like lama sort of apps as well.

Mixture al kind of is in a weird place, right? Like it's is amazing. Like you know, open source having is open source right is designed to be open source in source company based in europe.

So it's a slide different thing, but maybe that had a disadvantage for them based on how the european M A I rules are. But I think they all have to start differently themselves a little bit because there's a real argument. I'm true.

Cover this on the pod, but ChatGPT bought that your chat dot com as there you are. All right, because ChatGPT is very quickly becoming the kind of clinics of of A I right words like everything might go through an eye for the mainstream. So mixture and all these other companies have to start, I kind of differentiating themselves in some ways.

So like, I wonder if this is their first step towards that to try to be like a lij at. So it's liat is like maybe A A an alternative version to what ChatGPT is trying to become. Where's anthropic is trying.

I feel like to be something slightly different. Like anthropic keeps pushing forward these kind of interesting, you know, computer tools and all these things that feel like they're gonna big for developers in a significant way. I don't know this is a such interesting thing to me, but like we're getting to the point, but consolidations probably gonna happen and a company like mixture could conceivably be on the outside of that.

I think yeah they they have a real chAllenge with respect to providing an A P. I was competitive with open eyes because it's a little bit of raised to a at this point. Yeah exactly.

Open source tools, right?

exactly. We've open source tools with API tools in general, right opening eye, google and on topic are all able to provide free, cheap, probably unprofitable rate is at this point, which yeah is very hard to compete with.

So you're write maybe one angle they could go is for more of this consumer and of everyday a usage and and that goes back to a point you made, which is one important aspect of the chat box that wasn't the case a year ago is the user experience with stuff and uh I guess image generation being able to uh do web search. All of this is kind of development in of the user experience of using elements as a child bot over the past year. So there is space to compete there. Bvs distance, they are these positioning that short as competitive.

Yeah my thing is I I look through web two o old enough to be there and be part of IT. And I think the interesting thing about what web to o was, as I felt like I was layer on the the Better use case scenarios on to what the internet was right, like the internet kind of came up in the nineties. And then web two o kind of brought forth interesting like dynamic U X in U I. And suddenly like you could have things were like dig or you could update at things and people can share things in IT feels like that the transition that were coming into in the air space a little bit is like where do we go from? Like the raw information being put places to now it's gonna get a little bit nicer and Better use .

cases and move on to the lightning round where will try to go faster. We have like ten more stories in the set of ah yeah yeah yeah. So first we have a some new stuff for microsoft.

They had their ign twenty twenty four uh presentation recently and they announced and the major thing was the notion of A I agents. So they're adding these things statical agents, which aren't agents in the same sense of we've ve talked about before. So some examples here are you have an interpreter agent that is a real time translation tool for teams meetings.

A facilitator or agent to that is basically a note taker. And summarizing for a teams meetings, they have one for share point, uh, that deals with dt data in an oasis. Some employees sell service stuff, so they are calling these agents.

These feel like just more features of these products. They rented agents in the sense of we've been talking a lot about, which is, you know, you have A I thing that you give some instructions and IT goes off and duster for you. These are more like togas features you can use throughout microsoft.

You know what this is about? This is about mark, a statement that said that like soft AI doesn't matter that the futures, all that agents so such an adella says, hey, i've got to start saying the word agents but I do think like A A entia I I feel like is going to be probably the story of twenty twenty five and an sam altman and dara tropic have said as much, but like to me, this is the thing to potential to we discovered this on our episode, came out today, which was like the three biggest trans of twenty twenty five. And I think that, you know, true agents are going to come and I think getting people used to the idea computers doing stuff is important. But I agree, these are kind of like more features or anything else.

right? So like technically, uh, trans, a real time translator is an agent in the sense that IT, like is there with you doing stuff. So you can call IT that it's not entirely wrong, but it's just different from how has been used typically.

And speaking of agents, we are gona talk about a start up now age, which is based in paris. We raised two hundred twenty million earlier this year, and they have now launched their first product called runner age, which is meant to be used for agents applications. So they have their own uh, compact LLM of two billion parameters and they have now a API for both prevailed and custom agents applications. So they are guess arguing that of this would be useful for things like e commerce and banking where you can apply um you know different h sort of business processes like dealing reforms and temples, squad assurance stuff like that. So yeah interesting again, an example there where it's not trying to compete on elm access exactly not intend ced a new chatbot introducing something that is more tailor to businesses, which makes sense.

This this is all gona happen. And I think it's interesting, like I will say, I didn't with a very large person, a very big person in the in the investment flash A I space a couple weeks ago um sensibly like one of the biggest and he was saying the agents companies that are coming out is just become very hard to invest in because there are endless amounts of them. I think these are always interesting stories.

I think a little bit that might be raised to the bottom as well with these things like I I worry for these companies that like why can't OpenAI ye anthropic introduced themselves in some form and then suddenly the entire company is gone. Now granted, maybe that these companies are working with these A P I. So maybe there are U X, as we talked about before, we will be Better than those companies. But I just see it's tRicky to imagine a ton of these succeeding um in the best ones will be the ones that .

just make IT easy I feel like yeah so yeah it's a interesting situation where in this example IT seems to be with this rebuild agents we are now targeting ing some more domain specificity like just bought cell lambs are all general purpose.

Maybe part of answer is you tailor your product to uh specific thing although I also think they want to be like knowing four of the best agencia I and restaurant yeah and moving on with more stories about l lambs next we have on tropic and they introduced uh pretty new, new feature and cloud that is styles. So now when you use IT in the web interface, you can choose how you wanted to respond. IT can respond and formal way, concise explantatory or a Normal, which that figures like a mix of the above.

So, uh, yeah, it's A A tweak to the user experience that I haven't seen. I don't have tragic t or hours. Have a makes lot of sense. And guess i'll be playing around with IT and seeing which I prefer.

Yeah, I make even play with this all either. But I think the style idea, so the big thing, the thing I talked about her show that I really excited about at some point, as soon as I love using ChatGPT. But what I want our instances, which means the idea that like, I want a ChatGPT that I can fork off and so control and have like the ChatGPT, that's my business ChatGPT or the chat P T, that's like my personal one, or the ChatGPT that I can talk to myself at age twenty, right? And like I want to list to those on the left and they are not really GPT in the way that they gini envisions.

