← Back to search

The AI ​​that doesn't sleep: from DeepSeek V4 to Hermes Agent

Risorse Artificiali AI Engineering in italiano · 2026-05-02 · 75 min
relevance 88 9700 words Episode page ↗ Audio ↗
Show full episode description
DeepSeek V4 taglia il 78% di compute sulla KV cache, Roo Code chiude e Hermes Agent gira da solo. Settimana densa: DeepSeek V4 Pro e Flash con 1M di contesto, Kimi K2.6 e MiMo V2.5, l'analisi di Salvatore Sanfilippo che fa girare V4 Flash su Mac mini con quantizzazione mista Q2/Q8, il post-mortem di Z.AI sul bug async cache di GLM-5 (con fix contributo a SGLang), Vision Banana di Google DeepMind che usa la generazione immagini come interfaccia universale per la vision, e il deep-dive di Stefano su Hermes Agent, l'alternativa open di Nous Research a Claude Code che gira con GLM-5.1 e gestisce mail, calendario e paper di Arxiv in autonomia. In più: le interviste a Demis Hassabis (DeepMind) e Andrej Karpathy sull'efficienza e sull'AGI, il keynote rotto di Jensen Huang con Dwarkesh Patel, e i 40 miliardi che Google investe in Anthropic. Follow Risorse Artificiali per non perdere il prossimo episodio. Versione video su YouTube: https://www.youtube.com/watch?v=qKl4Vkb6BMw&utm_source=spotify&utm_medium=description&utm_campaign=ep50_drop #50
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
Reviewing the wave of new Chinese open models and the shift from research to inference engineering.
Benefits
  • Token efficiency lowers cost for same output quality
  • Open-source inference optimization lets anyone contribute
  • Adaptive thinking tunes reasoning depth to cost
  • Mixed quantization runs huge models on consumer hardware
Use cases
  • Salvatore Sanfilippo ran DeepSeek V4 Flash (284 billion) with a million context on a Mac mini Pro using mixed Q2/Q8 quantization
  • Google promises a 40 billion investment to Anthropic; Alibaba to invest in DeepSeek
  • GLM annual subscription at 20 dollars a month degraded after ZAI super-quantized deployments
  • Installed Hermes Agent to handle many tasks replacing closing IDE coding extensions
KPIs / results
  • DeepSeek V4 is 1.6 trillion params with 49 billion active
  • DeepSeek V4 Flash 284 billion, one million context
  • Google to invest 40 billion in Anthropic; GLM plan 20 dollars/month
  • Evolve.js comparison context: Kimi K2.6, Mimo 2.5 released
Tools / build
0:00 / 0:00
🌐 This transcript was automatically translated to English from the original.
Hello everyone, welcome back to Artificial Resources. This week we talk Chinese about models, we talk about DeepSeek which came out with the 4, Flash Pro, but there are also Chimik 26, Mimu 2.5. I don't know if you've seen it, so I'll give a little preview and then we'll go into everything, you've seen that Google promises 40 billion to Antropic, that is, a big investment, everyone is investing, strangely enough, they're not competitors in some way, but this big investment is being made, but the Chinese are also making investments because Alibaba says it will invest in DeepSeek, so even there we're starting to see this flow of money, but then there are many other interesting things, there's Vision Banana to name, then the two or three things that I was most struck by the scaling of the lessons on scaling of GLM5 that Paolo mentioned to us and the interviews I heard yesterday by Assabis and Carpati. Although I must say that the two most beautiful things are, that is, the most beautiful, one that left me like this, Rucode that closes, the Coding extension that closes, because it says that, and then maybe we'll talk about this from our point of view, that it's no longer time for IDEs, for IDEs and then oh well, then I installed Hermes Agent which is doing many things for me, even at this moment. That interests me. Know that when you said that this episode the models speak Chinese, I thought it was like when my Chinese models go crazy and give me answers in Chinese to questions in English. Of course, but in fact that's what we'll do, we'll give random answers to the questions we asked ourselves, because even if I have to say that I was saying because we don't have any comments, that's not true, because this week we also had quite a few comments, including ether. Chinese Mandarin? Chinese Mandarin, absolutely. No, come on, let's start from that, come on, let's start from the Chinese, DeepSeek, Mimo, Kimi K2, we have now decided, Alessio, that you are the one with the benchmarks, so you say everything about the benchmarks of this work. I can basically tell you that last week it almost seemed like a competition, let's throw out the Openweight model, the new State of the Art, in the sense that DeepSeek V4 Pro and Flash came out first, where Flash is a slightly smaller and clearly faster version, Faville, a whole series of new features, a new type of attention which please, talk about it Stefano. Well, ok, we have the new best Open model, except for the fact that after a moment Kimi K2.6 came out, which does even better, at least in the benchmarks, and Mimo 2.5, which more or less aligns with the other two, but it shows us something new, that is, new, it highlights an interesting aspect, which is that of the efficiency of the tokens, which is a topic we have talked about little, in my opinion, in recent times, that is, it essentially generates fewer tokens for the answers, to obtain the same level, let's say, of quality in output, in passing tests, etc., and that has an impact because less tokens means less cost, basically. Let me digress, I already have a second to divert you from this conversation about saving tokens. You reminded me that in the last couple of weeks I've come across a lot of videos or articles that talked about a revolutionary extension of cloud code, called Caveman, I don't know if you've come across it, where the idea is to make your model ball like a caveman, so instead of giving you long and detailed answers, it answers in monosyllables, like, I don't know, to a child, when you have to introduce him