Episode notes
Claude Code's 512K-line source leak, the OpenAI vs Anthropic beef, Apple's Ask Siri pivot with Gemini, and agentic tools heating up.
Chapters
- 1:00 — Discussion of Claude Code's source code leak and its implications.
- 8:00 — Exploration of the rivalry between OpenAI and Anthropic.
- 15:00 — Apple's decision to integrate Gemini models on their devices.
Transcript
Note: this transcript is auto-generated and lightly edited for readability.
what's up what's up welcome to another episode episode number four actually already i hope you enjoyed the last one uh as we said each time a little bit better it's great weather today is it actually we haven't been outside but the last time i looked outside it was great weather actually what do we mean we haven't been outside claude code released their whole code for claude code cli i don't sleep tonight well yeah oh i'll take you up on that one we'll see we'll see well again lots of things happening uh we got a full list again uh let's try jump right into it and yeah who wants to start first i guess with the biggest news that just came out today claude code yeah they made a big ass mistake or we don't know if the engineers did it or i think seven hours yeah nine hours yeah not even yeah it just came up on our timeline four hours ago um but they released basically their source code for claude code so basically now we the whole world knows how claude code is so good and how effective they are and even the more interesting the whole world can kind of reverse engineering in a way which open code the the open source alternative for that work in their cases no but they will they will i hope but i i know they will do they will hopefully just copy that into open code and an open code is on a freaking whole different level but to be honest they were already actually really good if you mix and match started using open code and take the secret sauce of uh claude code i even saw that they had like a special uh a special system in it in claude code just called undercover mode or something that literally was doing nothing else but trying to strip any uh official names from the gits uh gits commits yeah and i mean until now it worked i guess now it went rogue and just decided to release the whole source code but that's fun but saying barrier source code what kind of information was leaked like in terms of everything everything everything it is the whole the whole like way everything no and the we cloud code is basically just yeah okay nevermind okay that would be even a bigger thing yeah but basically their their cli tool how they are okay interacting with those models i was laying behind yes it's basically it's the internal architecture security layers and how the tool collects usage data it is the full architecture they release i think no user data so it's not nothing like privacy wise but yeah company wise it's uh super fun well probably uh open ai tomorrow code green code green code green code on another level yeah yeah no i mean it's good for the whole community because for the ecosystem was perfect i mean cloud code was the i would say leading kind of cli tool by far so if the others can the only thing that actually came close to to the um cli was open code but open code is still a bit lacking behind i would say not that i i really like that yeah open code and i work really well with it security layers and the internal architecture exactly and especially i think the memory layer of entropic is super well designed and that is i think one of the secret sources but let's see tonight let's listening uh go on github try to get that downloaded before they kind of delete it all they probably did already right they're trying but of course the community is faster and just forking if there's something in the internet you have fun trying to delete that especially in ai times good luck that was that spread especially in a domain like that yeah and i mean it's the fun part was the the first guy that tweeted that his tweet got 12 million impressions in i think two hours or something in two hours maybe we should we should adapt like a really a marketing uh campaign and just spread fake news in that domain no we just hopefully not no we just spread our own tools basically and call it then open source i guess i know but it is really really interesting especially because they were just in that standoff with the u.s government and uh true action they also just copy pasting that over yeah i guess palantir too but it was what i just found out it was not the first time actually that it got leaked there was another similar not not that much of data but uh similar leak late 2025 yeah the the thing was late 2025 was it let me see early early 20 yeah i just wanted to say there was at the time where i think 0.00001 percent actually used cloud code and it was still in the in the beginning shoes and nowadays i mean look at how cloud code evolved the one question actually do you do you guys think it was a mistake of a human engineer or of cloud code itself basically a mix i'd give it a mix because they got so used to obviously they make it quite public that they basically the mix is a boring answer you need to wait they make it quite obvious that it's being like 100 of the code of cloud code is being written by cloud code so even even the engineers kind of not 100 but they're using it daily twitter post today that i saw one guy that said 100 of my code was written in the last week of cloud code and i just say i obviously there's a human in the loop but i think if you do that for months and months you kind of get lazy and you're like and you might miss one step because it's just enter and then it's but i mean it was leaked through npm no like i don't know how that works to be honest i'm not that advanced but good question somehow it got leaked through there so maybe it was just a mistake no i think it was through a dot map file and that dot map file kind of indexes the the repo because you need the repo or the codes in order to install it through npm so if you install cloud code then it basically goes to npm and then npm and goes back to github or wherever the the code lives basically and then pulls it through there and i believe that inside of the npm thingy and there was a dot map and that included then the whole architecture or the