Then there are more like personas, right? And this is the kind of step towards that that I think A I is going to take us even when you think about agents like I don't want an agent that like, you know, if I want an agent that's gna go out and research something for, I don't know my kids, I don't want that to be like business guy. I want that to be like an emotionally charged person who isn't understanding what I want for my kids. So I think that's gonna a weird future where our own a is are gonna fractured into a vigilant different thing based on what we want. And I think that's for like the agenticity havisham, the chat behaviors kind of come to to each other, like when you think about agents, not just as being something that can go out and do something for you, but agents is being something that are a part of you that feels like kind of the future that we're looking at here.

exactly. I think this really speaks to sort of the other meaning of asian to one we tend to focus on is the reason, and they can go off and do so on their own plan. But we ever kind of fact of IT is this notion of having a persona, of having a persistent identity exactly.

And that's another aspect of agents that is definitely on the rise. And actually, speaking of that next story, very much related eleven labs s is now offering their ability to build conversational AI agents. So eleven labs s is the service that h typically does the text to audio, text to speech of their, the leader in that area.

And now users can build that is a conversational AI agents uh and you can customize, return a voice and response like so kind of similar to, I guess um yeah we chat P T uh chat experience with advanced voice. Small rate ah this presumably is the similar use experience of being able to talk in a very sort of life, uh, real time way to an A, I will have a voice. yeah. So we looked into this little bit .

and where big fans of level labs on our show, we've been using IT for a year and a half, mousse ce, we started really um uh I think this is going to be the future of many things.

In fact, Kevin, I are working on a thing i'll shout out to the studies because there might be probably developers this thing if you are developer who is interested in working in A I voice, we are looking to connect with somebody who is interested and kind of putting the pipe together of this stuff, because we have an idea for IT. The thing to me, that kind of the reason and thesis that I think this is a big deal, is I think IT is underserved. How much voice is going to be the U. X.

Of the future? Meaning that I think people are gonna start getting used to talk to these A I S. And in the same way that I used to feel weird when you would see somebody talking on their phone in public, and now no one gives a crap because everybody sees at all the time on their headphones.

I think that's what talking to A I is gonna feel like. Pretty soon it's just gonna be like a first can be like, are you talking to that thing? And we'll be like, yes, I am. And then five years, literally everybody will be talking to them. So I think that is is just all of this is kind of aiming towards future where yes will still type onto our computers. He s will still type into our phones but based on the where everything is going, including even like you know visual A R stuff, I really think voice is the interaction uh element of of the future and I think this is like everybody moving towards that in some way .

or another exactly. And uh yeah I think voice like a lot of things intersecting here in the sense where you'll have agents in the sense of persistent AI programs that have an identity and memory and sort of like are your personal assistant but know you and know what you want and you'll be able to interact with those agents both via text as you would usually like you, you know chat via coworkers of a slack or you can talk to them. I think that is very likely the future. As you say.

it's very simple. I like, right, like the one of my favorite type I series as the altered carbon series, which just only a couple books in. But getting Richard came over again, like one of the things in his book as he talks about, like, a hotel has its own A I that runs IT, right? And like if that sort of interaction where IT has a personality, but it's there a kind of interact with you feels like it's a weird thing.

is to production famously referenced her the movie, of course, very much yeah, if you watch a movie.

IT seems going and movie does not in well, by the way, really for the .

person we yeah and once again, smoothly transitioning to a related topic of the next one is about germany and is now having a little mary feature. So the german child box will now remember personal information preferences and context to enhance user interactions. A seems like you the same sort of thing asked tragedy ity memory, which has been there and for how long but quite a while now. So now users can instruct gina to remember a specific preferences um and you know be able to have games be more teller to them yeah I mean, again.

it's like, okay, geri, your shipping products that I have been around and these other things for a bit, I guess just give me more jan chest, right? Like let's get a shift that's to shift up. Give me more gunch's, but it's cool.

I mean, again, your point earlier about the idea that google products are all deeply integrated with this. Now that's probably where this makes the most sense. Like one of things I love about ChatGPT right now, as IT remembers, when I want to write something about for eye, for humans, for the show, and I will give me that kind of like a sense of IT.

So I can imagine if I in a google dock and I say, hey, i'm writing an email to my boss, I need you to remember the tone that I used for that. That would be really cool. And I think that's probably where I will show up in the most interesting way.

definitely. I think, uh, be able to tailor how IT does email other completion is a pretty big know. Sometimes you need to be conscious and not have all for now.

Yes, exactly. Moving right along. The next step is complexity. And they are introducing a shopping feature for pro users in the U. S.

Perplexity once again is an A I powered search engine, kind of a big game in town when to A I search and now they are going to be able to uh, see, shopping, accommodate and place orders directly within search results, which is somebody you can also do, I think, on google. So very interesting to me. The shopping tool is integrated with seller sites like shoppy and says that VISA on by the accumulations without sponsor slots. Uh and um yeah could easily save as being a motorized uh route for complexity, which I do think is pretty important because subscriptions are probably not gonna cut IT.

I have a lot thoughts on this. The most important one I think, really is, is that yes, I think perplexity is one of these companies that either gonna a flame out quickly or kind of take off. And I think this is probably where they see their reviews stream coming from because as we know, like the big unlock for google, was google unlock ads right in like IT specifically like very targeted ads for people, a large part of which is shopping.

Now, not all of IT, right? A lot of other services and other things that are you can targeted to specific people if perplex, I can make this work if there are if there is real value being delivered in this way to the people that they're buying from, they actually have a business, a real business. Now the question I come back to and this is a you know perplexity is very famous for kind of controlling how many things come back and you know what those links are.

So is there going to be a wall set up so that like you only get a couple things through? And what does that mean for the rest of competitors in these businesses? Two, when you think of algorithmic results in the past, the internet as I um I guess we're on this part as I don't like canada, right?

Yes, good. okay. I think we think of our golts.

we think of a algorithmic results of what the internet was. It's kind of become a ship show, right? Like there's a sense that like people figured out the seo that kind of bubble up those results.

So my worry with a perplexity shopping APP in general is if i'm getting just three results to choose from or maybe even just one because the dream of the A I world does that delivers knew what I want. How do I really know that that's the best one? I don't know.

Like, you know, the wire cutters made a really great business in your time. Wire cutter, a really grew business out of like kind of figuring out here's the best thing and here's the most budget thing like that was really useful from a shopping perspective, will perplexity have the backbone in the information to be able to deliver that. That's my big question.

This is a business like this is a real business. They can figure that out. It's also a result space business. So if the results are bad, it's not a good .

business definite. Yeah, you're right. In research engine game, you always kind of trying to get ahead of people hacking research engine for own benefit. And yeah, one of our respective is to me is it's not very impair how A I can provide an advantage this kind of area like obviously for research purposes for in depth search, perplexity is villager because they have really nice stuff. But if you're just looking for noise cancelling headphones, I don't know that they can provide a much aside from like summarizing the articles that do compare them already, right? I'll see.

Here's a business that I think would be really interesting in the plex or whoever. I think this is somebody I would like bigger B, I agents instead of thinking about a place where they're gona, where you go to search on and bring forth that stuff.