to his distant uncles. And nothing, people were very excited, saying, ah, this is beautiful, I'm saving tokens here and there. Let's say, that is, it was curious, it was against the trend, it could have made sense, except that with a little more time people looked into it a little and said, yes, look, most of them, if you apply this process to most of the cases you ask them to do, then they give you inconsistent answers, because imagine that, I don't know, you're asking them to make a diff between two code sources, you may want a more concise answer, but if they answer yes or no and it doesn't mean nothing, you're not going anywhere. So there was this trend, the name probably evoked who knows what things, the simple concept interested people, but then on the practical side it wasn't that useful in large use cases. Yes, but I'll tell you, in my opinion there is a, this is a metric that has been a bit underestimated lately, because if you go and look at the models that do reasoning, thinking, in reality a lot of time is spent in that phase and sometimes we are inclined to underestimate this aspect, but then for example when you try to run the models locally you notice this aspect even more. there are probably some applications for which, all things considered, you are willing to accept perhaps some errors, a slightly less quality response in exchange for greater speed. So much so that, for example, the cloud code itself but also other tools allow you to set the level of thinking reasoning you want to have. Clearly the quality changes but the response times and consequently the costs also change. Yes, look, it's something he said to Sabis also in the interview that I anticipated then maybe we'll go into a little more detail but now I was looking for my point he actually says that there is an architectural efficiency problem on context windows currently what has been done for deep reasoning etc. and he quotes himself he quotes Gemini 3.1 pro deep thinking the one that won I can't remember which mathematical competition he says that the efficiency problem is big because essentially brute forcing has been done up until now that is it generates many tokens and take the good ones he says when we talk about efficiency but also effectiveness it is not necessarily true that brute forcing always pays, in fact experience teaches us that when he starts optimizing he gets unexpected results and in fact on that thing the interviewer who is the CEO of Y Combinator asks him but so how far are we from EGI from a scientific point of view I speak to him rightly as a scientist like Sabis is and Demis replies we are not far away it could also be that we have already discovered everything we need we just have to optimize it it could be that we are missing those two or three ideas but no more revolutionary than Attention or so was to be able to truly emulate an intelligence different from human one but still something that we can call truly intelligent in fact he says that in Deep DeepMind they have two completely distinct lines of research which every now and then they compare with each other but which they also deliberately keep distinct as groups, one is on incremental research and one is on breakthrough research so they really explore the two possibilities they have an extremely scientific approach in this is one of the interesting passages but it is also to confirm something you say about efficiency also comes back to us, many of the Carpathians are coming back to it at the moment about the fact that they slobber the models a bit too much is one of the controversies that also came out on opus 47 that someone complained that it is extremely verbose among other things the same cloud code now has adaptive thinking that is they tend to try to propose an automatic system in which based on the type of question you ask but also GPT has had it for a while it decides how much thinking to do because maybe the question seems relatively superficial so it's not even worth having to do excessive reasoning just to give an automatic system that improves the experience in general, however, going back to DeepSeek 4 we were saying that there are some interesting improvements from the point of view of the algorithm of how it works Stefano yes yes no there are many interesting things first of all let's start by saying that it is 1.6 trillion yes exactly gigantic models absolutely gigantic model but a mixture of experts now I'm looking for the data and I've made some notes so as not to get lost the pieces 49 billion of active token parameters which yes is a mixture of expert but we are used to those with 3 4 5 in short it is larger the active parameters are larger than the average of many models that are around at least among the open wait let's say when you think about the models that run locally but it has some interesting characteristics this thing of being so strongly mixture of expert because someone and in this case Salvatore Sanfilippo is the one who wrote about it the most to stay in Italy he started to say attention but because this model is interesting here because not everyone the weights are the same the ones that matter most are the routing ones I think and then I can do different quantizations so much so that on some groups of parameters he managed to quantize Q2 on others instead in the attentive Q8 and this allowed him to run Epsic V4 flash which is a little smaller 284 billion anyway not exactly a joke with a million of context he ran it on his mac mini mac mini pro with I don't remember how much ram but I don't remember how much I don't know 628 196 but still it's a result he says with good performances he's been talking about it for a week on 'in the era of the dawn of mp3s when we were 20 years old or less in which there were various forms of compression to have the best quality we started from 128 then it was 192 and then we spread towards the variable byte rating so different parts of the file were sampled with different variability and this