whole source code into the npm records registry the npm website where you can find everything and on there was just a button with code and then well someone something or someone yeah true do you say someone to an agent too no at some point yeah i mean at the ajai point for sure but when do you guys think that will be the first when when when we achieve the ajai no not not that that is not as interesting and no on that top okay go ahead when the first time is actually when when someone talks or it tells you that the ai agent is a person for you weird topic but i mean there is this test i think we talked about it once that you basically you you just and sometimes you're talking to a person and sometimes to an ai and if you realize you can vote after each one and you that just test if it's actually yeah let's see believable once i see one of those em dashes yeah true true true true we love it but i feel like it's gotten worse it's gotten there was a time where there was no em dashes and now i feel like everyone has it it's like anthropic back in the day was just open ai jgbt all the time don't do em dashes but now anthropic gemini never use em dash and mine as well but they're still there i'm not sure why you know why huh i think they trained the new models on synthesized data or in clean data and that clean data came through an llm and that llm put the em dashes in and they were probably too lazy to exponential curve exactly and then it's oh that's why at the beginning wasn't that i mean there's a limit on real data yeah unstructured i don't use it but yeah you had a topic you want to dive into yes yes yes yes it was just yeah and the topic is also touching on the agi theme where jensen juan said we already have agi or he thinks that and i came across a i think it was i'm not sure who who researched it but there was a um a journalist team which went into the whole beef between open ai and anthropic right they i don't know if you guys have seen the the the videos but where they were at the ai conference in india and sam altman and dario didn't want to touch hands oh yeah yeah but that's already i'm surprised you guys don't know no it's that personal yeah i haven't seen it is insanely personal yeah no no he they were like they're the only one like all the ceos take there and do this and they're the only ones in the middle they're right in the middle they don't do anything they're just like on one hand and they looked into the research where does it come from because there was a lot of rumors because the the main beef is actually not between slam altman like the original original og beef it was between uh brockman so one of the co-founders of open ai and greg brockman greg brockman you're right and dario um so the ceo and of anthropic it all dates back they all lived together in one so his sister dario's sister him and greg brockman they all lived in the flat in san francisco or in the bay area i'm pretty sure it was san francisco and it all comes back to a a beef they had in 2016 in their living room on the couch actually where um greg brockman and dario are um had a discussion about the publicity or who should know first of the development of ais so for example if we have agi who should know like who should know first and now this is the interesting point i think it's interesting yes um i know where where you're going go on good one dario's his initiative was the government government should know be the one number one person they should know first before everyone else but greg brockman had the opposite opinion and said the people the public he was apparently he wanted his one of his plans were to um give some of the major architects whatever plans to the un like council yes that's interesting very interesting but dario and that's where the whole beef uh comes from uh dario actually um yes that would be treason i mean i mean he's probably more of the american uh in that sense but he considered this as treason and that's where all the beef comes from i think super impressive uh interesting both sides have a point actually in a way yeah but i think it's very interesting that uh apparently dario's perspective is government should know i just wanted to say and now they said to the government yeah don't do it i would have thought it's the other way around yeah yeah but but if you piece one more puzzle piece together with with entropic now with the mythos mythos yeah with a new model maybe maybe and when they when they have to have to explain that i think yeah yeah too known yet and this it is mythos i believe and some other greek name but not as basically the the project name and that is a new model yeah and i mean apparently they're they're most powerful mystic you know almost agi level important the the biggest yet recorded training run so that model of entropic the new one that i didn't release yet um we don't even know if it's exists rumors yeah but but we the the community is pretty sure that exists and that it actually is the next step into agi and but the the fun part is internal blog posts and stuff got leaked and they they said that that model is super intelligent that it went i think 100 percent or 150 percent over their own engineering team's expectation of how accurate the model is going to be how good it will detect um security flaws and whatnot and is it a specialized model you know is it like a i don't know i don't know that is the thing they don't tell you but the thing is that they trained it on such a big massive amount of data and the interesting part is that some people believe that they broke the scaling law so basically the scaling law just means more compute yeah equals more more intelligence basically but now after a certain point some people some researchers are hypothesizing hypothesizing yeah exactly that that one broke the the scaling law that entropic got the architecture so good that you can now throw so much more compute on it but you get it 10 times higher or more accurate output i mean and that kind of aligns with the chinese models right which we talked about in episode one two two but it's also interesting that they keep it kind of very secretive and very like for themselves ish in their kind of agi wars but if you look at google and turboquant from last episode they like i mean it's also quite a big step to have eight times