I think the more interesting thing is what if I had an aig of myself that went out and look for stuff right and like that a slightly different business model is that like i've trained my own AI agent, like I control what IT knows and what IT wants. I personalized that and i've sent IT out to the internet to scow and then come back. That feels like a much Better business model for me as the consumer versus one company kind of controlling a pipeline through. And that that's the thing I think could be fascinating.

And I think we only have a few more um yeah for more stories here. Uh next is suno music generation and they now have suno v for so uh the music generation space has been moving pretty rapidly now with snow in audio. Both offering very good at text to song uh capabilities and very user experience has also been rapidly evolving.

So now we have a new feature called the remastered enhancing older tracks and also remi and A I powered lyrics assistant for egger lyrical content. And I will say I found lyrics generation of these tours to be vange an eri C2Be bor ing. So v four is, of course, now gonna lead to a Better at your generation. Ah you can use IT with existing tools like a covering, you reimagining existing audio and also personas with musical alderby goes, yeah one of these things where you do have some space still to improve the quality on the uh sanctifier ation as someone who uses the sound of is a lot so prety pretty often you will see weird I artifacts and like yeah sometimes you see go crazy and like you know you read the lyrics but vii is just like producing sounds that sounds like words but aren't um so i'm i'm taking v four. V five will start to see that go away.

I is a wife in a lot time now. We've talked about a ted on our show. I just resubscribe provi for my experience with that so far has been mostly good, not not perfect, I will say. And I know soon aware that there's been some issues with the force so far as I ruled out.

The thing you said, which is really interesting in me, i've had that experience probably like one out of every six generations, which I didn't have for three point five, was where, like you hear the audio in the words are literally just like a, it's like human words, but it's just garbled stuff. I feel like that the beginning of a of a new model and they're still find tuning and the releasing IT IT is from a in general, the audio fidelity is insanely Better. Like IT doesn't start IT started sound, not like you don't hear that tininess that you heard in the audio of such of the voices before that kind of gone.

I also agree with you, the lyrics generation is not great and I think if you really want to write your own songs and this, in fact, of funny thing, i'm working on A A dumb Christmas album of for a to release at some point about A I like Christmas albums just to have that experience. You really have to write them almost entirely on your own like even I even use chat p and chat chat p is OK at IT as well. Give your ideas.

Writing songs is not easy, right? It's not like it's not like an L M time to fully gets IT. Um so I think that if you're gonna use these things, you're best. But I am playing with the words yourself rather than trying to just kind of spit IT out entirely.

Yeah, yeah. I generate a song now per episode for for an ultra. And you know, I tried to just give IT to the topics and says, last way to have podcast generate physics. And IT just goes back with the same stuff over and over, like the same words, the same phrases. And IT gets so general borrower.

So it's it's definitely case that if you want to do something more interesting and personalized, you pretty much have to uh tailor IT a lot to make IT actually have good interesting lyrics. exactly. Next, moving on to image editing a generation.

We have flux point to one tools. So black force labs has been on a roll with flux. I think just last week, we talked about flux pro and raw editing.

Now they have this sweet of tools that has kind of usual things you would see in these kind of editors in painting and out painting. They also have a flux point on depth for structural guidance, with depth maps and with Kenny stuff for edge based guidance. And also, we have a tool to create variants and restless of images.

I love that. I love lucks. First of all, flux is amazing. The thing of that flux, people have, remember, is IT was a stable diffusion people that started IT.

And remember.

back in the first one, is still my favorite AI model, a video or sorry, image model all time, because you can do incredible things. Like back in the early days of civet, I if you were on severe eye and you were downloading like laureus from there, like there was some really amazing stuff. And you can tell that team that was disabled a fusion back and really was pushing forward and then the stable of fusion kind of blow up happen.

You know, i'm sure people on this forecast remember like the kind of the CEO left and also others have happened and those guys went black force lab and like IT is very cool to see a cutting edge frontier image model. And by the way, to post video model at some point coming forward to that is really pushing the the industry forward in a way. And I think that's great.

And the gain my journey feels a little crazy in me right now, and I feels like they are not looking as fast as they should. And another of you played around with my journeys tools, but they feel little funny. And sometimes the results are not that grade.

So I think this is disrupted in a significant way. I the interesting thing is I always wondered with these companies, what does the business look like on the back end like you know, the API for images and the API for video like is IT a business? And I don't know what IT is or not.

I almost think it's like I love they're doing this. I think it's incredible, but it's hard to figure out where the business is persue. That's my only question. And i'm sure somebody can tell me. I just don't know what IT is about. If you do if you do know, i'd love you to just at me on A A A for human show and just talk about IT because I would love to hear more about what the business on open source video and audio looks like.

right? Exactly a good correction. Just so it's clear, I may have made a sound like there's a web tool here and having a deepen, this is actually a series of models.

So it's in painting that induced all these things are individualized models that seems that are being added to v API. And IT does point to I think, for black forest labs, maybe we do have a pretty clear uh, business case. They also mention that these models are going to be available the of their partners to give her out a ei freebie. And now the cat users flux .

for imagination crock and a partnership if they are paid.

So yeah, IT seems to me they aren't actually targeting consumer space. They are targeting to be a provider for other tools that building a top of that which is resting next opening. I just gave a chargeback ity, a major creativity update so that the story is a new version of GPT for oh, which is said to have a Better creative writing capabilities. And if they surpassed goods germany in villa marina briefly, I think that vender was a new germany that went back to the top. So, uh, IT points to this interesting time in the vie chat boats, where they are kind of doing more tweak and more like little updates um to improve things more gradually though have big releases yeah I mean that to .

me is the best thing is like if they can just start drip dropping the updates, I think it's the best thing by far. Um it's hard to know how much proved like that's the thing. It's always tRicky with the open eyes updates like hey, that writes Better and i've tried IT and like I guess I don't know it's it's never been to me as somebody who's written for a living and done things and been creative.

Like it's always been OK and often times get stopped. And so I I wish there were a little more specific about like how they see IT doing. Like I want one of the videos be like this is how we see that but are not going to do that.

It's just more of a of an easy drop. I still think this is interesting to think like what ChatGPT has cooked up for, like the second university of ChatGPT. Like is there something coming because you know it's like you just never know what's going to be coming from those guys.

Like this was kind of out of the blue, so who knows? We will see what happens. But it's no, it's great. I love that they keep improving, which is nice.

And just one last story. We have one about runway and they have frames, which is a new a imagine generator. V A, has enhanced style list c control and of course, also nice looking images. So I guess the name of the kind of points to where we are going frames is like presumedly talking about frames of a movie, the individual shots you have.

