Episode notes
Jensen declares AGI is here, OpenAI kills Sora and pivots to world simulation, #QuitGPT flips enterprise share to Anthropic, and Google open-sources TurboQuant.
Chapters
- 0:30 — Introduction and discussion on the weather setting the tone for the podcast.
- 2:15 — Discussion on Nvidia CEO Jensen Huang's claim about achieving AGI.
- 7:45 — Overview of Google's advancements in AI, including TurboQuant.
- 15:00 — Discussion on the challenges and future of AGI.
- 20:00 — Analysis of OpenAI's and Anthropic's market dynamics.
Links
- https://www.nvidia.com/
- https://www.google.com/
- https://openai.com/
- https://anthropic.com/
- https://github.com/anthropic-ai/anthropic
- https://openrouter.ai/
- https://perplexity.ai/
- https://fauna-robotics.com/
Transcript
Note: this transcript is auto-generated and lightly edited for readability.
three two one welcome welcome to our next episode episode number three of digitalized lab we got a lot of things to talk about again unfortunately no sun it is quite depressing again but yet we're coming from a high good good sunshine days every all yeah i guess so actually but now we're back to the normal dutch weather i would say the four days five days i think they were in total pure sunshine there was a once in a lifetime opportunity here but anyway i love how we always start our podcast because we don't know what with the weather it's the same like having a small talk with a person you don't know what should i yeah that's great today yeah we have really good weather tonight literally okay then let's start with the biggest news what did chanson huang say yeah actually yeah he said uh we achieved agi or he thinks that we achieved agi crazy i mean to be honest wasn't there different ceos that also already mentioned it or was it the first one that really like kind of pushed the topic no there were a few people who said that we will reach agi in the next five six months whatever but no one said yet i believe that we reached agi especially not such a high figure in a way yeah true true true true maybe not everyone knows who he is to be honest he is the ceo of nvidia and i think also the founder one of the founders i believe so yeah one of the co-founders yeah he started it with with a few friends in the wendy's no in some i don't even know inside yeah it was actually a really nice story um but they met in or they they met up in a restaurant the bar or whatever and to start nvidia and they had i think is that why he always goes eating chicken wings yeah yeah that is exactly that place and they had the meeting i think this year or last year where they met up again with the same table yeah exactly yeah and they they they had so actually really really nice story yeah but who do you think has has it i think there is only one correct answer for that one who is launching a new tool and destroying a new business every tuesday oh to be honest you you're thinking of entrothic yeah of course but you're thinking of google yeah i mean to be honest they've been lacking behind heavily in the ai wars like not in a piece like in let's say six months ago not as bad as apple they're pushing one tool after the other look at google stitch and what we talked about last time look at there's a million tools i also got a couple hot takes from google this round like i couldn't decide between both of them maybe maybe maybe they both cracked but just don't tell anyone i just want to say maybe maybe they cracked it together because i mean also what is agi like when yeah when did we there's benchmarks on it yeah but let's say we reached 100% on the benchmark that some some guy invented and we reached agi i don't know yeah but to be fair i think in the interview he also says that um most of the the systems are extremely capable that we already have so on that i guess on that takes he says we already have agi but yet the systems are still far behind building a company like nvidia for example autonomously get it not really possible true he said that yesterday as well no that he or the two days ago that uh he wouldn't think or he doesn't think that any ai model could build a company like nvidia like yeah yeah i guess agi is also not yet the question is what what's the hurdle you know what's in between what doesn't work yet what couldn't it do and wait first of all the the jensen i mean the the phrase that we just said or the hot take we took out of the podcast that he did with lex friedman quickly to round it up and and i only listened or only i sadly didn't listen it completely through but one and a half hours i'm in so i haven't reached the point where he says i think we reached agi but there were a few things that he touched on before and one of the things is we are now not only limited at the scaling law for one ai agent but rather like we were um bottlenecked a few years ago with scaling with humans so basically the management of humans was or needed to be developed in order to get to a further point and stuff like that and he he said that we are right now at the same point but with ai agents because one ai agent or with a few agents the systems are already really good but there is that swarm intelligence that is missing and different perspectives different agents or humans and then that scaling part so how you manage actually ai agents on a scale that is the next hurdle i would say in order to achieve something where nvidia could have been invented by ai agents i still don't think that it could have been because the the human intuition and how jensen huang steers the company and all of that i