#186 - Adobe AI Tools, Tesla's Cybercab, Nobel Prizes
01:33:54 Share

In a world of one and zeros, pain and adult dreams take flight, protest the future, ruling down streets captions with all the a shake no, the patient ary surprise stories of a week will take inspiron, begin any eye.

Hello and welcome to last. We can at podcast we can hear that about what's going on with A I as usual, new epsom, we will be sumi zing and discussing some of last week's most interesting A I news. And as always, you can also check out our last week, a newsletter at last week, in that I, where we have weekly emails with even more articles, even more news and also everything related to the podcast with link and so on.

I am one of you, a regular hosts on a crank for background. I, the stanford. Now I work at a start.

And once again, journey is still out for eternal leave for a little while longer. So we do have a guest host once again. And once again, john cone is feeling in that role.

I'm your regular irregular cohoes.

Yes.

yeah, yes. It's great to be back on. Thank you so much for inviting me again on ay, and we hear jeremy is doing well. Everyone's healthy, baby healthy, mother's healthy. That's great.

That's right. It's yeah going well. And from what he told me, he will be back routine soon OK ever in an episode or episode that he will be getting jeromy back which ensure any older regular listeners ah will be excited as am I.

I mean it's an invaluable income stream to be co hosting the show. Every episode that i'm in here, i'm like a million dollars go here. Germany is missing out on that so .

he's going to get back yeah you get paid fun and new awareness ah do get .

lots of news about as I do like that about IT. I guess in case people haven't been listening um to episode I been in before, I will try to keep this short this time i'm cofounder and chief of the scientist at N A I company called nebula. I wrote a belling book called deep learning illustrated.

I host what i'm pretty sure is the world's most of new data science podcast called super data science. Andrea and jerrem have both been guessed on that show. And yeah, so I mentioned a little bit last time that I was on that there's TV stuff developing and that is continued. There's like two separate TV projects that are that i'm really excited about, and I can't wait till I can share some things publicly on those.

Wow, that's really exciting. All I will be keeping an eye out. And i'm here like we did promote the podcast for quite a while on the show. And so I would hope that a lot of listeners of last ki are also hands of super data science.

There is at least one because I got an apple podcast of view maybe two months ago, that somebody, although IT, wasn't necessarily related to us having having sponsor your show because they said that they heard me on the show, they came over. Now they love super dt. Science as well.

So I all try to look up while you're doing your apple podcast updates. I'll have a look at mine to see what this person's kind of name is and what they said. So I I know that this is one go .

you're you're else getting paid and listers hope. Well, uh, before we get into the whole set of news, I do want to try something a bit new, a just a quick news preview to let people know everything is coming up since we do have like to our episode that is so smart.

that is something as a listener. So something that i've got to say is that this podcast is the only podcast that I listen to you last week. And A I, and so that is something that I, I, I man. This is on my wish list, is listen .

to your users. So to give that preview, we will be A A bit lighter I news this week that isn't quite as much heavy heating topics. We're gonna a lot of news regarding doby.

They just had an event who are they covered a lot of new tools coming to their sweets of creative tooling. We will be probably talking about tesla for decent while as far as business goes, they had their big we robot event with a lot of cool stuff. Maybe not uh, well, we will talk about IT.

Uh, then there are some exciting open source projects that will be uh maybe a pretty big topic will be talking about the noble uh prize awards, of course, which happened last week and had quite A A lot in the I and will be talking a little bit about policy and safety is not quite as much going on, probably talking quite a bit about anthropic and what I posted. So they go it's gonna kind of a mix week. Nothing too huge. But I definitely some interesting things going .

on should pause on all safety and policy stories. So jerrem comes .

back then just catch up. Yeah, yeah, yeah. And then before diving in, as usual, quick shot out to any listener comments and corrections. I did not have to catch much that front this last week.

So in law of having those kinds of comments, I did just wants to say on apple podcast, I took a look, we do have two hundred twenty six ratings now, which is, I don't know. H I used to be like two hundred a couple weeks ago, months ago. I yes.

So I imagine, yeah, hopefully IT IT seems like some listeners are leaving reviews, uh, just star ratings without reviewing IT in text. And thank you as well. That does presume help in our government.

I don't know. IT feels nice. Our average went up by point to one four hundred and seven out of five.

So that's really knowing I am going to have to go in and give a one star because we're still at four point six a super day is science. So you guys going up to four point seven is completely .

unacceptable to me. I'm sure if you got more reviews that go down, you know you got a revert mean. So on right so .

um and I did find here. So on August twenty third, someone I don't have. Do you know this person? D three, two, one p is that in front of yours?

You know I probably would forget a name. Yes, he was.

So d three, two, one p wrote in an apple pocket of for superos science, said, i'm a loyal last week, an eye listener and hurt you as a guest on the show and figured I would hear your episode. I'm glad I did.

great.

That's nice to hear. thanks. D three, two. D three, two, one per really rolled off the town.

That really does. Okay, let's get venues. We start with tools and apps. And as we promised, the big top here will be adobe. And they've had the adobe a maxed when you when you for create the conference and as usual, as we search events, we got a whole, a bunch of news regarding new stuff.

And a lot of IT was A I and the big news was there A I video model? So they've had their image generation models for quite a while under the firefly umbrella. And now they have A A I video model.

We've had an beta for a decent amount of time, uh, just for people to use, uh, i've I believe in h website form. They canna previewed IT. Now he is available in the beta in the premiere pro, which is their program for editing a video OS.

It's one of a kind of leading products on that front. And you can extend footage by up to two seconds or make mid shot adjustment, and that's just one of the tools here. So they have that uh general extend. They also have text to video, which is only available as a limited public beta on the web APP.

This would allow you to use uh you know the usual kind of thing like sora like recovered last week of movie IT is able to produce five second clipsed with uh not quite H D resolution but pretty good looking quality. Uh, just looking at the examples they put out, IT is up there in terms of the footage not being obviously A I for the most part. And they do have also image video.

We have a reference image along with a text prompt for more control. And they say, uh, one of things you can do is camera control, like saying we should pan left or right. Last note, as with all the firefly tools they do save, this is a commercially safe, meaning that they didn't use code data to train. And so you should define to actually use IT to produce a commercial video.

I suppose. So I know that you're an adobe fan.

Have you use this entry? I have not a because I don't do much. Aside from editing a and the generative expand h so far is not very useful.

Maybe I add some special effects, uh, which would be interesting. They do showcase their ability to generate uh, kind of nice overlay of special effects or I want to call IT. I did play on in photoshop with their tools that are I based.

They really have some very nice things, particularly when IT comes to cropping, like the ability to crop an individual object is just totally different. IT used to be like a time suing process. You had to let go all around the borders, an object.

Now A, I will just do that for you and it's great. Similarly also like the um infill of images to reduce noise, just a lot of nice stuff that is directly integrated. And I would imagine that would be kind of four videos ers people create more kind of creative videos. This could be useful for b role and other things like that.

Yeah, it's super cool. I watched the video of this AI video model. Of course, there's going to be chair picking. The ones that don't look at generated you to have do how could you not do that? But similar inly, like you say, some of these features are really cool, like being able to have a existing video clip.