So they highlighted ability to control and get very stylized images, things like one nine hundred and eighty uh makeup or uh miss on then like digital portraiture in the classic um individual language of classic cinema stuff like that. And again, as i've generally found, a runway, presumably of the idea here is that this would be something that people could use in practice rather than necessarily a general purpose imagery. Or this is something that in your workload, you may find more useful? I don't know. I think .

that's about right. And I think to be longs with you, somebody has done a lot of AI video generation. I think the turning point has come where text to image is fine, but it's really image to video is really where, sorry, text to video is fine, but really image to video is the thing that is going to be the bigger deal.

Because with image video, you can control consistency. You can make sure their characters the same. You can do all this stuff you need to do as an artist if you're making a film. And when you think about, like when a movie maker story's boards, uh, a film, what they're doing is you're putting the entire thing kind of in some version of images before before he. And so I think this makes perfect sense.

Any runway needed one of these because ultimately, my first step, if I making a runway video, is always, I go create an image somewhere else first, and then I drop IT into image, to video. So this makes perfect sense. I hope it's as good as some of the other ones. The prominent we've had the runway, sometimes that I know this is the kind of back and forth is all of these companies and larger companies are so controlling over like what you can generate in the end, sometimes to an extent of like.

You want to make out on this video about like a woman sexy dress spin around and it's not like it's anything no solicited but the word sexy is like too much for them, right? know? Or like things like that would feel like it's really hard to kind of find where that line is. But I find myself with runway often times not being able to generate something without any of reason. And and I think that there's gonna this kind of fight about like what you can and can generate is gonna be an interesting thing going forward.

Yeah and I guess once they have paying customers presumedly what they want to generate, uh, is got me one of a major factories. And I totally agree.

I think image to video and maybe even most of video to video, I think that's kind of underrated tex to videos is like where of all the hypos but if you want special effects, uh that's video in painting that is just saying, you know um you know remove any blurring or remove this car from the shot, remove a pedestrians that's the most useful thing I can think of and then you have image to video, which is if you want to can to get a specific visual um you get can really control things as you say, get exactly what you want and vent text to video as maybe like the food, your of these kind of h applications already Normal tools, stories and and we have a pretty short applications and a business section that will begin with a pretty fun little story. Ah so once again, we have some opening drama to cover at this time. It's all the drama going back to a while ago.

So in the ongoing a court case between elon mosque and some open the eye, there have been many males between elan, sam alt men, ellia sugar and greg brock men released going back to information of open eye. And then until up to when elon mass basically left and became a no longer part of open I in uh twenty eighteen or twenty seventeen h one of those. So you can actually read these emails, and it's pretty interesting to see how these people talk to each other.

You can read the like initial email from the sam altman to you a mask and may twenty three of twenty, uh, starts to have been thinking a lot about whether it's possible to stop, stop humanity from developing A I I think your answer is almost definitely not. And if it's going to happen anyway, IT seems like IT to be good for someone ever and google to do IT first, any thoughts on over to be good for Y, C, to start a manhattan project for ai then? H, A mosque response, probably worth a conversation.

Uh, sam altman has another long email that basically describes what OpenAIr w as a t t he b eginning A I l ab a nd t he o n m ask r esponse a gree t o a ll a t s ea a s o p retty i nteresting t o r ead t hrough. And then later on when you get to conversations that um were going on with opening I changing from and nonprofit to a four profit and potentially being emerged with tesla and making in on musk with CEO, that's where you get to some more drama with kind of back and force. And essentially greg brock and india are not agreeing to our plan that was being discussed that we on mosk basically haven't control A C E.

O. There was a very long email. We've got a rational and argument as elon mosque at some point h gets fed up.

Uh, he says in one email, this is very annoying. Please encourage, jump to go start a company. I've had enough.

Uh, guys, i've had enough. This is the final straw. Ever go do something on your own or continue of OpenAI as a nonprofit stuff like that.

Very a direct and somewhat, uh, confrontational language. So yeah, interesting to read. Nothing knew. I don't think as far as what we know, but I mean.

I will if I say that hearing sam, say the manhattan project, whatever ten, twelve years ago, when now an actual manhattan project being worked on for A I in amErica will be worked on soon as pretty interesting. Yes, I about this, I just had the book character limit, which is the story of kind of e loans takeover of twitter and x, which doesn't paint a great picture of elan.

But and I think elan a complicated character, and I think you can usually have to have to rest all the idea of the incredible things he's done, the incredible risks that he is taken with his own money to do these incredible things, plus a personal that I think is is fair to call personally I and and difficult, right. And I think what's interesting here is elan has made some of the biggest things in our world today possible, right? Like and that's a pretty big thing.

Like I had a test up for a while. I just returned the lease. But he has cars that are electric everywhere, and that has become a thing.

He now sense things to space, where the space race was kind of dead. This is elon kind of like kick starting this conversation with sam and all these people, and yet the the drama of elon to seeped into this. So I mean, it's just it's an interesting thing just to look back and to think like this is how IT all started. But also it's an insight into elan's personality, I think and kind of show some of the complications that exist within .

exactly and uh, IT does get somewhat nuance here and discussions of particular opening future back at twenty eighteen, uh, when IT seemed like google was poised to really be leader and in v emails you see kind of interestingly that google is dominating with of paper submissions.

This is actually from under ka athi who emailed in a mask and uh parents very talking about having A I C O and initial coin offering also to collect money that elon was opposed to, I guess he would need to loss of credibility。 A good idea. Mart, that's the book I want to read.

What I want to read is the book. And I, hopefully somebody he's writing right now, how very good google go wrong. And I know, and by the way, like that book, I think we would have some interesting stories to tell.

Now whether not you get the stories out of google, I don't know, but like somebody should write the story, which is like basically around this time, and maybe it's the rise of the AI age you want to call you, but like hearing the stories about internally at google. How did they let? No, I think no one was known Brown.

How did they let him leave to go to character and started character? I had not like, kind of was like, in whatever that guy's name, we ve got fire. Blake lemoine, who got fire for thing in the day I was alive, like all of that was happening during this time, that they should have crushed this. They should have absolutely crushed this. But like what happened behind the scenes that they didn't is is a really important thing for sure.

I'm sure someone is writing a book on the kind of the history of l lamps. And that basically is twenty seventeen up for now, right? Start transformers.

And well, there are language models before transformers, but large language models kind of are the story of the past a six, seven years. Yeah, yeah, yeah. Particular google, I keep kind of mentioning that will break the moon. People forget to now.