think that's the missing kind of link yeah we can emulate how humans kind of think and act exactly but you still don't have that touch or that gut feeling i would almost say i guess there also needs to be someone who pays for the compute and everything else right i mean i mean it could be the model like if if the company earns money under the ceo being and trophics cloud or whatever well from the bottom all the way up 100 in a way but yeah it's interesting to think about no no no interesting topic for sure and let's see um let's see how it goes open eye kills sora what do you guys think about that didn't hear of it yet i just heard from it earlier i actually love it yeah so just to round it up again i mean it took the loss now i guess so but the sora is uh open eyes um video generation uh app they launched i think a couple month or a year ago maybe i think like four months i think sora one was released a year ago almost um but the sora 2 the the actual model that is picked up yeah that is getting used or being used was i think four or five months ago so already old but but quite interesting because there was a huge hype right when when i launched it it was i remember i still remember seeing all of instagram was full of those sora badge watermarks yeah the first day was full of watermarks the second day someone released the water sora watermark remover everything was gone yeah thanks yeah but then google came in with the nano banana i think it wasn't called nano banana back it was flash 3.5 based on what whatever i don't know you know is that already a year ago no that was more it was it was after sora got released that i'm quite sure about but i would have to check on the rest but the app itself i think was released a year ago could be could be but i think sora 2 was the point where it really picked up because then it was actually usable like and the most of the people that had a plus subscription of chat gpt could use it and that was the adoption that was why there were so many thingies i probably costs like probably cutting costs because it's a shit ton of computer i think it's just it just just also shows in what kind of trouble open air they are i mean they they had the code reds couple months like two or three months ago and i think this just shows you know they're not making enough money i mean image and especially video generation make a year i don't know way too much even with all of the revenue that they have quite a quite a bit of revenue to be honest but the the video generation is extremely expensive in terms of compute that's what i think that it's so i i i was cutting yeah i mean and they also i don't know if you guys remember but they they scored a big deal with disney right at the start where you could use their yeah use their um disney characters with sora so you could never saw a video though on instagram either with or maybe that's just me i think they they were in talks but i think the deal didn't get through or the deal was over the next few years or something but i'm also not really sure about that one was a while ago but do you guys actually think that it was because of compute or because to save money or i think it's cross-cutting i don't know i can just imagine that video generation is as tim already said it's hella expensive like compute heavy and i don't know how many people actually still used sora to be honest so yeah yeah but maybe also um i actually have in my notes source team is shifting towards word simulation for robotics so they're probably also pivoting to to human robots and i guess it kind of brings it all comes together with the code red that you recently had you know probably the pivoting open air has huge problems with usage currently everyone is kind of shifting to anthropic and what was that what was the hashtag you quit gpt love it the whole military story i think that that shows exactly yeah for sure they they yeah it's getting more serious what did i tell you about uh i don't remember i was like 20 like basically they always had like a 50 50 split on enterprise customers like open ai and entropic but since that whole military kind of debates and open ai stepping in for and that quit gpt kind of movement happened uh it switched to being 75 percent uh and traffic market share 25 open air which is quite a big deal in a couple weeks like in two or three weeks it completely moved around so that movement was real especially especially in enterprise i would have thought it's just b2c yeah yeah i mean they they had big or i don't think it was a big protest but in front of um the the headquarters and the the office of entropic and the morning after they announced that they are going back from the deal with dod and they were all hearts and nice messages with kreide with chalk true on i mean i can imagine that people like it like obviously it's not a bad and not a nice look if uh something like what happened there in in the war technically but it's an open ai steps in like of course we're happy to provide our model it is fun for me to see because i mean i'm using cloud way more than than chat to be honest because it's for me a nicer architecture feels better whatnot um but it's fun to see that that engages people more to shifting than actually the better quality of the output you know because i mean cloud or entropic was before there and open ai too open ai were the first mover okay they had more and more market share but then when they had the 50 50 split and then because of the deal they get a 25 decline of users in the enterprise yeah yeah i think what what is also