I mean, you could take this photo to we have very now the youtube version of this part guest, and you could put like of, you know, flames behind me or you know, we talk about firefly, fireflies, I guess, or around this or whatever. And so that's a pretty cool functionality, being able to change the camera angle. And I haven't seen that. I don't think i've seen that in the other generated .

video tools yet. Yeah, we IT could be the case. Maybe that you can try IT with movie je, since I do support some fancy like video editing of H I think that was primarily for changing russia content of a clip, not to camera.

So that could be something unique to us IT wlink make sense. Uh, because this would be pretty much, you know, meant for actual usage by people creating a sort stuff. So IT might be different from sora and movie jen, for instance.

Moving right along, we get a couple more stories waited to adobe, and this is about upcoming things that are more in a previous stage. So they previewed a couple experimental A I tools. One of them is a project scenic, which will h generate free d scenes based on user input, and then use that as a reference to generate a to the image, something that I find pretty interesting.

So are being that you do have essentially sort of, uh, a set of objects, you know how tent, something like that, and you lay out a scene, and then you can use that as a reference to generate to the image with your typical A I model that will kick start that process with that pretty structured free scene. So not something i've seen and something that is interesting. They also did review project motion, which is for creating animated graphics in various styles, and project clean machine, which is an editing tool that automatically removes destructions and images and videos like camera flashes or people walking into frames. So they go they've been announced at the same conference as the m sneaks in development projects, so will probably not be seeing this integrated for a while. Maybe I want to come out, but you know, as someone, as people who track a research, some of these things, things that are pretty interesting and necessary, something we do see in other tools or in research.

And I was saying before we started recording that thanks to this podcast, I do get a good sense of what's going on with doby innovations. And i'm really glad I get that because we hear, for whatever reason, whenever opening I released something everyone hears about IT anthropic same kind of situation. IT doesn't seem to be as IT.

IT doesn't seem to get as much reach when adobe makes his announcement. But these are big deals. These are eat at one of these kind of capabilities sound's, super useful. And the key thing that adobe is getting right here is that assuming that these capabilities work well, having them in the flow of somebody using an adobe product without having to go off and use some other tool, having everything integrated like that, that in my view is the key to success with general AI.

exactly. We were talking a believe just in last episode. De with movie um you know movie there is a cool just like sora, but on ever hand it's not released and it's maybe not even practical to use because of the amount of computer, the time that will take, which was also so sorry, I to take way too long to actually use in practice and also IT IT likely isn't actually targeting a major pain point for many people you know generating uh, video from texas, cool.

But now on aside from people, it's not necessarily the case that IT would be useful. They do have video editing as a major feature where you can revise aspects of video. That's definitely one of these more practical things that would be useful and these source of things like um you know this project motion, which is for creating enemy graphics project clean machine, we definitely like sexy, right, but we're presumably much more a useful in those kinds of workloads.

And the one last story about adobe, they have another experimental protyle called the project supersonic. And this one is meant to generate audio effects for video projects. So this would use a text to audio, object recognition and voice invitation to create background, idea and side effects. Sound effects.

So uh so this is something we have seen from other companies like eleven labs with a aud effects being V A short snipe's right of like and our ball high ground and seeks like that that you obviously in videos to make them really be realistic and which is you know decently less hard when generating full on music. So yeah, this is, I would say, definitely something that seems like IT will be coming to, uh, adobe create sweet of occurrences. IT is just a demo, as you described .

having this as part of the creative sweet again in the flow general vi and addressing a pain point like he said, that's another really important thing to ham home here. If you're not addressing a pain point other than being cool, how is that going to be a bit success? So yeah, lots of a flame for a job. Keep IT up. I don't .

lan round. We are moving away from a doby, but we are keeping on the views kind of tools with her story is about to to expanding A I audio generation tool to all U. S.

creators. So we've had this dream tech, dream track tool that was first introduced in twenty twenty three. IT would allow users to generate short instrumental audio clip based on text prompts.

Uh, so I would basically create royalty free soundtracks, about thirty seconds log and as per vertical. Now IT is being released to all U. S.

creators. So that would presume help a lot of people making short clips on youtube. Youtube does have their shorts thing, which is similar to tiktok and instagram reals.

And you can see how having something that do you free could often be something useful. So very machine keeping with youtube, also releasing tools like vio for text to video also on youtube. IT seems .

like we're almost just staying the same story again. It's like exactly it's exactly the same story as with adobe. These are in the flow tools that are useful for people uh exactly like being able to generate shorts easily, clips easily uh instruments, audio. And now yeah this a dream tracks yeah very cool. I'm sure it'll do well and it's great that there offering IT to all the creators.

exactly. And um I think gets its interesting to have these stories now because you know IT was just at the beginning a year, but we got sore right. IT was last year a tex tivo was very much nicer.

Sora was very mind blowing at a time for the quality video IT could generate. And now you know, almost near the event of a year, we are getting to the point to where these things are actually being deploy and commercialized AI video tools. So that tells you something about the, you know, the speed at which AI is being developed and roll out by different companies.

I guess a key distinction here, my assumption is, I don't see IT explicit, but I believe all of these tools being offered to us creators by youtube, these are free to use. yes. So that is a distinction relative to adobe and IT also. So the assumption here is that alphabet downstream will get more revenue from having slicker uh slicker videos and creating an ecosystem that creators want to be and using and publishing videos in uh to continue to maintain their position as if I if i'm still correct on the stat. Youtube is the number two most visited site in the world.

Yeah, youtube is massive for people who don't go very much. Is in terms of time usage, you know, is above instagram, a above tiktok, above many of these things that you think of as places where people spend time. And my oppression is youtube shorts haven't been quite a successful as those our things.

So could be also a reason for google to be pushing with to try and push her along. And speaking of goole and rolling out tools, more users, we've got a similar story now. All germany users can generate with imagine free so we have had imagined free for quite a while.

IT was pretty limited and now you can use IT in games similar to uh, judge bet. You can generate images in germany. But just saying prompts with like draw, generate, create uh, x image.

Uh, interesting, these generated images do come with a cent I D watermark. So if you try and detect whether this came from a gam, you can actually ventilate an image as A I generated or not. This is not currently available for free users, so this is being made available to people who are page have I advance and enterprise.

are you it's interesting this word which you very confidently pronounced is imagine i'm always like, I am never sure I guess I could watch I O twenty and twenty four google I O twenty twenty four to get how they pronounce IT. But it's interesting because for people who are listening to this podcast, and if you want to look IT up, you don't type imagine into your google search, you type imagine, uh, I M A G N, which is is a clever name, because it's like, imagine because you're creating something, you're imagining something while IT is also image generation.

So imagine I imagine that probably makes more sense but doesn't roll off a time quite as easily as imagined so um and I just quick correction spoken so the feature to generate images with people is not only available for free users, of course, jim, I had quite an embarrassing with regards to that when I first rolled out imagine so they have reintroduced IT, but they're so limiting to people who are paying moving right along.