But they think he was talking about was this chatbot platform where google was letting their own the internal people, can I customize chat box and create individual chat bots via and an alarm? So they like, had IT and then a charge party came out half a year later is and that's IT right? Maybe in a different universe, maybe of their first to get all the hype and recognition from the mainstream who know.

by the way, you would make a good movie too because he can already see, like the version of the social network where you've got Larry and sergey on their yachts, like not entering their phones about this, I think and then six months later, they're like in the office, like pulling off their jackets, like we've got to get serious about this man, like it's like it's very, it's very dramatic. I feel like yep.

And now onto a few small stories. First up, we have a confirmation that amazon is investing another four billion in on topic, I believe have covered that this was an ongoing discussion a little bit earlier, but now IT is definitely the case. And so now amazon is coming up to a total of eight billion, uh, in anthropic.

So when you go on topic, looking to raise more money. So this presumedly is a nice when for them and um yeah still there is this kind of partnership or alliance between amazon topic that seems to be ongoing. And uh you know the topic is in a hard spot basically trying to keep compete with opening eye, trying to directly had on a beat on both the consumer front and the A P I front.

They're doing really good of the technology. Many would say I would say personally, I prefer close to tragi t, but they are not as well known. And IT is a real question as to whatever they can survive long enough to be kind of a second player or even the first player. Do you know how .

this could get fixed immediately? And i've been asking for this for six months, and I really don't understand why they haven't done IT make club the voice of alex a make IT just do IT and do IT. And this may be know your audience might be like a lexus, so stupid.

But on forever, I have five relaxes in my house that we have one in our bedroom. We have one in our itchen. We have one, my daughter as one in our bedroom in B, S, B.

Sorry, privacy, people apologize. Whatever I got, all I want is a room device that I can ask a random question too, so I don't have to pick up my phone. I don't to do anything.

You just say, hey, whatever. Can you give me this answer? Alexa is so dumb. IT is based on right now.

Often times, if you ask you a question, I will give you like a blog spot answer or some random thing from the internet poll. Integrate cloud directly into alexa, and you will immediately have a use case for everybody. And I don't understand why they're doing this.

I must be some sort of like legal worry that alex a that amazon has. But to me, IT just feels like IT is a IT is the dumbest thing not to do. We all have these devices sitting in our house that we talk to. Why are they not actively doing this?

I have no idea. Well, yeah, I just google this to have a fresher. And I guess you'll be happy to note that there wasn't announcement that almost on will in corporate clause in a service in a seemingly paid service. Therefore, I saw that. Yes.

I remember that and you're going to pay extra money for which again, okay, but what if I play for a claud subscription? Each month, which I do. Shouldn't I be able to just pt that over like IT does feel like and and it's like i'm going to use a crazy amount of tokens like i'm not gonna be over using and maybe if you want to put a cap on me, that's mine. IT just seems like a dumb thing that these giant companies exist and they can integrate this stuff like that feels like I just an a nightmare.

Yeah, you know, that's another book you could write. Maybe someone already has written IT. What happened with alex a as I think alex a was maybe the first one like maybe prety theory, IT was sort of what he was around .

the same time. He was around the same time. But yeah, even in elexa and again, I think a lot of people have them in their homes and what we use in four hours, setting timers and playing music, right, that's about all we do with them.

And annoying for a lamp control .

and have even more advance, right? Like the vast majority people have relaxes are using them for a lamp control because .

they know about the plugs and and speaking of amazon, the next story is about their robots. And there's report now that amazon's autonomous robots are seemingly struggling to keep up with human workers. So this is, uh, we have robots that not to being to be used sorting, loading and unloading package uh packages, and they're pretty good.

They have a lot of robots in their warehouses that drive stuffing and delivered around. There's a lot of robots that they have deployed with a last decade. Now they are trying to add of these robotic arms and and more kind of advanced I that seems to still suffer for an example of targeted picking.

So you need to go into kind of a clutter, the set of stuff, and then get a specific thing out of IT. Uh, still not quite ready to August. Replace human and work Better, which I think this was worth taking a note because this is something will be seeing over the next five, ten years, I think is will be excited about where a lambs and that will be kind of the general purpose.

A I progress. But then in the background, in warehouses, you know in trucks, uh, you'll see this automation start really impact of kind of low end of the job market. Yeah, I mean.

this was the fascinating I honestly, the robotics to me is going to be the future of the next probable ten years in that, yes, we'll still have these AI agents. All these things are early, important in terms of this.

But like when you have a robot is not a humanoid robots that can actively work as people, that's gonna be a big deal in terms of jobs in something, anything else, but more than anything else, like I think we're onna start getting used to having them in our homes. And that's another weird thing, right? Like it's like the idea in the same way that we're going going to get used to talk to these A I S. We're onna get used to some version of a robot in our house that can do are now we've ready have these roby vacuum. But like pretty soon there will be a robot that conceivably can kind of a walk around your house and then plug itself in at night and then you know, do a few things for you and then do a lot of things for you like that that's the world that we've moving into, which is pretty exciting.

right? yeah. And you know that's it's going to be one of these things that are maybe more .

disruptive than people.

Change is like amazing. And technologically in in terms of what you can do with IT is is like a new old era, really. But at the same time, automation on the physical front may be even more impacts. Al, early on, like you have one point eight million people working in warehouses, uh, in logistics and you know that's gonna start falling and we will see how fast that goes. But what about construction?

Like I was just in my brain. Think about construction. Like you can easily see construction jobs being overtaken by robot.

S because it's a precise IT has to be precise, right? So like, imagine you at first it'll be like the one or two robots I have to lift the. Y that I can lift the heavy things for the Normal people. Then suddenly you going to see whole teams of robots putting up buildings because, a, they can work without breaks.

A, B, they can, you know, do the things the same way every time, whether the human gets lazy, sometimes I am human, I get lazy, or you might get sloppy about certain things, or or maybe you're not as good as the best person in field. A team of like robot construction workers who are very good at doing the thing. We'll just go to go, bang, bang, bang, bang, bang.

So this is the argument that, like, yes, the a lot of the White color jobs have been kind of sucked away, at least in some ways right now or even some creative jobs. But the the labor jobs, which are really a vast majority of the world's workforce, I think is still in labor, that's it's gonna get disrupted pretty quickly. And when you think about what that means for the world economy and and like, yes, will the economy go up because more opportunities and more things exit. When you think what those people who won't be part of disruption or be part of that rise, IT gets to be a little scary in terms of what we could be looking at no.

ten years from now. And I mean, we've always seen robots do this right of last years in things like car factories that you you've seen robots. And the distinction has been that they can only do essentially kind of the same thing over and over.

They can be intelligent there, literally preprogram motions and a part of a their built and they are really good at like a wilding a thing, a specific spot. But um we think that they are missing that people do have in areas like construction is the ability to um kind of be flexible and um get a rise to specific needs of a given place. So I would imagine construction and you have a lot of situations where different construction sites have different layouts, have like a bunch of stuff all over the place.