San Filippo approach reminds me a lot of this which obviously makes sense it increases complexity because you make the system work in a different way depending on where you are or when you are it makes sense from point of view of software engineering yes yes yes I agree and then it is interesting that from this point of view there is also an interest from the open source world and we noticed more than anything else it is something we noticed that perhaps Alessio put on the table a couple of episodes ago as in part we are moving from pure research to software engineering from engineering in general and it is no coincidence that actors like Antirezza Salvatore San Filippo enter the scene he obviously has a notable background especially on CIC++ in general on programming I would like say and enters the field by also putting his nose into this kind of thing which is yes on the models but it is optimization of the inference to go and do different quantizations so the thing is interesting from two points of view both for the result that he has obtained which he talks about and which is certainly interesting but it is also interesting to see this shift of skills which makes the part let's say pure research on the models but it is the engineering of making them work with the hardware we have which returns to making the open source movement become more protagonist than it could have been with the open weight because let's be clear about an open weight model I'm very happy that they gave it to me and that I can install it etc. etc. but it's always something that comes down a bit from above, that is, how do you contribute to that stuff? You can't if you're not inside one of the research centers that release it. Instead, if we're talking about optimization of inference rather than agents and optimization of tools, then we'll also talk about that. The open source world becomes the protagonist again and we use we but certainly I believe in it as strongly as open. source had an impact on operating systems in the 1990s, the end of the 1990s and then on everything else in the 2000s, I believe that it can also have an impact from this point of view, I don't know if you have a yes but then there is also a different group of people who can contribute at a research level yes that's fine ok we'll make it open but how many are able to and how many have the resources as well as being able you also need the resources if you're not inside a research center good not good competent not competent whether you are you can't do that stuff there if we come to software engineering it's absolutely different if you want I can digress in this regard in the direction of the post-mortem of easy AI in which they published why exactly that is the point ok so what is the context easy AI is one of the providers that provides GLM the model that we use for how I use it of which I had an annual subscription because it cost very little and gave good quality results at a certain point it started to give poor quality results it was hallucinating let's say using a somewhat vintage term at this point since it happens less and less often but not only he was hallucinating he was inventing words he was mixing things up it was as if he had gotten drunk or for those who have spent enough time in front of a screen it was as if the memory had been corrupted when characters come out that should be there and there is no explanation for why they are there internet Reddit specifically obviously she was infinitely outraged by the fact of not receiving a level of professional quality for their 20 dollars a month that they paid and that they said I would never have another quarter they will not have never again my money obviously that is these things she doesn't know who I am these scenes here in short and nothing since there was no transparent communication those people seemed competent the thesis was that these people from ZAI were a bit greedy or in any case they were a victim of their own success so they could have had more demand than they could serve and so they closed the tap in the form of deploying their models in a super quantized form which obviously degrades the performance in terms of quality and this is the result tate and this was a thesis that I took as good because I didn't tend to have other elements and people believed it so much so that someone thought that ZAI's was a business that was going to die because the quality wasn't there it was interesting sorry another reason why people had made this assumption was because the same models provided by other providers specifically Alibaba worked very well so people told you no I'll leave it alone ZAI are stingy they let them go there it works well and nothing and we were like that for a a bit ZAI didn't bother to comment some time ago a comment came from one of the ZAI engineers who said ah we had a bug try now and things should be better things have been better since then so let's say the story checked from their point of view and they promised a post-mortem the post-mortem arrived yesterday or the day before yesterday and in my opinion it was more interesting than I could have imagined for many aspects meanwhile for transparency I would say yes partial transparency I would tell you now I'll tell you why from my point of view as an engineering manager basically a cynical person and not very optimistic towards the stories you told us basically basically they gave very low level technical details explaining what the problem was and therefore from the point of view of technical nerd satisfaction they explained things as they were and the answer is credible but it is the opening let's say that I find it slightly more questionable in which they basically said we are happy to share with you the resolution of this problem which was very rare and we had a bit of difficulty chasing it we never managed to catch it yes a couple of balls in the sense or not were you using it well or were you not looking at it because I and all those