And we are keeping with the theme of A I being rolled out more widely. Next up is made, and they will be launching in six more countries today. Of these are brazil, belize, a guarana, bragi, Philippines and V, U, K.

And they have said they also plan to launch in many more countries in the coming weeks, countries like a algeria, egypt, indonesia, macco and saw so i'll be, of course, adding support for new languages. And these meta AI tools are all these things we've introduced through out facebook, instagram, whats up, messenger, chat bots and so on. IT is not going to include the R, P N. Union due to regulatory concerns that you've seen that with a couple of things and I just met a eye. So uh, people in europe, uh you know not getting made A I not sure we're gonna be upset about IT personally have not used wings and haven't benefitted too much either way that I is being rolled out more widely.

I wonder if you could just like change your location and get access. That seems like a pretty easy thing to do because I don't think for most of these things, I guess like with WhatsApp, you provide your phone numbers that as a country code, but for facebook, instagram, presumable, you could just change where your country is, I guess, and get access to these tools. If not, then great work.

U. K, you finally got something out of break. set. Economy has struggled for years since leading the E, U, but you get that A, I, nice.

Yes, yeah. Your ability to regulate less harshly. U, K, doesn't have V, U, A, I act, of course, because exit does lead to more business friendly results like reads for sure.

And the last story here, finally, we are doing something a bit different. No more A I being rolled out to more people. And this, what is bit intriguing if you are more of that, say, nerd, when comes to A I.

So the headline is opening unveils secret meta prompt and is very different from anthropic approach. So IT might first sound that this is the system prompt as what anthropic revealed pretty recently, the showcase. This is the actual thing that we uh give claud.

When you give you an input, you there's a whole match of text instructing cloud how to respond. We actually know what that is. We have opening eyes.

We don't know the system prompt. And this is different from the system prompt. This is a matter prompt that is used in a tool wares that can optimize a prompt.

So you instruct a the allam here on how to improve the prompt. But anyway, even if it's not a system prompt of the article goes into quite a bit of detail comparing the style of the prompt. So this one is pretty structured.

They have in various sections with guidelines, steps using sort of markdown formatting the versus on public wish. If he looked at the system problem to was a bit more narrative style, lots of paragraphs, stuff like that. So uh, it's into a single look at, again, if you think about things is like system prompts and do prompt engineering is really cool.

And this, if i'm reading this correctly, this they've released this meta probe is this specific the o one model family that's not being released .

IT I would have to check but that won be surprising for me facile case .

yeah that seems like that's what this is um which is interesting and that also explains things like actually i'm pretty positive. That is because there's things in the meta prompt around steps and how those um should be broken down and revealed uh users. So I think we talked about one last time was on the show was every brand new um a couple of episodes ago on last week.

I and I i've been using a one to tremendous effect actually just in the last few days. IT constantly blows my mind and ah I don't know IT is cool to be able to see under the holiday IT is also nice to see OpenAI being of IT open um with what they are doing here. I guess know if there's kind of an openness arms race that anthropic or other players encourage in the space.

That's great. Keep IT up. That's right. And the these prompts are from where are so you know, not quite as public as and topic and topic had more of a blog post and a whole much of kind of discussion of revealing the system problems.

This is in the internals of the opening platform where they also discussed prompt engineering and guidelines, stuff like that. So this is just an example of a matter prompt you might give is not even something that opening and necessarily users, although IT is likely that this kind of prompt engineering is what they do moving right along to applications and business. And we begin with tesla.

So if you're a big fan of autonomous driving and robots as I am, we have that pretty interesting developments from tesla va. Last week, they had a big wee robot event that has been kind of previewed and uh schedule for quite a while where in a moss has been saying for a long time, you know will be revealing a lot of stuff regarding our autonomous driving taxi uh service that is set to be uh major focus of tesla. They held the relevant last week and there were some very cool things, so they did show a new cybercafe truck, uh, sorry, uh, a new cybercafe car.

This is a futuristic c looking car, kind of, if you take the cyber er truck, can dense IT into a regular ish sedan with two doors, a no steering wheel. Two seats, the doors are, you know, go vertical when the open, not a regular door. So lots of fancy size.

I lukie stuff basically in the idea being that these kinds of cars will enter full production in as a two, three years. Twenty, twenty six was the promise is said before twenty twenty seven and this would be how they roll out the taxi service. In addition to that they also had a robot van unclear why it's not a cyber er van. But either way, robot van kind of a similar idea of a futuristic about capable of sitting twenty passengers instead of just A A couple on on top of all that.

Also at the event, they had a whole bunch of their optimist humanae robots just for out the event, they had them serving drinks, they had them chatting to people and you know where a lot of promises being made or projections of these, uh kinds of robots being available for 2 fifty thousands dollars and being able to do all source of things like being personal assistance。 So when you go, lots of cool things at the event, but not very many specific details. And even the demo of the um of this new cybercafe was on rails.

IT was held at a movie studio. So IT was a very safe kind of demo and the optimist robot seemingly were tele Operated, very not authority. People were in control and and talking to people. So the stock rides actually went down by decent, mild by eight percent because, uh, investors did react, thinking this might be more style of substance.

Yeah exactly. And when you you just got an inbuilt meme really nicely there when you do on the hollywood set. And IT seems like what your pitching is, a fantasy more than a reality.

IT really worked all too nicely for that, for that hit in the stock Price the next day, much higher expectations then we're realized with this, evidently from an investor perspective. And part of that is that if you think about a big competitor in the autonomists driving space, o is the leader here. I think it's fair to say question.

And wao has a alphabet rather has invested thirty billion dollars in amo in developing a complex system involving radar lidar as well as very detailed 3d mapping of environments that works in conjunction with just video cameras。 Whereas tesla is proposing here, is that the cyber cab? And I guess later, the robot van, in your rate, why isn't that the cyber van? Cyber truck? Cybercafe robo van? Anyway, maybe the cybercafe going to be something else, the cyber vans going to be something else.

And so what tesla has been trying to do for years is to have their quote unquote full self driving, which is a marketing term as opposed to a reality. Um and so I I didn't episode actually on the five levels of self driving cars. If you want to check out super day to science eight one zero epo de eight hundred and ten.

I specifically talk about the five levels of autonomy that you can have in cars. And while full self driving is described, full shelter, and this is solely a marketing term, IT is not full share driving because you still need to have somebody standing behind, stand behind, sitting properly. If this are very short, they'll be sitting behind the wheel of a of a tesla vehicle.

And and part of the big hurdle that tesla is facing is that their AI team is trying to develop some kind of way to have true full sell driving, not just marketing full sell driving where you don't need to stern, whether you don't need peddles with only video cameras, and that may never work. So you they talk about about the cyber cab having roll out in probably twenty twenty six, said elon musk, but certainly before twenty twenty seven. And I don't know how both of those things can be through probably in two thousand twenty six, definitely before twenty to seven.

Um so I guess yeah around twenty and twenty and six hundred and seven. But that depends on regulatory approval. And having just video cameras could be really tRicky way to do that. The gambit here is that IT will be a lot cheaper.