interesting in that is or the underlying is that how fast everything is moving not only the tools and ai as such is moving fast but that also translate into the user somehow i mean what you just said you have to you have to move fast to keep up like if you're actually you know yeah but also the loyalty is not really there right everyone yeah open ai true i mean i i the better first year i i only used open ai i didn't use anthropic at all cloud same i mean i like i tried it out but it wasn't there yet and now it's the other way around i always use cloud um and other models but it's i rarely use open ai for me to be honest because of quality it wasn't even like way before the movement happened and stuff it's just because cloud for me outputs a better quality uh output yeah but i think it also shows of how fragile everything is a little bit you know like and there's also like these trends again you know like how social media can influence that type of hot take i think that is also first of all really interesting observation with a mindset and not really the loyalty but i also think that because it's such a new thing that people haven't made up their mind yet because they i mean you see open ai on one hand they say they're doing what not and on the other hand they are doing the exact opposite of that so the users or the people that are actually using ai don't know who to trust interesting to compare that to other trends trends like this mega trends back in the day unfortunately have no no insights in that but be interesting to see if there is parallels or if this is something new true i don't i mean it's for me it's quite similar to the movement of for example tesla recently like where basically tesla was being hated on in the us and all of the tesla's were like lit on fire and stuff there was also fully social media induced and even trump had to say guys calm down so i don't know if it's new but it is like obviously it's new because social media is still i would say new like yeah but still i think the the model quality is there also a really big thing and especially because people didn't choose their their i don't know favorite yet because i mean i i chose my favorite car brand when i was six or something because influenced by my parents also an interesting hot take actually because many companies have like invested heavily into like already picking up like marketing basically to children so that those children when they grow up super smart always stick to that brand bmw whatever it's not only car brands it's different kinds of and i think that doesn't work in that yeah but not not yet not yet i think that that we will be definitely there in a few years where where the models are basically almost all the same they have a good intelligence exactly once they are aligned and then you have the little minor things will they be like a ceiling probably i i think not not a ceiling for the technology on its own but a ceiling for the understanding of the human and how fast or how good you can interact with ai agents because i believe in a few years the ai agent will outpace us and i mean you see it already um if i if i have four cloud code instances running then i'm running behind the ai and not the ai behind me basically and there is then the bottleneck is the human or the human computation i mean you can spin up endless teams obviously then at some point you're lacking behind or you need new agents that just orchestrate but yeah but that is then exactly that point again where jansen huang said okay if you now scale up the agents how do you make and manage those agents because you can spin up a million agents in one minute but they all produce an output they all need an input in order to do something and then how do you orchestrate all of that data that comes in at one point do you yeah exactly but do you have then one orchestrated agent is one orchestrated agent already too little for a hundred agents currently you don't have to use that hierarchy in agents of course but you can you can technically say there's one ceo main orchestrator there's one ceo and then there's teams below those or you use a spotify model or whatever i also just wanted to say i think the school brand is with a spotify model but yeah no that will be interesting there's also many many companies i see recently that pop up they kind of build or they promise different ways of how in the future we will interact with the ai and it's not of course chat windows like every bigger company does nowadays but for example link the building platform that you can integrate it into i don't know social media whatsapp or whatever then you can obviously like telegram or whatever just send messages to the orchestrator and the rest is being done in the background but i don't know if that's the future we'll see could be just a intermediate interesting times ahead yeah definitely let's see yeah but yeah i have a one or two things on stack or skip if you don't wanna let's go yeah first one obviously i told you guys earlier about it google released a new uh quite cool tool for us turboquant it's called and basically how i understand it it's brand new um basically it manages or allows you to to summarize or to kind of how would you say it's quanta size yeah quanta size um your context window size by six times or eight times i believe it's eight times actually yeah and uh delivers up also delivers up to eight times faster attention performance so basically meaning the model stays exactly as as kind of yeah the quality as it was the quality stays but the context window doesn't