And that's where humans are much Better. And it'll take some time. I think you you will. I already know there are kind of drain specific robots that can be used for things like painting.

Are you working a longside humans? And we will have definitely more robots alongside humans for the first kind of while. But then eventually you will have human robot. I can do what humans can do.

and and that's corn painting. That is the worst thing I hate painting, taping corners and taping walls down. So the robot can do that is going .

to save me a lot of time already. And now onto projects and open source, a couple pretty needs stories here. First one is about deep seek. And they now have a reasoning model called r one light preview. So this is a reasoning model in the same sense that, or one from opening eye, is a reasoning model.

IT is a train that is optimized to when you give IT a prompt to have h twenty ten, twenty thirty seconds of thinking, right, breaking down the task, planning out the way to answer IT and then doing like individual steps, uh, thinking until that gives you answer. And so they have this now open source version of that people can get for uh, commercial use uh, for not commercial use. We haven't released yet, I don't think, but we are going to we said and on various benchMarks that deal with a sort of thing not quite as good as a one preview on on the more charging ones, but a significant Better than a stupidity for over non reason, various.

So they seem to be replicating a lot of a performance gains of a one preview from open the eye. And you can go. And as you play of this on of their web U I and um yeah very similar IT seems to be in practical.

You play with that. I spend some time with .

you with all I not know. I'm curious what you found.

So it's super interesting. Um the thing that's different here, apologies if you if you just mentioned this this but like IT actually shows you the reasoning where opening eye hides IT right. So only one specifically doesn't not show you step by step what it's doing and is just a really interesting thing to read through that.

One of things we mentioned this on our show last week um somebody pointed out that IT that IT look through the reasoning as IT asked how to count how many hours were in strawberry and when IT comes to the third r the reasoning line goes, wait, there's a third r like it's I got this really like human way of like discovering this thing what I found this thing about this is, yes, a first of all, it's a chinese company so you know that they're working in a different M O. That the american companies are um too. I think there is a quote out I see if I can send this year later but there is a quote from the CEO of deep seeker, one of the cofounder ers that just came out where he basically said, like, look, the mode that opening I has is not that big.

And our mode is our team who we are. It's not the technology. It's what we're doing as the team. And I do think that's an interesting thing to think about going forward is that the people who are going to continually develop new things on top of this stuff might be the mode, like, do you have the right people inside of what you're doing? And you know, opening eye has lost a lot of people and now theyve run a lot more people in.

But like, that is an interesting argument in overall, though, you can do this right now, like you can go to deep seek dot com and play with that. You get about fifty three a request, which is quite a few, and it's really fascinating to do so. Definitely takes some time to spend with IT over thanksgiving weekend or take some time to spend with this and go try IT because it's just fun .

to see how the A I reasons through your request yeah exactly and you can use IT through their a deep seek chat interface.

which right that's where that's their website and maybe not I started that part out.

Yeah yeah it's it's somewhere you can google check deep yeah. Also, if you haven't played around levies like fought to models is very interesting in general to see kind of what we are capable. I've had like for all one has been very, very useful for me as someone who works in programing and have have been working on a some of complex project.

Uh and i've had like you know super long conversations where we are write like complex programs. And we go back and four, and you know, it's not perfect. You can do IT right to first time, a lot of the time.

But uh, IT can do a pretty good job and definitely speeds me up by like days of work, right? And next up, we have just one more open your story. And IT is open scholar.

It's an a eye system developed by the Allen institute for a eye and videra washington. And IT is designed to help peecher access and sympathize scientific LED literature in that domain. IT is seemingly more effective than dupuy for or other child bots. So as you might expect, IT uses retrieval of, uh, they say, forty five million open access academic papers. So you can ask you a question.

You can look up among all these a publish studies and and research papers, and then I can answer your questions with stations to those questions and ah yeah I guess it's not entirely surprising to me that IT is possible to build something in main that is the main specific and works Better by just dying IT directly to a good a database of research and really optimizing IT for correctness and the precision so yeah very exciting. I I I know academics and people working in research can use this uh for answering questions after you do need to kind of bring yourself up to speed on a given topic that he may not be very familiar with. And these kinds of systems will make IT much easier.

I hope they're trustful. That's the thing I keep coming back to. And again, I know we'll get there.

The hardest thing I keep coming back to, I worked on a strip for a youtube video couple years ago and I used a one and did a bunch of things. And and and I think this will help be a trustful. But I had to check things again and it's mostly there. But there's like one thing if I had said that would have been very wrong and like that does suck stile .

sometimes you know yeah uh you know obviously for people doing research, uh it's pretty much .

ant is very important .

that thing that says is actually true ah and there have been funny stories about you know like things at hosinia tions finding their ways into research papers and stuff. So still you know if this is done by university in partnership with this kind of research baLances for um i'm guessing they really optimized for those kinds of people and and probably is a tested and and optimize to not lead to many stuff. Moving on to research advancements.

We have a pretty short section on this front uh today and will probably not be covering in quite as much of that as we tend to with Jerome, who like goes on for five minutes to really break down the details of every paper. Uh will be a little more high level. And the first paper is there's a blog post from one proc called the statistical approach to model evaluations, which is for paper, which has a exciting fun title, adding arrow bars to eval's a statistical approach to language model evaluations.

So the gist of a paper is in general, when we cover benchMarks and we can cover the numbers that you get on a given test, we are uh telling you the average of across all the problems, uh, what is the average correct state, for example? And that is not ideal, according to newspaper. They argue that these aren't sort of the best way we can report performance.

Uh so if you look at statistics and you can have more precise um kind of estimates of the actual outcomes of a given experiments for evaluations, uh, you can, for example, evaluate multiple times and different subset of the questions. And instead of providing just with average, you provide kind of average average and you can also provide the distribution over different scores. So that's one recommendation.

Theirs is to do that using the central limit fear. And we don't need to get into IT but registers um um evaluate multiple times on subsets and and look at the distribution over those different evaluations. Uh a couple more a kind of suggestions of that type here where're talking about standard errors, reducing variations of questions, doing pair differences.

IT gets pretty in depth and intervened groupies. But at a high level, I think the interesting point here is that there is a Better way to do evaluation that is more precise and probably more accurate and less uh, sensitive to differences in benchMarks, so that aim to do the same thing. So very interesting to me, I think, to see if people will start doing with um it's definitely kind of pretty sensible that you'd want to use proper statistic techniques.

Yeah I mean, i'm going to speak for journey on this because I this is not my area of expertise, but IT sounds superfast cy, I mean, I was reunited that frontier math story for a couple days ago, which talks about the idea like how you know it's really pushing for like what these elements can do by creating very, very hard problems. And I think benchMarks in general have always been a weird thing.