So we mos approach while IT has created these beautiful jaguars that are really nice to see in um you know you have a really luxurious experience and the drive is amazing because thanks to the light, the light are the rate, are that detailed mapping as well as video cameras, you feel really safe. The experiences is amazing. But those cars are very expensive to build.

And so IT actually means today that when you're in a amo, IT costs more to create that experience for you, then IT does to have a human and driving the car. Now of course, over time, those costs will go down. You get economies of scale.

Um but yeah, I guess the idea here what what tesla is hoping to be able to do is to be able to leap frog all of that by trying to have video cameras only work and then that allows them to sell these um v cybercafe s for just thirty thousand dollars or less where's uh from memory of my memory of the of a major about four times that um so yeah, so I guess we will see how works out. Definitely a big risk could be big rewards. Um I guess shareholders at this time are betting that the risk reward ratio is at a baLance .

exactly and tweet fair. I mean, a source of noting that tesla is, I believe, valued more than all other other makers combined like this is not a Normal car company that their main business is selling tesla cars, but their stock Price reflects the optimistic future where they go beyond that. They at least you know get A A big share of the self driving taxi business or they get a big share of a humanity robot market, things like that.

So that's already Priced in essentially in the stock Price. So if they can deliver on these rings in next couple years, the dark Price would presume ly take a big hit. And as you said that currently, you know, if you are test the owner, you do have this fsd full sell driving, uh, I think it's called beta still.

And as recovered earlier this year for a long time, F F S D was really bad. Head IT was like scary to use with fsd twelve. And the latest iteration of IT is gone a lot Better. So they have moved with fully A I approach that uses video and training from their giant amount of data from people driving around.

Now, if you use fsd, IT, is everything much Better, much more human like, much more reasonable? But there are also kind of twink of the a terminology, so we now call IT supervised F S D. That's what you have in a tesla is not just fsd, it's supervised fsd.

And so they also promised in this uh, presentation a fully unsupervised fsd being launched in texas and california by twenty twenty five in their cars, right? So that's uh, kind of less sexy detail. But a very important detail is, as you said, they need to get on supervised fsd, right? You know what that means is you can let the car do its swing.

You you can just give up control and not be scared of what it's gonna. And that's the case of lamb now. And they've had to working in such a isco for quite a while.

The testing IT in the lay we're trying to expand to more cities. Sla doesn't have that yet. And it's going to be very interesting to see if very able to you deliver on IT by twenty twenty five.

As I say, next story not quite as exciting but still pretty notable as we've covered the opening eye has been making a lot of deals of this year with very a uh media organizations. So they have now announced a new one. They have made a deal with media congratulate wish owns outlets such as housing chronical, the sanford cisco chronicle as quire cosmic polton.

And so this is adding up to a bunch more deals that you've done. Now these uh GPT and surge GPT will display content from over twenty magazine brands and over forty newspapers as part of this a partnership. And they have many more a deals that have secured over the past year.

And as we've said, you know, IT might not seem necessarily like a huge deal, but this could be a real differentiator from something like perplexity or cloud even right? If you have to pay for access to latest news via v uh media organizations and uh you have to make these kinds of deals open, is the main player that we know of making those kinds of deals. And they seem to you know, continue doing this with this latest development. So would be testing to see if all these stories, uh, if all these deals pay off.

And just two nights ago, at the time of recording, I head dinner with some Peter gold stein who he speaks on, generate A I in in public generally uh, really well spoke and highly educated guy and he happens to be the chief AI strategies for her. And so this deal had just been announced. And so now I can tell you all the secret, think no I mean tremeau festival I got absolutely nothing um you know no behind the scenes .

uh in the details absolutely .

nothing at all. He was uh a constant professional um but yeah that is interesting. Yeah this is just that trend that we're seeing across generate a ee where initially these alembic were able to train by just striping everything. You know, nobody had the robots txt set up in a way to prevent that kind describing because IT was a completely new way of a completely a novel approach to scraping and generating intellectual property. And so now everyone's woken up, including host a neurotic times, tons of these big publishing organizations, springer fair og um yeah internationally, lots of publishers waking up to how much value they can provide to uh the super scales in particular.

And at a time when you know those same super scales are threatening their traditional business models, because generate A I tools prevent you from getting, say, a google search result that caused somebody to click through to an asset like cosme paulton or acquire the sentence of scope onic's instead of needed to click through to that article where the same for school chronicle has ads, a display ads that allows of the general business instead, the geni tall just provides the answer and we don't today have. I mean, jani tools can do retrial al augmented generation and act in a kind of agents way to be pulling information in real time over the web. Often when IT does that, IT does actually provide you with the source, maybe you'd click, but I suspect there were going to, as time goes on, have alams be increasingly real time updated their model ates update a based on real time information so that when you're doing, say, a google search german, I will bring back for you and L M exact answer to your question in real time, as opposed to you need to, you know, click on some goole suggestion and see the result.

And so that yeah again IT eats into the business model of um you know historical newspapers and magazines. Um you know obviously they always had ads, but you were also paying to have a show up at your door. Not a lot of people do that anymore. That's already really een to their business models. So now they're more dependent than ever on ads, specifically digital ads, and now generated I poses a threat to that model.

And so i'm glad they are able to get some a some revenue back from being able to these the the high value content creation that they do um because yeah if we end up in a world wear, you just have gena models learning from gender hated content online. That might not be ideal, though maybe we can figure ways of making that works. So yes, so make sense from the perspective of someone like OpenAI to be paying for jenni content high hold content horse get to make back some of the money that they could lose now in the future.

I guess by not being able to have people click through as often um on sea google search their ads. And in the long run, this does pose interesting questions for not only publishers business models in terms of ads, but also someone like google because you know their own tools like a german I search. They have there's there's rumors that google devine had similar kinds of capabilities to to ChatGPT, but IT wasn't released because that eats into google's core business model of display advertising quick through advertising.

You know, uh, digital attribution when you have G A I models that are just providing exactly the information people want right now, mostly in a text display where maybe you can find ways of sneaking some display ads in on the side. But in the future, you're going to have more and more interaction with these tools through just voice and audio. And how do you then insert sponsorship and there, like maybe you can have okay like kind of boosted results and I guess that some way that people will go. But IT seems like one way or another google's business model will get eat into you know there decades now of absolute um effectively monopolising dominance in search um where that's all paid for british advertising could be a into here.

Yeah so you know lots to save her. I think this is one these stories in general and these friends, not what you sexy, you are not getting like the main headlines and they are not me being topic that's getting all the hype on twitter at sea.

But it's actually like sudden a very important thing to be aware of and think about because IT speaks to really the future of the internet and the future of search, right? And so, you know, comparison to obvious deals from opening, I plexi, I has also done something kind of similar. They have a revenue sharing a model.

They have this publishers program where we've also had companies like fortune time entrepreneur at sea join on and it's it's very much uh, on point, as he said, perplex I D A I source GPT will be uh showing their sources, will be providing links. But for most part, presuming you won't be clicking to go to go sources, you'll just be wearing what the eye says. So we do need some kind of new business model for the people actually writing up for news, right?