take up that much pram anymore so in a sense you can run bigger models on smaller kind of hardware sizes and smaller devices um but of course it's not limited the model size as you probably want to to mention is still the same uh that can't be quanticized yet and so they kind of just introduced like a zip file right for llms in a way yeah in a way it summarizes way way more efficient than i we had that we had that technology already like even cursor all of those vibe coding tools they all used it to summarize in between the context but then you always knew afterwards that stuff was summarized you always knew like you saw that the agent doesn't know anything anymore it would forget stuff and then it makes mistakes but in this at least what they promised i didn't try it yet uh is that that won't happen that it's eight times more efficient than what was here until now but the model would shows again that the the the the area of uh local local ai and local lm infrastructure and private infrastructure is definitely in the in the yeah in in the uprise yeah i mean back to jensen huang he said it in the gtc and the the relation to to and the apple iconic scene where he says we're gonna put a computer in your pocket yeah yeah basically the same but now the computer is an ai model and that handles basically everything that the phone or whatnot did yeah even even for phones it's interesting especially interesting because you can run models on phones quite well now super awesome with the gamma models etc or quend any ones but um yeah no what i wanted to to add is what i didn't really understand until i read it now the speed up time like the load up time when you use for example lama swap when you switch between models is eight times fast as well because it doesn't need to inject that much context window it's also has tokens of course but it doesn't need to inject that much context data anymore into every prompt especially on speed and on startup and uh yeah smart also eight times smarter yeah faster true interesting right crazy yeah i haven't thought about that one but yeah of course yes they open sourced it yeah it's also amazing of course best for credibility of a company to be honest but uh please go on google and different different providers just open source stuff because they could have gatekeep that for for themselves they probably built it for themselves but yeah but yeah i mean again you see how big of impact open clause and i think that was a good step also for the open ai or open source community and ai on that note i came something across something on twitter the other day or on axe as elon likes to call it um it's more like a meme thing but death by claude dot com do you see that a website yeah death by claude dot com yeah um so it's uh it's like open claw hiring hitman or something no not quite yes like it to be honest but a website which basically gives you a detailed scores and does like a deep research um on so you put in it's just an input field you put in uh a industry or job a description and then it will tell you give you a score of uh how likely uh your the chances are that claude will taken will take and then it also gives you indication of how long it will take so it's like those ai taking a job bench benchmarks thingies like those statistics but even more on crack because and it will also tell you it is i think it is based on the research that entropic launched i think three weeks ago because they did exactly that they they made the spider web graph where they say which industry will be mostly impacted of the new knowledge worker era whatever yeah but take it take a look it's it's it's definitely a skip but it's it's a fun it's a fun meme thing but it also gives you the the indication of how the how long uh cloud bot things it will take to replace that yeah it will replace the human in that sense true but wait wait wait we have to google is definitely a stack yeah for sure i mean to be honest we have to test it it literally came out this morning but uh from what they promise and i also read that it doesn't need any training it doesn't need any fine tuning it's literally plug and play at least from what google promises and that sounds like it's not like one of those flashy demos that we see on a daily basis but it's actually something useful if you of course work in that in that space wow i think we will take a deeper look and then we can update next episode we can we can mention but i think it's a stack yeah for sure cool but i also like your your the the website because i feel like sometimes interesting most of the times ai or the people use ai to make money to make more efficiency and stuff sometimes take a step back it can also be fun just this vibe code something in 10 minutes and throw it out there i i also i also saw something in in that kind of meme space era there is a website where you think you chat with an llm but you basically on the other side there's a human so you can go onto the website you can choose if you want to be a user or the llm and then you have and then you have basically one user texts the llm and the other user doesn't know if he's texting in an llm or a user what no they don't perfect way to kind of do that you're not supposed to ask the hard question exactly that was again a wipe code but it's super fun you know like it wasn't for ai of course but it was like uh i don't know what the name was but i have to reach like look it up but it was basically of if the person could realize that he's talking to an ai but in that sense it would be an ai or not and in that