Going back to video games, if you remember, is a big video games. But like you know, it's never really one set thing. It's like how you test IT, what IT is. And just for the Normal personal they are like, of course, we want Better versions of this like that seems like the pathway to like Better understanding how these things work.

right? And the this brings a kind of reminds me of there have been kind of multiple papers like this in AI academia over the years as we've gone into a very benchmark heavy paradise. So in reinforcement learning, for example, you know, it's even harder to benchmark than in supervised learning in your first man.

And you like start an agent in environment and then train IT and IT. Turns out, like tweet, one parameter completely changes your performance IT may be that IT isn't your all government Better as just your random seed is Better. So there hasn't kind of movement towards that.

Now when you report to the performance, you don't just report the average. He also reports we kind of arrow bars of variance. And this plays into that, and I think makes us a sense. A second paper here, this one is a few weeks old that I just wanted to note that I will probably be digg into this more and scaling episode. Whenever we get around to recording IT to jeremy, the paper is scaling laws for precision.

So after we talk about scaling laws, right, that's the idea that as you increase with parameter account of the dataset size in a one training, you get a pretty predictable change in performance on the core metric of how good you are at predicting stuff. And here they show that, uh, you know, in quantization, which we also mention a lot, where you train with weights that have a lower resolution, or you, I should convert a train model to one that has a lower resolution for meter efficiency, you are using the space faster. What I show, uh, in short, is that reducing quantization is effectively the same thing as reducing the number of weights.

S in your model is similar. So there's no like free launch here. Uh, quan quantized your model. Making IT smaller does lead to a predictable loss of performance. And so I take away here is that, you know there's no way around scaling. We are not going to be able to train smaller models that are Better indefinitely or smaller models by wave conversation, for example, of having smaller resolution of raids like at some point you're not gona get four hundred billion uh lama, four hundred billion promoters worth of performance.

Let me question about this only because this is something you i'm not a supertanker person when you look at this and sort of say, you know that inference computer argument that the more the more of the scaling going to come from inference compute IT makes sense to me that it's not about smaller models, but IT could be about smaller specific sets of data or specific AI agents running on the larger models that could give you the sort of behaviors that you are looking for in this sense. Is that kind of what this is saying?

Well, this is saying for a single model, uh, if you quantize IT either previously to or post training, and I think you're post training, one is the important one where, let's say, you train a four hundred billion parameter model like lama, uh, a lot of times is a bit shown that you can vent quantize IT, make a smaller but retain most of performance well, yeah, the point is like there is now you can't yeah like you're not going to be able to compress IT indefinitely losing performance. So it's a pretty important insight in terms of the current trend of scaling up data centers, right? Is just we will need to keep scaling the .

centers in power, baby.

That said, that's right. And a couple of more stories in the lining around the first we have survives, which we haven't talked about in a while. They have a new record breaking chip that has really good on these big models on metas allama four hundred five b.

So they have of this specialty chip with giant way for chip that is very different from standard gp user. Even views and allama are able to achieve really, really fast performance, uh, seventy five times faster than GPU base solutions. And this is by using a thing that is a optimized for new or that essentially there CS free system.

So super expensive ships, they are are gonna like cost millions, I think, to get one of these. And they do have a drawback, which is that you are not able to batch a bunch of inputs. So typically when you're looking at charge you pretty or something regret, they take a lot of individual inputs compute from in parallel to get a bunch of outputs.

And that is part of how they can be fast here. This is for a single input, your much faster. So if you have like an on a premise reason to have lama and L M, maybe that will benefit you. But uh, not necessary gonna be important for GPU users yeah I am so curiously .

know about this kind of like A I don't call novelty egypt, but like especially egpc business because IT does feel like you know there's a world where these could be incredible, important, and there's another world where this is the data max to the V H. S, right? Like which is always a tRicky thing to figure out.

Like if the vast majority of use cases is on the vhs, if IT slightly worse technology, is that not gonna the one that just wins out? H, I think both reverse, and I think rock IT does is something similar to this, right? Like or they're doing a similar sort of like their own chips.

And like I used rock a fair amount gearoid rock. Um and it's incredible how fast IT is. You know IT really is like if the speed of its amazing I just don't know enough about the back end of like the scale and will people wanted to scale like that?

What IT feels like to me exactly igg Q. C, H has their think it's like language processing unit. L pu, this is like google TPU, but more tailor to language models.

And they are providing A P, I and competing in the A P I space for fast, indifferent, satish, cheap. So it'll be yeah, it's a very, I think, big question and interesting question. However, brock can sort of take the lead or or carve out of space for itself by having the advantage of customized hardware or G, P, S.

Well, exactly. Moving right along to policy and safety. Just two stories of this time.

The first one is, uh very kind of specific to my region of the words of the bay area that sam allotment will coach share the uh sun from cisco mayor s. Daniel Louis transition team. I don't know if this will be interesting. A lot of people, personally, I guess I was uh kind of interested in this so opening I is based in some france disco and we had recently had elections.

There is now a new mayor that is coming in to um you know will be in charge of sanford this go and per of a story sampled will be cochero his transition team meaning that this pair presumedly will have stronger ties with the tech industry. Um and a lot like a sco in some way, is the hub for a eye startups, especially vender's ay area. If you're working a ye, you're probably up in sanford, sco and not down here in paralo or mondeo, whatever.

So this actually could have some significant implications for A I startups in the area. Although i'm not kind of new once enough in politics to know how much you influence some albums have, but some altman also has been going to washington a lot, has been talking to a lot of politicians and trying to influence, uh, policy. And this seems to be in mind with that.

Well, under is a government california. I have a lot to say about this. I think actually I do have something say, but I think is interesting is so i'm considering a move to go suit right only because i'm an L A person.

I've been working in L A in new york and i've been in media. But and that deep in A I and I think A I is more specific around services so that IT is the place to be. IT feels, again, back to the web, to our world.

IT feels like the two thousand, two to two thousand, six time where you saw a lot of people move to the bar to work for companies like google, facebook, other stuff. I think sam cisco has been through the P. R.

Nightmare of the last three to five years, started really before the pandemic. But the pandemic mean IT worse. Obviously, homelessness and all those problems have become a major thing. I don't know a crazy amount about local centers sc politics, but I do know that Daniel lis, one of his big things, is trying to like, make the place safer and make people feel like it's place I want to come back.

Same goes an incredible city, by the way, like I lived in new york city in brooklin and IT is close to new york as the west coast's, like much Better than L. A. I think sam go is going to go undergoing renaissance. This is just like kind of good timing on sam part. This is the time that I think it's going to come to fruition.