compression of your context uh yeah but i think you cannot open source yeah you cannot i think you cannot compare the turboquant with me force model that is a whole different level i just thought that like the the whole you know the company how they work like entropic kind of seems like the ones that would maybe yeah but yeah exactly but now the question is did it align with the new model and the government deal that is my whole question because if dario went back and had seen that the government used a specialized model that the government had like a really really high uptrend that of course was not on par with the public one but for me i mean this is all almost conspire a conspiracy theory for sure but the good kinds the interesting ones yeah but like when i read this i was like oh shit they actually might do have age or some some type of intelligence maybe not agi like god level but already such a jump what you just said because maybe they already because with this maybe they already told the government before but then of course they're yeah they were like yeah please don't use it on the battlefield yet and maybe that's why they said no please don't we don't want you to use it for mass surveillance and to kill like kill kill decisions everything else i think they were fine they're trying to monetize it's because the government doesn't use it anymore but i think they're still they ended up using it yeah yeah the government you mean yeah i don't know if they're doing it internally of course i'm pretty sure it's already wrapped up defense dod yeah the department of war no it's hexa i think yeah department of war they like he basically yeah told what was it like all of the government agencies are not allowed to use any entropic products anymore so in theory they're not allowed yeah true but that but that was how long ago two weeks three weeks i would i mean they they can't afford to say no i don't know how it is now they don't make a big fuss about using a i think i or whatever from my standpoint it's just played it's the playbook of the trump administration you know you always take out the biggest hammer bounce it on the table and hit everyone on the head and then everyone is like okay it's not that bad let's come all together again probably uh and i'm pretty sure they're all having lunch right now or dinner knows there's something especially if they actually reach that kind of intelligence with the new models well if it ever comes out if you look at yeah what happened in the past you know we had that yeah but yeah but if it ever comes out we talked about it we had we we knew it super interesting yes that's why you should listen to our podcast of course this is actually why indeed but i think we slowly have to move to stack or skip or do you have another i have another another um two little points to touch on on the new side um because i'm a little apple fanboy of course the apple um well we asked tim cook right in the first episode where is where is the usable ai features didn't come back unfortunately no unfortunately not yet but i heard some news for one that apple now is actually pivoting to a asked siri so kind of jet gpt style app to integrate their like to into have like an a uh yeah so you're saying they're giving up on the on a smart siri and they're just building yeah so the the app apparently it's a system-wide chatbot layer and it's supposed to be called ask siri i hope it's more than just a chat ask siri yeah i hope it's more than just uh but it's a standalone ios app that's what i think was was quite interesting why is it not just integrated in siri i mean they bought them yeah i'm guessing it is but i'm guessing so we both will be but apparently they pivoted to that standalone app also ask siri yeah no no it's a smart play everyone is now used to typing in in chat gpt so get the users involved into that but in apple product and the biggest point for apple is that they have such a big market share with phones and phones are still the the most used computer i would almost say for sure for sure for sure for sure now and yeah definitely um and if they get the user then more and more into the app and everything works basically from that app because apple has the deepest integration in their own os no one else can go that deep i mean you know it with pwa how much headache it was to to get the notification done and then they they have all of their data they know how their tools work and if they then just integrate it but you have a little layer a ui layer on a basically app and you can do everything through that they don't give up on the idea of having an actual smart siri after i mean no i think no no no think one step further for what is your your second camera button on your new phone i mean they already use it siri yeah i mean it already is they already use yeah but that's like that you get like the visual intelligence or whatever you do yeah i never used it actually one same sorry once but it's just not smart yet and i don't like do you even need the standalone app if you can literally just talk to siri ask your question i think it's more about making it accessible for the users because the users are already um using jet gpt you know that's the problem we always think okay yeah and you have to we have a different uh yeah i get it and one more one more thing in the apple area that i thought was quite interesting right we talked about that apple is buying resources from uh google for yeah for computing yes custom training apparently they now got leaked they also bought the rights of some gemini models they i'm not sure what the exact legal classification is but they can actually do their own stuff with the models and apparently they're working currently on having gemini models running locally on your phone exactly and that's supposed to also be that's what they're building on top of with all of the new chips and all of the the m5 chip the a19 chip yeah but what i was surprised by it was that it was actually a gemini model and not like in-house if you were they gave up on that months ago they just said we can't do it and we're gonna buy or like buy a specialized imagine how much money they saved exactly i mean it's actually smart like they did they saved probably give up and get by bias no no no i don't think that they couldn't make it