Uh, and this increasingly seems like a business model and not just that, but you've seen also read IT and twitter, you know please read IT for sure sign deals to license or data to, uh, I believe a google and that's another aspect of internet where you know any content that we provide on these platforms can be modernized for data. So it's one of these realities that is kind of interesting and h speaks to like the development of internet at large and our information ecosystem at large. Uh, even though it's a so little of dry on the face of IT.

one thing that was reminded of as you were speaking there, you touched on IT a little bit or if LED me to this thought, which I think is really important here, which is the ethical contrary as well, where as this kind of business model, where your the way that you freeze t there, where you were talking about how you know who's going to actually pay the journalists who are creating the high politic content.

And so once you are a newspaper and magazine, where maybe in the not so distant future, most of your revenue comes from gene I companies, how does that affect your reporting about gene I companies? So if a player like open an eye and thrown c cohere becomes like, you know, this unavoidable monotheism of culture in general, really influential. And simultaneously, they're paying all the bills at all the information sources that we get our information from. There's an ethical contrary. There is interesting are you get .

on bias reporting exactly a lot of interesting applications like for the ad model. You seen uh, quite a bit of you know shift to click bate essentially and and getting uh headlines and things that drive people to click through and read the article with this shift, right? That's not kind of as important.

So maybe people will start covering things that are just trendy, but people will be asking such a GPT about right? Uh and there's a whole lot more to say on this. Like new york times, for instance, they i've moved to a subscription based moitie ation model of large part.

So uh, more and more media publishers need subscription type revenue streams, which little bit distinct. Anyway, I think you're said enough openness, but where you go open the eye, paying even more media publishers onto projects and open source, we've got a few in neat stories here, starting with open r and open source A I framework, enhancing reasoning in large language models. And this is actually paired with the paper that describes the whole system, not just open source project.

The broad idea is essentially to provide a framework that allows you to do all one type reasoning with open models. And this is acceleration from a bunch of organizations. University college london, university of liverpool, the hong kung university of science, china logy and some other ones.

So they going to a lot of detail about how the model works. They treat IT as an M D P, uh, right to where you have sequential steps, essentially avail a lam. And for each step you can provide a reward. We have whole party for training the reasoning to be able to do a good reasoning step by step.

And they do show some interesting empirical results where you can see that using of this reasoning um as you increase the generation budget, something you've talked about a lot of inference scaling what you get as a post training, you can now just scale them out of time and resources. You gave your model to be able to generate a Better answer. So here we also do an experiment and show that in fact, this approach results in Better outputs.

And even that's Better than just simple things like majority wrote. Unfortunately, this is coming from some universities. So it's not really evaluated at scale as be small models.

But anyway, I imagine we will be seeing a lot of a sort of stuff with open source efforts to replicate or one type reasoning moving right along. We've got another open source this this time. It's a benchmark.

Is the M L E. Benchmark evaluating machine learning agents on machine learning engineering and is actually coming out of open a eye. So this benchmark is meant to look at machine learning engineering tasks.

And if you're suffering to you, you might find this kind of fun and interesting. This benchmark includes sanding five machine learning engineering related competitions from cargo. Chicago is a platform has been around for quite a while.

They do have competitions where you can submit a an answer to a machining problem, where if you're in the top performer, you can actually earn money and been a many participants. It's it's a pretty big platform. So in this benchmark, we uh allow agents to try and win in these competitions and they have some interesting comparisons looking at, uh, different scaffolds that take actions by calling tools well.

So have a aid thing which is purpose built to perform a research over solutions on cargo competitions. And we actually say agents run autonomy y for up to twenty four hours in their experiment. So if we look at figure for A P T four oh m lab, they have total steps two hundred sixteen at one time, two hours for G P O aid.

They say they have a fairly nodes in the research for a long time of twenty four hours. So pretty no, it's getting very autonomists at this stage and the benchmark pretty hard. So even for the best performing model or one preview to get to a metal, which is to say to place well, to place Better than most people, all one is only able to do that seventeen percent of a time, roughly.

And they, these models often make invalid emissions. So even the best one only gets to eighty two percent valid emissions, which hopefully most humans are able to do. And if you don't do of this aid thing, which is previous built for cargo, you do much worse.

You you know going to single digit percentages on the metal placement and near fifty forty percent valid mission rates。 So a very intriguing new benchmark in the realm of agencia. I soft engineering, ai and IT will be seeing if we can, you know, destroy IT has been the case of many A I benchMarks. Yeah.

this is great. exactly. yeah. We're just constantly having to come up with more and more complex benchMarks.

A couple of years ago, this would have seemed like an insane benchmark. What are you creating this for? Obviously, machine get do this.

I don't know if the machine will be able to do this in my lifetime. I mean, this is something you in your super ty designes pop gas ever sold only. We talked about how mind blowing for both of us, the release of P.

T. Four was and for both of us that was a moment that we were like, holy crap, A G I N our lifetimes, A S I N our lifetimes. This is probably happening.

Um and so you know prg P T four, if somebody had said they were making a benchmark where you're evaluating people on these extremely difficult chine maning engineering tasks, they require a lot of thinking, a lot of outside information, a lot of steps. I would say why that's a waste time. I don't know if we have machines that can do that in our lifetime.

And now you know if this AIDS scafidi I D E A sounds like a key part of the success um that opening I is getting here on the benchmark and so to be getting, you know as he said, o one preview, which of course it's o one that is vastly outperforming all the other models that the text with the eight frameworks that they also use GPT four o they use the biggest lama which is four or five b uh huge and close three point five minute. You know the eight framework with those models, as you said, that was the only way that you could get kind of any decent performance like the other than that you know other than using the aid approach, the best score was four percent ah four percent chance of getting any metal in these competitions. Um or maybe you know maybe a good thing to be looking at here, be looking at the above medium performance.

So um you know you could get seven percent of of your performances above medium um above median human performance human sumissively mance without aid with aid you're able to get seven percent or Better with any of the lens that I just said GPT for a long a three point one four of I B class three point five senate um and o one preview won't be shocked at all to people who have you use the one crushes, uh, any of the other models uh, where I was involved. And so you know the next best model was GPT four o which have fourteen percent above medium scores with aid. O one preview got double that twenty nine percent.

And you know we know that the full of one model will be out soon. It's interesting that even the internal researchers that opening I weren't able to use that. I guess it's just too early in development.

Um and so yeah when that comes that you can expect that to continue to go up. So you start to have OK. You know these are relatively low percentages.

You know the best approach here only gets a third uh of its answers above the median human performance. But these are extremely difficult tasks with so many steps. And I bet you in a year, this is the kind of thing we'll be looking at, not thirty percent, but we will be looking at eighty .

percent exactly. I think that you know almost certainly the case. And um as you said, it's it's pretty minded ying to consider what we're getting here, right? You getting fully identical that does basically justice de how to go about solving a problem and does IT with the case of aid.

They do you run IT for twenty four hours and they you know do all sorts of search processes and so on. So i'll be only getting Better and Better at this benchmark over time. And it's kind of interesting like, you know, the best system is able to get ten percent of gold on these competitions.