An and then probably in a lot of ways, has a huge vested interest in this because his company conceivably is on the path to becoming the next google and the next or the next larger company where there are hundreds, thousands, maybe in tens of thousands of employees that would come to work at open the eye. And I think he wants to stay in the city. Now, I maybe not, but like, I think the city itself is a special place in in silk valley, obviously is special in its own way, but sometimes goes the city is pretty magical. And I just think in general, this is good timing for everybody if IT works out well.

Exactly yeah go has been struggling in recent years. Uh post covered with like uh office vacancies as an example, a lot of tech has been leaving, stores have been closing. And I has been one of a bright spots in terms of kind of IT being really centralized in the city opening. I was there since the beginning, I think and with their accent, kind of I know a lot of founders. Can I just go there and, uh, rent apartments and, uh, there is even like a cerebral valley, which is one neighborhood and like many people are, yeah, yeah, yeah and I say, don't believe that we r nightmare.

It's not like the apocalypse of of their some rough areas, but every way will be interesting to see what this new mayor of a some times ago, who defeated the previous mayor pretty significantly in the election just a few weeks ago, what, if anything, will be doing for tech entrepreneurs and A I I guess we will seek. And I just one more story, a pretty quick one. The title is bidens final meeting with SHE ginning reaps agreement on A I and nux so uh after they met recently in person, uh, they did announce that they agreed to avoid giving A I control of nuclear weapons systems, uh, which I guess we all want that to be the case. So nice to see that IT came to be and you know I guess not a hard concession to make hopefully for these leaders. But nevertheless, it's is good to know that you that is the plan.

absolutely. I think that makes less sense.

And just one more thing to discuss and a synthetic media art story. And this is kind more of block post when new story, but I do think it's interesting. So astral codex ten, which is A D major blog kind of thing that is read by a lot of types who live in the bay area, rationalist and so on, recently had a little experiment, uh, what they call the A I R turing test.

There were fifty images, half of them where A I and half of human created, and they had eleven thousand participants. Try to see which ones are A I or human like, try to distinguish them. And the result was sixty percent correctness. So, you know, random chance will get you fifty, fifty. The outcome is most people can't really tell if it's coming from a eye or not that people actually preferred the eye slightly ah and I guess this is not surprising to me i'm i'm curiously as surprising to you gain but no.

not surprising me know no interest kind of follow up on that. They saw that cope AI commercial. They came out last week. There was a big story that like you could, I know one of the people that made one of the, there are a couple of them, one of them.

And he told me that that commercial tested through the roof, like the people didn't really care that he was being with the eyes. Don't think that a lot of them didn't know. Obviously, I think this is where the stories we tell around our matter more than the art itself sometimes.

And by the way, like if a pizza moves me and IT moves, I don't care if it's A I generator. Not like to me about what my interpretation specific is of IT is now some people might very much disagree with me. They might be like, I need to know that there was a human being that did IT.

And again, I wanted know their story. I think most people, when they see a visual thing, they just react right. They're not like trying to ask himself first, is this a computer that made this is a human that made this, are just looking at something.

And it's been in part, it's now mostly hard to tell what is real and what is in. Now there are some people who can very specifically say, like that is not a real person because I can tell that they're shading on this part of their faces off. But like they fix hands, they fix a lot of stuff.

I think we have to stop like bar, like maybe we have to stop thinking about IT does IT matter. Ultimately, the matter is like the the, you know, the credit in the pay, all the stuff that they use, the training. I think all that matters, art is art.

And I think this is the weird thing that people struggle with forever. Like people argue that, like bosky ott stuff was not art, because IT was scratches, and like collages, and like because what I was out and IT is held up to this day. Or or picasso, like what this guy is making, this weird tube is looking stuff that doesn't look like a portrait and like that just kind of the world that we go through as humans.

And weirdly enough, I miss the A I weirdness aspect. Like sometimes A I art isn't as weird as I used to be, and I miss that. I think that was more art. Like sometimes.

yeah, I have kind of a pet p, where A I art now is tony mous with text to image and, you know, looking, you could go back to two, eight to nine. And the people are doing really cool stuff with gangs. Yeah, amazing stuff.

And that wasn't text to image. That was like training your own model and and doing kind of more outward things. So I think there is no actual artists who do A I and you know, with this experiment, again, not too surprising.

I think it's you time to admit that A I stuff looks pretty in general. IT doesn't look necessarily kind of like the high end of art, but IT can look very nice and exactly now have the questions are more to do with the effects, the economical side and not so much wealth metics. Although I guess one other problem is the potential for IT to flood the internet and kind of end up with too much AI generic kines. And that would be an issue.

I wonder they i'll say I hear from our from our audience from time to time is the idea around A I films and movies and and people that are feeling scared about the idea that they gona take their jobs. I think a lot of people talk about the race to the bottom in terms of cost and and how that can go back our conversation before you know that all that everybody can make music, that sounds pretty good.

What value does recorded music have in in a big way now that for as soon everybody might build to make prompt to movie, what value does like a story have? And I just hope that like fandom still exist for specific things, and by the way, that could be an eye creator or a Normal creator. But like fandom is important and having like real care about things.

And I think people who make A I R will still have those fans in audiences. I think the generic stuff probably. One which is okay. But if they're so much of IT, IT might make IT less valuable.

which is a tRicky conversation. Yeah and you know personally, i'm sure you found being able to make little songs of the eye to be super fund. I know you know ever since I started doing for the podcast has been pretty delightful and entertaining.

So you wind up like unlock king new types of stuff that you would would have never had before and exact exactly and that's IT for this episode. Thanks to everyone for listening um as I like commercial, you could go to last week in that I for the text new letter and thank you again for cost. And I think your experience was useful on this one.

Thank you so much and is always fun. And if you if you have a chance, come check out our show, A I E for humans. It's on youtube. You can find IT there or you can find IT at our at our website, A I for humans dot show, which has links to audio and .

all other stuff. Yeah, a lot of fun. Go check that out and you know if if you need to visual, and we do appreciate if you share IT subscribe to IT bait, all those nice usual things.

But more than anything, we like to see what people are tuning in and getting benefit. So hopefully we will be another episode next week. A sometimes i'm bad at the scheduling stuff, but, uh, everyone, enjoy this A I ultra song.

Step into the little dream where pixel come alive, A I paint species in real time washed the future, the code.

For the c reality, what the new group.

take. new. In you know you. Rhinds fast.

Is fire away?

电脑 却 比 现在。

Public story IT is A I don't arise. There's nothing Better. together. three. boys. High per last day are amazing. Pretty air in a small should prediction is where we are. It's not very.

Enjoy the right.