i mean it was just be too too expensive and too much time yeah like until they wait all of those people in why would you the whole architecture yeah exactly just buy it yeah and i think they're betting um big time big time on the on device computing in the future i think this is their bet like also privacy security wise i we sure think so yeah i love the the connection between um the first iphone basically what your your personal computer and now jensen huon your personal computer it's basically the same sentence but means completely different things because first we had actually the computing layer in our hand and i don't know how long we actually have the computing layer in our hand or if we do something with that or rather we have one computer at our home and that is basically our brain or our agent surprisingly both sentence make complete sense yeah in both times it's just different timelines yeah but but it makes so much sense in those two decades it just makes sense actually that they picked gemini not even only because of the quality but also because google was the first one and still kind of is the first one that really plays around with small phone size models like gamma the whole family of gamma models is based or built for running on phones so they're probably also tapped into that kind of area and also google is looking into in the pixels etc running and locally using those gamma models which is kind of smart but still i think it's crazy that apple doesn't have their own yeah yeah i mean they make it their own now they buy the architecture basic training and i guess they just adapt it but the question is going to be anyways is there one model or are there going to be a hundred specialized i would say especially those small models probably they have yeah but are they are they all from one company let's say like like from apple the iphone because you don't see much more we don't know phones out there i think it's only gemini currently like that they work together with google in a way to to get that but i don't know it's unofficial of course no no i mean more and more in the kind of sense if we have in 10 50 or let's say in 50 years if we have one model basically one underlying architecture that is the best model that all of the people here could figure out in the last 50 years and everyone uses basically that and specialized versions of that or do we still have those million different labs and everyone tries to reach agi first or not agi in that sense then anymore but rather you have a healthy competition yeah exactly that's why i hope it will be will be dispersed i think i don't want to be be told to be using one model as long as the biggest company still leak the open source as long as they are using their their own tools to build it true true true yeah but also this last thing on apple because i just remembered it they actually uh published yesterday or today or something they went live with apple intelligence in china before actually getting the accreditation from the government like the legal kind of you know and then they pulled the plug again after like an hour or something but i don't know what's gonna come to them it's like they felt the pressure of oakenglore maybe but maybe they just forgot you cannot forget that probably you can but how does that happen how do you go live in a country like china where everyone knows that you have to kind of watch out especially when it comes to ai yeah well but i mean there probably was more involved than just we'll probably have more probably not just someone yeah for sure well um any tools we should talk about this week stack or skip let's go you start yes you have the best ones always or not yeah you put me on a pity here right now sorry um start off let me let me think for a second i go for i mean it's not a big one to be honest because uh turbo quant was still my biggest one i think i actually read that it's one of the biggest kind of news of the year but let's see everyone is kind of for for local ai definitely for local yeah um but no what i saw is small kind of niche thing but as i worked with java a lot back in the days it's still actually being used quite a bit um so much my my my homies from uh germany they're still studying that yeah it's crazy yeah okay germany is also still yeah it is being used still and google released an adk for java to basically integrate gemini in it natively but it's a really niche niche topic so i don't think we have to expand too much on it uh for me it's a stack but it doesn't need to be for everyone so to be honest no good master if you say stack for me it is for me it is if you work with java have a look if not then you don't have to so yeah proceed whoever wants to go ahead yeah i have just a little nice tool nothing nothing really fancy but i think i showed you guys already the apple pass basically it's a it's a little website um what's his name apple pass yeah apple wallet ah yeah the thingy that i showed you yeah it's basically a little website where you can configure your own apple wallet pass whatever and then you can basically send it out to someone and it's basically like a you can share it i think also like a drop wise no i think you cannot share it but you can you can have a qr code on top of it so you can basically show people and they can scan it and it's basically like a digital business card yeah okay so super fun but a bit buggy still to be honest because apple has are the days no they're trying to make it out probably no exactly apple apple has such weird um roundings and padding uh around it and the the writing is cut off and whatnot so it's still early but really nice tool good really really looking forward to it the only thing that got me a bit was i can send messages through that also i don't know but as soon as you have that card on your on your phone basically i have a wire to your phone i can send you messages through the website on your phone and it gives yeah it's super nice secure yeah but then i i'm thinking it's nice for you as the guy having the yeah but then i'm thinking if someone could build a tool in a weekend that can do that i mean how many how many applications are doing that just in general i what i'm always surprised is like why is our qr codes like never taught no no one talks about the dangers of qr codes you can