So maybe you could use some money by getting way I wanted to work well, although that might be more effort than actually winning with competitions and now never release from opening. This time it's more of a package, and the package is swarm and experimental A I frame for building, orchestrating and deploying a multi agent systems. So this is pair with of this kind of cookbook post that is basically example of something I do called orchestrating agents, routines and hand offs.

So under the opening organza on github, where your code they did release the swarm package, which is very bold, front says, experimental, educational, right? So I found that kind of amusing. How much of emphasize that this is just a sample framework, is not meant to, uh, be a stand alone library and is mainly for educational resources every way. The idea with swarm is that you can have agents that you know, for instance, hold an ongoing with you based on some instructions, and then they can do hand off so you can say, okay, now you go interact with our agent, kind of like if you call you know, uh, your health insurance or something and you talk to one person and we transfer you to another person, that the basic idea here and so yeah very much we know not a huge deal in the sense of this is experiment and educational, but dos indicate, open the eye, continuing to invest in agenda I and and having more autonomous A I as the future.

Yes, this definitely bit released. I'm glad that we're covering IT in this episode. This is certainly what i've seen is probably the biggest splash in the past week gone my social media channels and yeah, I think agenda A I is a big it's it's the most exciting topic right now that we have an A I these systems becoming increasingly tony ious multi agent systems being able to work together.

And actually on that note, i'm going to a plug something that i'm doing, which is on december fourth wednesay december fourth at uh ninety emaciated c new neston time. I'm running an online conference in the iri platform that is all about agented gi. So we have a number of experts coming in all great uh, speakers in this space talking about multi agent systems and how you in his hands on sessions, you can use uh open source tools in python to be able to develop your own multiple ent systems for particular task.

So I think this is the most and I when I was thinking about what topic should I be covering in this upcoming conference online, I was like, agent ta I, no brainer. And yeah. So I I think it's super exciting and no debt will be talking about swarm. Uh, they are on december fourth .

and onto research and advance ments. And we have some kind of usual stories. Usually we go to papers so on. But at the only place I could think to put this news is in this section, and when users is of a noble Price, uh, being awarded to some AI people.

So first help if you got to the nobel physics prize being awarded to two scientists, john jay hop field and Geoffrey hinton, who basically were very instrumental in the development of neural networks and also lead up to deep learning, large null nets. Uh, you know, jeffrey hinton is one of these big names that has had a massive impact on the history of eye, and not just in the way you might think. So he has been a big player over the last two decades.

He was part of what uh kind of the surface and repopulated ed new onest with some work in two thousand and six where he used some ideas from previous work to have an initialization scheme and and really demonstrate for maybe the first time but you can get a very large newness work to be very performance. But also going back uh decades to the eighties, Geoffrey hinton arguably was a big player also in the popularising union networks event with the a release and with the um kind of documentation almost of this back propagation algorithm that is a backbone of new or network training. So this was not exactly new at the time.

There has been you know previous developments of the same algo room but the paper that Geoffrey hinton published with some other offers any popular zed IT made IT very accessible and well understood, and made in onest kind of hyped up in the eighties. And later, a hintz work also LED to IT being hype up into two thousands and two thousand times. And that's how we got here.

So you know, many people, including, I think hinton committed, that is kind of funny. But I got the physics prize. There's no computer size nobel prize.

So, uh, you know, uh, it's kind of funny. But every way h does speak to the impact of this. And john j. Hop field, not that was a big collaborator of jeff re. hinton. They worked together on the notion of boltzmann machines, which are a little bit more physics like, and turned out not to be a big player, but they do with some ideas in the development of .

neal networks. yeah. And so I dug into this initially for a social media post and then now for um an episode of my podcast and super data science. Um they'll be coming out soon. I don't know if it'll be out before this last weekend after that not so I won't even get into that.

I'll just basically give you all the great content from that here, which is that I dogged into, you know why was this a physics nobel prize at all? And so as you are you mentioned there, there's no kind of computer science or computing nobel prize. The closest kind of thing is the turing award, which jeff anton yh bengie ya arty one in two thousand nine, eight or eighteen one together sounds right um for their contributions on developing deeper learning and for the nowhere Prices.

I'm kind of imagining this kind of discussion where the nobel foundation this is completely made up like i'm imagining this right but IT doesn't seem stretch of the imagine imagination for the mobile foundation to be sitting around thinking, wow A I is doing some really crazy things lately um yes, there are certainly some risks as jeff intends himself has been talking about a lot the license you like google, but there's also huge, tremendous benefit to humankind. They're like we would love to be confirmed some mobile prizes on people for their AI work, but we don't have a category that we can do IT with. And so little like how can we do this? Okay, for physics, jeff enton, who is arguably the most important single player in the development of deep learning as you ready, went over in a lot of detail on he collaborated convention with this guy, john hop field, whom I had never heard of before personally uh having done a lot of research and writing on e learning.

But this hot field guy is a physicist incident. Him collaborated a lot and you know he did make significant contributions to early relatively early artificial neural network ideas um together with hinton and so you know some of hot fuels research um is what they call bio physics where you are trying to emulate the physics of biological systems. And this research fell into that is a new network researched in some way, fell into that. And so I kind of ties together physics with hinton in some way um and allow them to justify giving the physics prize um to hinton and it's great to see him uh get a nobel prize. The next story that we're going to talk about with the noble present chemistry that one is more or straight forward to understand, you have to make a less of of a leap.

Always fun to hear from hinton. He is very direct in interviews. So of this new york time peace has a quote him.

So he said, in a place if was noble Price for computer science, I would have clearly been more bit of that butter isn't one uh so and just to be clear, the work uh from the eighties on boston machines and have been known as hop field networks, which is actually by a doctor hop field those do have more relation to physics, as you said the way. Um if you did begin to a math and uh relates to physics and the kind of connects to physical somewhere. So not totally ridiculous, but somewhat strange.

But a bit less strange is the nobel prize in chemistry, which is also deeply related to a eye. So this went to a dis husb is and john jumper from real deep mind and David Baker from the university of washington. And guess what, it's for their work on alpa fold and things like that.

So as we've covered now in the past couple years, demand has done a lot of work on making A I models for scientific simulation and understanding. And big one from them has been alfa lt. To be able to model how proteins work.

So that makes a decent amount of sense for this and what Price uh, alcohol has had a very significant impact on to the field and they represent a pretty massive amount of progress on this pretty cal problem. Uh, so yeah, lots of novel is going to people from my eye this week. And you can expect .

you can expect more in the future, no question. I think that this kind of sets a precedent that just in the same way that the chemistry nobel has in A A fair bit in recent years, gone to biological advancements, big, you know, gene editing kinds of teaching, where is kind of like gaggle biochemistry, in the same way that you could, you know, say, jeff s.

Hinton stuff was kind of vaguely bio physics because he's working with a BIOS ysp st and like you say, the ofir networks. So you know there is this argument in the past to being used for chemistry um you know to uh to be able to award lot of biological advancements. And if this fall squarely into that category, you've got, you know, we have superhuman abilities to be able to take the sequence of a protein and predict its three dimensional structure, which is so mind boggling ly complex, like humans cannot to do this, or you know, on on any of the kind of complex structures that uh that alpa lt.