way it would be the perfect test sometimes it's a human and sometimes it's a bot yeah and you don't know i think that that website is more on the fun page because you have then a timer as soon as someone sends you a message you have i think 60 seconds to answer and people can also basically let an image generate and then you sit there with your mouse and like the draw io thingy times you just draw a little horse or something and then send it back to the other user it is user against user but one thing is in lm and the other yeah i need to look up the url but that that is what i mean sometimes also just have fun with it for sure sick but yeah sorry another tool that we have in stack and skip or skip pardon to be honest for me that was the google quant was the the biggest but i'm very surprised that i haven't heard anything from savvy true is there not i mean you probably use so many tools this week to be honest i have top of your mind still auto research auto research is still the tool that that keeps me up yeah yeah i think even in the first one that was just when it came out super awesome and what people are using it for that's exactly what i meant people are not only using it for reinforcement learning for what it actually is being was developed but rather to optimize everything you just need basically an algorithm or a way how you want to do something and then you need the kpi at the end what you need or want to increase exactly a goal a task and just let it run for however long let it run yeah i actually have to play around with it yeah exactly it is super awesome but i haven't had enough time for that but is it just me or does that also underlines again the local element for sure but it also tells you especially because i mean the the the founder of auto research or the creator of the repo i should rather say is a legend in the ai world and if he uses already ai in order to train his own ai i mean that is there is already the big step into the right direction and then we touched up on minimax last week about how fast their development cycle is or not development but release cycle and that is what i found out what they are basically using in the background is auto research they are scheduling 30 to 50 percent or not scheduling but 30 to 50 percent of their reinforcement learning loop goes not to a human anymore but rather to agent yeah and that is how they could exponential cut their training time of their models in half it's crazy there's a super nice chart uh minimax m 2.0 if i'm not wrong was three months seven is the news yeah 2.7 is the news and they they started with the auto research loop basically with 2.7 that is why they're so fast nice but in the in the basic of the architecture of the llm they're all the same that is also the fun part they are all one trillion parameters they are all the same moe architecture so mixed basically that not all of the weights are activated in every run mind me again which model was not moe is it and no i believe so i one of them is completely different architecture that's what i remember next next episode yeah yeah but that is the fun part they are all the same but they just started with a basic architecture and then tried to find loops or ways on how to iterate way faster than the competition and also it's not it's not a trade-off with a quality the quality each point is way better that is crazy and especially if you compare it to claude claude flats out nowadays because they are taking so long and i don't know what they're doing maybe developed agi who knows um but mini max is the release cycle gets um way less and the model quality gets so up it is crazy how good and how fast they're moving forward perfect and especially also open source so it's actually interesting that the bigger players like it's a it's a small company compared to the rest but now i have a i have a debate question that that fits perfectly to it it's basically for like basically what wins in 2026 and it's the company who has the smartest model or the company who has the best stack around the model so let's say for example google in my opinion has a crazy stack around their gemini models they have nano banana they have vo 3.1 now like they have all of those models that are integrated into i think i think the biggest part is not what you're saying with the different models but google has the data google has the machines google has the knowledge and the talents with deep mind that is already the complete cluster what they have and other labs like entropic or a smaller mistral that i also really like they don't have it of course they don't have the possibility yeah but the bigger players also try to kind of buy in the data you know obviously they did already and especially if you look at alexa or whatever it was probably in preparation probably to just get the data in advance nowadays no one is really using it anymore but but you you were going somewhere sorry i interrupted oh i actually forgot about it but with the models the the combination if google with the broader system i mean it was part of it yeah no just uh yeah which which kind of thing wins like the the company who has the smartest model just the model or the company who has the full stack around it that makes it actually actually usable actually interesting i think we need to go but the stack by a stack you mean like integrations or in a way like tools you can use for example if you go into the germany app you see immediately image generation you see video generation you see all of those tools that are kind of integrated on the phone