literally embed anything you can literally and then you just like pay your parking ticket for example next to and everyone will scan it yeah and no one will notice anything yeah yeah i've even read a story about people making money of putting fake qr codes on parking machines yeah i told you about it yeah they basically take like one dollar on top of the normal price and they just have a website that buys then off the official website but so it's technically legal but it's not illegal at least not yet so they may be no one patched it yeah so make sure guys if you scan any qr codes be aware before you scan them because you know who know who did that yeah yeah you can do a lot you can do a lot of things with qr codes you actually can no yeah but those were not too many i mean it's been five six days since we recorded the last one so that's uh the main new tools i feel like there's already so much happening yeah i just want to say even entropic released two new tools i believe the computer use fucking awesome actually i mean let's let's talk about computer use overall it's not only like who started it's i suppose perplexity or someone i guess i guess they're trying to catch up with open claw right pasting each other right now like the big new topic agents agentic kind of workflows computer use like models that can actually make use of your controller who was the first actually open claw i'm just no no no there were there were or i mean people already built built that a long time ago there were i don't know i think but like very niche yeah only the browser thingy yeah they but they gave you basically only the computer usability and you had to build the computer use basically what your model does with the information that it gets from that tool and it needs to reason about that and then does the thing you can build your own exactly so so the the connection between the llm and the browser was already there for a long time but i haven't played it integrated into their own kind of systems i know but it is also you see that they they don't give out free credits no not on none of the companies gives out any free credits yeah it's ending it's ending i'm not sure what exactly happened because it kind of started with everyone at the same time the investors probably just want to see no money coming back i don't think that i think i mean if you're looking at the ram prices and the computer prices overall everything gets so freaking expensive and ai produces so much data you need to put that somewhere and the the the the the screws are being tightened in terms of investors want to see some level of sustainability of their investment yeah i mean look at open ai they still make what two billion losses a year good that you mentioned it the very last uh fact i wanted to bring on this episode what do you guys think so we talked about that sorry god's got put down let's put it that way how much do you think it costs them just to maintain the architecture architecture the numbers it's just one number i'm not i didn't research it too much although like there's probably some level range would already be interesting but wait wait wait so so how much it costs to prefer them per day to keep it up and running only up and running yeah i don't think for usage no usage okay no not i mean because that was dying down yeah i have no idea per per day per day oh let's give it like a five five million i don't know i'm going higher they're losing so much money every every year i think i think 10 million i think this is fun we need to do this i think it is between 10 and 15 million that is my range you need to say one number no range okay then i say 13.7 million i go up to eight congratulations oh what shit 25 15 million dollars per day per day per day just on the infrastructure i don't i don't think it included usage that's why they so if if people actually used it yeah i mean i mean the the the second i released the gipfy gipfy you know the trend where everyone just posted their picture into the chat and let it do a certain style yeah the servers were almost burning yeah i mean we talked about it but in that kind of time period like those few weeks every social media was full of sour watermarks until the one guy released that too but i don't know interesting actually i mean i would have cost but no i would have cut cost as well in that sense but yeah 50 million is crazy 50 million is insane i don't even think that many people were using it anymore that's why they needed to put that's why it was just a base cost but not even more usage or yeah no and it was included in the gpt plan right so even in the free plan i think you had a couple you had three images per day or one i think okay but images but i mean sora is videos i don't know about yeah but the underlying architecture is basically the same sora is basically just 30 times a second it produces an image no that makes sense yeah that that is also why sora takes so much money i believe because alone an image generation takes so much energy so much power i mean we we wanted to test it on the our local server and the pcie lanes we need to figure out which card we put in which pcie lane because it depends then how much throughput the model can actually use and those image generation models they actually use a lot of throughput and need that so alone the the data centers behind that is yeah envi link i recently checked like envi link the newest kind of generation generation is crazy compared to the first crazy and it's been like what two years since the first one came out i mean it shows you again how fast the whole the whole tech industry is moving now it's crazy yeah and we are here to report on it every every week well i think this is a pretty good moment to wrap it up right we're all quite cooked a long day for us long day for you if you hear this in the morning have a good day if you hear this in the evening hope you had a good day and sleep well i'm definitely not gonna sleep tonight claude we're gonna come all right all right thank you very much for listening uh please make sure to leave us a follow uh give us five star review yes please give us feedback also uh big shout out to the guys behind the cameras for setting us up doing everything so uh very very much appreciate it please give us give us feedback and keep building peace