Can succeed on. And so this is an example of an artificial super intelligence. And so IT isn't the kind of A S.

I that is associated with a singularity, because it's not general, is very nearly in its application. But this is one of my favorite examples. Alpha's lt. Has been for some time now, been my favorite example of an artificial super intelligence where we now have, not just in terms of processing speed, but in terms of some intellectual capacity, a machine able to do something that humans cannot do um and that's pretty dam cool.

And not to give the only attention to deep pine here, David Baker was also awarded the prize. He is a professor not filled to deep in as far as I know, and he has also had a long history of a research and proteins.

Uh his work like to the creation of first synthetic protein and he was a big part of the creation of roseta, which is a similar a computational tool for things like to design uh of a small molecule docking so uh and other related sort of effort to create Better computational tools for scientists. And you know again, IT IT might seem like a themselves subs john jumper. Aren't chemist per say? Aren't scientists new field? But that pointed out in statements, especially in chemistry and in these fields in general, uh, experimental progress is still progress.

IT doesn't have to be conceptual, and that's what to be computational efforts represent. And onto lying around going back to talk about an agenda. I and from some called the water on IT, we have a story l EMS.

Can't perform genuine logical reasoning, apple researchers suggest. So this is a newspaper released by people from apple titled the G S. M. Symbolic understanding limitations of mathematical reasoning in large language models.

And what we did was a modify gsma k, which is a standard, uh, bh mark with over eight thousand great school level mathematical word problems. They modified to have this set of symbolic templates that allowed for regeneration, I ve diverse said of equations. We um covered a similar effort, uh, not to long order from scale, where they also generate variations on popular benchmark.

And when they tested the models on this new benchmark, once again there was a big dropped in performance of the drop in performance was between point three percent and none point two percent depending on the model, which means that essentially these models uh to some degree trained on the benchmark or very optimized for a benchmark h in fury. And this is the same, a level of difficulty. There shouldn't be uh, any change in performance on the this new variation.

But in fact, there was and you there's some deals here like for instance, adding a single clause that seems relate to a question causes significant performance rops even caused doesn't contributed for reasoning chain. Uh there were also high variance across different runs of A G S simula with different names and values. So various details are created at once again demonstrate benchmarking is hard and we shouldn't trust numbers of uh benchMarks is necessarily because there's a lot of uh.

complications there as you talk about all the time on your show, benchmark cannot be relied upon. And this is exactly the kind of thing that you see that the kind of concern that you have when venture mars come up, everybody, when they released their l am, they wanted to be the state of yard across all the big benchMarks.

And so IT IT will be very hard to resist the temptation not to find tuna st um to to be trying in some way to be gaming those benchMarks which are publicly available. And you it's not like the it's not like you you don't have access to the training data to be old making model form. That's unsurprising that the mouse performers um and yes, IT should also come as no shock that alliance can perform genuine logical reason. They are not designed to. They are just predicting the next token and it's amazing what they .

can do given that. And as you ve seen before, we've kinds of comparisons, the drop very quite a bit. And when you get to a larger, more sophisticate models, the drop is less bad.

So actually for GPT, for oh, we have a drop of h less than one percent point three percent or on me point six um or on p view two percent versus something like jamming and uh mistral which is seven or nine percent three big drops which is to say that this can still be relied upon broadly for general evaluation in less. But the exact numbers, the exact ranking of models that subway where there things like with these kinds of match Marks cannot necessarily rely upon. And the next paper actually adds to that.

So I found the fund to include both of these of the next papers, and not all alone. Reasoners are created equal, and they do something a little bit different with the same benchmark. So they look at great school math benchmark, G S M. And instead of creating a new variation about what they do, is have this interesting test where essentially you need to get two questions right in a row is supposed to just want.

And so what you would expect is your performance on this variation problem is your performance on the standard benchmark squared, right? You multiple your success rate by itself because there's too uh, problems and row, and that's what you want to see ideally. And as with the apple result, that's not actually what you see.

There is a reasoning gap emerges. And similarly, the gap is is very comparable. We got you know G P, for all these bigger models with roughly small gaps and huge, massive gaps for things like five free and gamma lama free, uh, A, B, you know, these smaller models, broad speaking, so they go there's another demonstration on this pretty good benchmark of that.

The backwards aren't quite as precise as would be ideal. Oh, ready. Moving on to policy and safety. We have pretty interesting development related to on topic. So on topic, C E O goes full technical optimist and fifteen thousand word B N T A I is the refine title of a stack crack article.

And we go it's it's about to this very log block khost that uh barrio M V C O of on the public released be unusual hasn't been you know like sam ta who goes like on all the podcasts and and has released already uh block boast uh over of his faults this isn't something a direr m does quite as much but now he has and it's a very, very detailed examination of the implications of the in particular positive implications so he begins with this whole thing thing of like why don't I go and and be more uh positive and IT feels a little bit like maybe we just talking to investors and and making IT seem like a propionic a bit less safety oriented or over is uh optimistic about I every way a lot of interesting notes. This it's it's very long, but I would say it's very a long of because it's very nuances. So he goes into a lot of detail and it's it's not sort of like rambling.

It's very well fought out for my read of IT register. IT is that omi defines a notion of A G I. So he says A G I is a very useful term.

Instead, there is a term powerful A I, and the definition of IT is A I that is smarter than a nobel prize winner and feels like geology and engineering escape of of performing tasks like moving on side conference and riding high quality novels. And he has a nice phrase for summary. The computations of IT is if we get these powerful eyes and we can run multiple instance of of them, it's not competition ly crazy.

We will get a country of genius in a dota center, right? And so the beginning of box sets up with idea sets up the belief that will be getting this and probably five to ten years and then the rest of IT may be ten thousand words or something like that is about the implications for different things. So he talks about implications for, uh, biology, right? For being able to make a lot of problems uh, progress on very significant issues we've been unable to tackle like um you obviously cancer during the diseases, halting alzheimer at earlier stages, all coming in the next seven to twelve years.

He says this nice term again. Um you can figure with as the compressed twenty first century the idea about after powerful the eyes developed, we will in a few years make all the progress in biology and medicine that we would have made in the whole twenty first century um without IT. And then there's what topic he gets into. He gets into, uh, inequality and the implications for economies, particularly of the developing world, where again, there some nuance. He says i'm not as confident that I can address inequality and economic growth grow as I that IT can invent fundamental technologies because the ology has such obvious high returns to intelligence, whether the economy involves a lot of constraints for humans you know that's what kind of writing you're dealing with is pretty sophisticated stuff. But yes, he gets into things like global economy, climate change um even this sort of like more philosophical things regarding meaning and work and of course he does have a lot of covets about to potential side effects you know dangers at sea of its a very long essay but I think a very uh interesting and and wealth fought out S A on the implications of powerfully I coming. This next decade, most likely no question.

Um this is super A A this is super line like hopefully your future eye systems with my own take of what's going to be happening in the coming decades. I am highly techno optimistically like darro oma day is and he hits the nail on the head with so many points in here. The country of genius is in a data center is perfect like this.