app also yeah yeah because i i think their their their interface is horrible that's a different discussion every time i'm on gemini i'm just like let's go back to perplexity or whatever but perplexity is lovely but uh that's horrible i only use it for for for image generation to be honest no you guys use germany only for image and video generation to be honest no i actually use it quite often yeah which kind of i still think big text summarization okay context way to go yeah i mean nowadays almost everyone has a million but yeah if you use the models through perplexity i'm quite sure you're limited to 200k or something because just day limited yeah but to be honest don't mark my words on that one i have to check it yeah that could be that entropic sometimes they don't give you the full thing because it's actually really hard cursor especially it's if you don't have the max max max subscription but i still feel like for for that price especially how cheap that one is on open router yeah for now on open router also hot take by the way true on open router and it's super fast it summarizes my text good the only thing is i don't like the formatting but that is okay that can be i mean playing around with the different things do you guys know the company fauna robotics no i heard i heard something about them i think they are the ones who do like also humanoid robots but i think smaller ones i have to i have to look like well i have dangerous half half knowledge right now wait perfect thing i found a website yesterday humanoid robots.diy where someone has actually listed up all of the humanoid robot companies there's a rating system so you can rate them and he also has a humanoid robots.diy if i'm not mistaken or humanoid.diy or something yeah you will find it and but he also shows basically the outline of a humanoid robot and you can learn basically about it super interesting and i believe they they were also mentioned in that one but i don't know more about them i love how i can just sit here you know and just type in tech crunch and i just see google launches there are three pro music generation mall they go into that direction quite a bit i didn't i tried it once to make like a jingle or something but uh they do quite a bit like if you it's actually a big topic like if you look at spotify for example you see so many artists that are officially of course not ai but if you actually look into it and their instagram presence and social media presence which is also fully ai generated they sometimes have four five six different million users or like listeners monthly i don't know i even read like an article about it that the biggest three or four of them if you actually dig into the files it's the same guy behind them which just makes probably a shit ton of money doing for real fake music which i guess it's still like i don't know it's a debate we can have i don't know if it's good or bad i mean if it has four million users and people don't even believe that it's ai generated i guess it's good music for some people you know what i heard last week that is also in regards with spotify and making money in that sense darknet diaries also really nice podcast shout out super nice yeah you you showed it to me it's okay you you can get your your your thank you thank you i feel much more actually thank credits yeah exactly thanks to my cousin who showed it to me a couple years ago for the credits yeah but but actually what i think it was one of the newest episodes where he explained how how people are actually pushing money through spotify yeah but they're washing they're washing money they are for for um for bad groups and all of that stuff because they are all based on the clicks per or the so they they basically just have so they the ai they have an ai artist right they they publish music then i guess have like a bot farm or something like that and i just farm money yeah yeah i didn't even think of that and it is a smart huge market i saw that recently people are yeah people are pushing a lot of money around there and you as a normal person that uses spotify every day i i didn't even think about that i think the biggest smart is called sierra you have everything yeah sierra something is like the biggest kind of ai of of course not ai officially but if you actually look at like there's instagram videos where just five fingers disappear or something in the middle of the video so obviously she's not real and she's never been seen and she wants to go on tour next week good luck so we'll see you next year i don't even like that's not gonna happen how nice making money with that nothing yeah basically with nothing just automated perfectly just a really good base prompt yeah it's crazy the file of multiple uh identities and so on and how the crazy yeah yeah well well i think there was uh most of my biggest topics for for this week yeah if you guys have anything else no no i think besides round it up we apparently achieved agi and sort of big topics this week to be asked i think we actually need to do the agi episode soon yeah maybe we'll do a special episode uh on agi specifically sure maybe maybe we can introduce one of our first guests then yeah might be a good opportunity true these guys give us some feedback on the podcast um as always um as you know and i think that's pretty much it right do we have anything else no not for this week well until next week yes we enjoy you uh we hope you enjoy your week keep building and keep building and keep building and keep building and keep building and keep building and peace have fun