others caught it after five minutes every day so this is a bit of my criticism obviously they have to maintain an official room so if you are vague on the terms you say rare conditions what rare means for you or for me is debatable and therefore it's fine so that is a bit of my criticism however once removed this façade of admitting a fault but up to a certain point they have reached the point of talking about the point and for those who have a background in software like you two specifically but maybe it comes from me think of San Filippo in this case because it is something close to him also try to guess what is the main cause of the problem they had asynchronous cache propagation ah asynchronous who has ever had a problem with synchronicity especially when you have to take operations into account what does asynchronous mean for those who are not familiar basically start concurrent operations and not always wait sequentially for each response before doing another action to go faster we decide that I can do two things at the same time like in real life every now and then when you do two things at the same time you make a mess and therefore you have to have a bit of let's say self-control to coordinate yourself here they attributed the main causes to the cache to the kibi cache and to the synchronicity of this thing for which there were conditions in which case sometimes if an operation was aborted the asynchronous operation of the cache was not yet completed and therefore maybe the next operation was in a context that was not its own and that's where the very relatable mess came out let's say from the point of view of the software of whoever had their hands in it that is everyone there we have come across this thing at least once and so their story is particularly credible in my opinion because they have identified a bug specifically in this context in one of the projects on which they depend an open project now whose name I am now looking for because I read it this morning but I no longer remember it a related project is SGLANG which I don't know well I confess I don't know exactly what its role is but they linked the PR that they opened to that project so officially there was a problem they recognized it in their stack they have fixed and they are contributing to open source so hats off to this side here then let that be the whole story who will ever know most of the story makes sense and is interesting from a moral software point of view you have to know enough what you are doing if you want to provide a service to a large audience otherwise there is no sp erance that you can ask Cloud Code to do it for you and that's it I don't want to say stupid things but it seems to me that SG Lang is like the inference engine on which they run the model if it were the Lama SPP ok ok or the VLLM so we were saying therefore you can also have the best model in the world but if you have problems keeping it up to scale to levels of competition of use like those they can have with millions of people from all over the world who want to use it it's not the same as saying it and I can't believe it because if you have ever even just tried to set up something local on your computer and you've already run into performance problems etc. imagine this thing taken 4-5 levels of size higher yes absolutely returning for a second instead to DeepSeek you asked me about the attention, how much the attention has changed and to talk about it a bit then as always I would say what it was like let's say January last year when the first DeepSeek R1 arrived if you remember the first DeepSeek moment the first DeepSeek moment DeepSeek is more interesting for the papers than for the model itself because in the end even R1, apart from having made great news for itself, it's not that it was adopted on a large scale also because then all the others arrived coding wasn't its strong point it came out when coding was on the launch pad etc. yes no but Stefano am I wrong or was it the model that deployed perplexity ehm v3.1 they used if I remember correctly it's the next one but yes it's still a DeepSeek yes yes yes it's still a DeepSeek you are right about this however in the sense that it gave ideas on which all the others then developed let's say well no DeepSeek's papers are very interesting because they are really very open in research and in sharing what they do and I quickly remember that in DeepSeek R1 or rather V3 which is the basis of R1 there were mechanisms that had not yet been seen of um sparse fast attention and flash attention which are two ways of putting attention on the tokens different from how it was done before now no one made them anymore as per the original paper attention is all you need which takes everything and looks at everything because it is too expensive however the sparse flash attention of sparse sparse fast attention now I get confused with the names of DeepSeek V3 was very interesting because it managed to have a multi-level attention which led to having a broad context at the level of concepts despite costing little now going into detail in a podcast becomes complicated and I won't go into too much detail even in the new one but it is remarkable it is remarkable it must be read re-read for those who make the models the inferences are already in particular from what they are already doing because they insert the concept of what they call hybrid attention architecture and they have three three big innovations one is what they call sliding window attention so basically the attention is divided into pages as if they were memory pages for those who are a software engineer instead of using the raw text within the memory pages they use a sliding window production to consider how the last pages overlap with each other this thing allows you to concentrate the attention only on a more limited number of tokens and ultimately have efficiency but the two most interesting are those called csa compressed and sparse attention and hca i ivly compressed