Is that something I talked about a lot around the release of o one a couple of weeks ago, where you exchange this kind of capability, where you scale more uh, in terms of you know some something like increasing the number of matters or something going to have that kind of ability to increase the nuance of these models plus at you scale in difference time. And it's not hard to see that we could have in a couple of years, AI systems that are like, what do ooma here calls powerfully I wear. It's like a nobel press winner.

You know, you have jeff hinton and Dennis savas level of intellect in machines. And then, like you said, Andrea, you can scale that up. You can have a country of genius in the data center, all thinking away at complex problems.

There are some things like, you know, he talked about alzheimer's specifically, and some things like that. Those do seem a bit tRicky to me because those still require real world experimenting. The algorithm can have a hypothesis, but then you need to test IT on you know, laborat and then eventually humans once you have confirmed that, that safe over a number of years.

So there's some kind of constraints. It's like that classic, you know, nine women can make a baby in a month. There's some kinds of things related to scientific and country that are beyond, just think you can infer base on information. There are some new information sometimes that needs to be acquired from the world with experiments that take time. H, so that will you know slow down some types of progress. The bigger kind of progress that a daro also has, the head on ear, is social progress where unlike fundamental discoveries, which are something that you you just have happened, uh, kind of unconstrained by social or governmental pressures, the the distribution of that equitably um that that is still something that there could be a really tRicky problem like you know you might have you know you might be able to create crops that would provide high school nutrition to every one of the planet. You might technically have that, but that doesn't mean that north korea is going to allow you to have those crops in our country.

exactly. So I also share your camp, take the um in general and more and missing front uh and this is a very grounded there, a library exploration of why we should be optimistic and sort of details of IT uh to be a bit more precise he defines powerfully I as we wings and saying is that I could come as early as twenty twenty six though also ways in which IT could take much lower and what have actually focused on is what happens in the five to ten years after we get powerful way.

So the idea being that, you know, we get all these very nice outcomes, ideally from A I if we don't mess IT up. And of course, IT does also call out that is not all roses, you know, there are dangers, association of iron we do need to think about our topic is heavily focused on a safety, uh, so you know, a very nice read. I will say no in a lot of discussions of agi of future of the eye IT IT tends to be a little bit more size.

I not grounded in definitions, not grounded and losses. This very much is. And just a couple of more stories.

The first one is once again about nuclear power. So google has partnered with chaos pier, a caro's power, to construct seven small clear reactors in the U. S.

To power of these technologies. The first reactor is expected to be Operational by twenty thirty, with arrest deployed by twenty thirty five. So the statement from google is that views would offer a clean, constant power source that could mean electricity demands with carbon free energy.

Uh, house powers is a nuclear energy, uh, start up, which has some kind of renewer technology related to nuclear power. Uh, of course this is related to a recent move by microsoft, also related to uh, using the undemonstrable actor at female island. So overall seems like A I is making us embrace uh, embrace nuclear power at a rapid pace, which hasn't been so much your case in the past century really.

And the other story we are gona cover is more under research front as we often do. We have some research related to safety. The paper is l lambs no more than they show on the intrinsic representation of allama hosinia.

So um what this goes into is how recent studies have shown that a allah's internal states in code information regarding the truthfulness of their outputs and this can be used to detect errors. And what they do is on lize this pattern a little bit more. So we show that we are a this information about truthfulness has more kind of clarity than previously realized， that that information is constructed in specific tokens.

So that means that looking at of these specific tokers leaves too much Better error detection. Although these arr detectors don't necessarily generalize entirely, meaning that truthless and coding is not universal. The encoders are some of different across different topics, for instance. So, you know, very practical implications here. Of course, home nations h allama just making stuff up is a big issue with using these tools, practice and these kinds of approaches that look at um the outputs for presentations to detect whether realm is fluctuating could very much be deployed by airline providers.

Yeah at important researcher being able to stamp out hallucinations almost entirely is key to all so much success in A I, including particularly when we talk agented AI systems, multi agent systems. If you even have a one percent err rate in a multi agents system that has ten agents passing off between each other, that one percent era compounds and becomes a really big deal so you know you need to be talking about fractions of a percentage um and the smaller than more zero we can put in front of that after the special place, the the Better our future AI systems are going to be.

And onto the session, synthetic media art. And this is also last story. Have one here. We started with adobe. We are ending with adobe um and this one is not related to A U I I tool.

IT is a free web up called adobe content authenticity and that allow creators to a touch content credentials to rare digital work. So you can figure this as a nutrition label for digital creations that provides information that makes IT easier for original al content to be traced back to its creators. So the swap up, uh, enables photographers and digits artists to apply these contact credentials to other content. IT can include a verified name or identity and links to their website and social media profiles. And this is the kind of thing, you know, we've been talking a lot about the gush to a eye that we would need something like this, some metadata attached to a files to media to let us know whether it's A I or not, whether it's by a real human uh or some products or in addition to waterMarks like sinne ii that are being built to germany and and also metal tools. This could be used by human creators when they publish their photographs or their images to eventually claim credit and on top of having us uh be a tool there also relies ending a good extension to allow users interact with and view content credentials online uh and of course as a using you can also inspect the contact that credentials have given file h to be able to see ve that and you know I can sort of act like a digital fingerprint almost this is now in free public beta uh, or will be rather this will be available in free public beta starting early next year and they do say this IT would be for integrated into their actual apps like photoshop. And Sarah.

ah nice to come full circle with a doby here the doby super episode not sponsored by doby in anyway .

only but yeah I mean now I think this is want to be things again not gonna get many headlines probably right uh hasn't been with the talk of the town in A I but h might be think that could have big implications for many people who work in this field who are concerns about uh A I generations for things like photographs or just this information. This is the sort of thing we need to embrace to tackle those kind of things.

And that's IT for the episode. Once again, we are heading roughly ninety nine mark maybe will keep doing this will see. So thank you as always for listening. We appreciate as always. You can go to last week in that I to describe to the newsletter that will also have links for all the stories here. And uh, you can also look at the description of effective for that will also have links to super data science podcast and to this upcoming event, uh, Robina hosted by john coming up in december. And thank you, john, once again for being a fantastic guest coast.

My great pleasure is truly an honor to always be on my favorite gas to listen to the only podcast I always listen to last week. I yeah such a true to be on here. Thank you on for the invitation. And yeah thanks all you listen ers, for your time and attention again today.

Yes, exactly. Thank you listeners for listening. Presumably if you're here you're listening to a very end, but which I always find pretty impressive. Uh and as always, video, appreciate your comments reviews. I try to keep a look out to make sure I don't miss any. Uh and of course, since you uh got to end of epsom hopeful ly, you will enjoy a full version of the I generate song I made for the epsom.

In a world of and series pain and dream, take light, protest the future ruling down streets, conditions with tomorrow be the. Patient no surprise E.

To the code with the big, so electric dreams on pot, every machine, the first thing, test those future rides, the dog is canvas problems. But then I, other guys, nobody mize the, we've come by, we've got take tail soring high last weekend. A ee is the recent why with. Every body imagine for.

Sorry, other week, would stone online business outside?

So we.

watching. The future. 我的 那。