attention practically they manage to make a million tokens spending 78% less of the memory computing power not memory because because basically instead of having a flat KVCash therefore key value but now I'm always oversimplifying if someone listens to us who makes an inference and doesn't redact how I explain it but to try to explain it instead of making a dry key value map with the kvcash they have essentially created a system hierarchical tree memory in such a way as to recover more quickly those portions of kvcash that are significant for the attention of that moment this level of compression is done for more or less large pages of kvcash depending on how old the context is so the first one I said is let's say in things that are no longer new but not too old while the therein goes in the older things so the older the context becomes the more I compress then in itself interesting it must be seen in the field because certainly in benchmarks and management tion of the context they have interesting results while I was reading it some light bulb on the impacts that this thing can have came on for me it is definitely an improvement compared to a dry compression of the kvcash intuitively it is as if we were saying that we pay different attention when looking at the things that have just happened compared to those that have happened in the past intuitively yes but if you want intuitively it is more similar to saying that we are paging on disk if you think about the RAM that is you don't do direct access but you do index access to the pages and you reload the pages you need then you do the attention on the attention page the real attention which then leads you to decide the next step in the layer you do it on the real context but the recovery of the real context instead of scrolling all the way back with KVCash go and get the page where it is most highly probable that you will find that information obviously it is highly probable and not certainty so in this sense you can miss some information that could be interesting the paper is read well I didn't have time to delve into it to the end I read it but once we say, however, more or less this is what happens inside the paper and it will certainly be adopted in some refined and refined way also by others because saving 78% of the computing power to manage the cash cables is fundamental and here if you want I'll go back to the interviews we mentioned, in particular that of Assabis because for me the interviewer was enlightened by that thing I heard him say about business choices rather than technology and the interviewer asks him in short you are becoming fundamental because now all Google products they integrate Gemini in one way or another you've seen they put it in maps it's in the whole workspace and he says yes and then there are all the AI modes he answers him not to mention the AI synthesis of the searches which are billions of calls a day or whatever because having integrated it into the search anyway it's the most used thing in the world and he says in fact for us in the last few months the search has focused a lot on distillation because all that stuff you see integrated uses flash doesn't use pro because flash we managed to get flash to give answers which are of the highest quality spending less than a fifteenth in terms of computing power and for us it was fundamental otherwise we wouldn't have been able to put it everywhere so it's interesting how much we can actually say that OpenAI sorry OpenMind is doing absolutely frontier things but of a different frontier that is they are making it efficient and effective instead of concentrating on the further development of Gemini 3.1 then I could be wrong but looking at the cell I see Gemini 4 coming in May but this is just my feeling but it's interesting how necessary it is on the one hand to have ever better answers to get there he also talks about it to get to the GIs okay later I'll talk about it better from that interview there because it's super interesting but how important it is also to be efficient and distribute the load as much as possible even on the Edge etc. etc. etc. and again engineering of the question exactly it's more engineering than research even though they continue to do a lot I also like to see that guy going back to DeepSeek every now and then they come up with these discoveries which are the ones that make you take the step the step and then after all the others in one way or another they refine and adopt and we will have more powerful models that can be used on resources more or less equal to those that were initially required for poorer models and then there is another thing that I think is interesting to say, let's first mention the quantization done by San Filippo which is a Q2 on one side and a Q8 on the other but the paper explicitly says that the training was done FP4 quantization aware what does that thing here mean that when you do the training you represent each weight of the perceptrons in FP16 normally FP16 there is someone who has done it in FP8 but basically it is done FP16 they have done it FP16 but making the training FP4 aware means that in looking for the local minimums which then bring the general minimum of the model because it is always a function of it is the search for the minimum of a function in the end towing a model trivializing a bit but here's the stuff to say that FP4 aware it means favoring the weights that can be represented in FP4 without too much compression compared to distributing the weights a lot on an FP16 so it's like saying ok I have all the power of the FP16 but I try to concentrate everything at the top at the bottom in the center in one of the FP4 cuts that a quantization could do already in the training phase this should bring about you can already see it in other models minimax has done it if I remember correctly among the first it was minimax 2.5 having made this choice this thing here leads to having quantizations that when they are done, it's less difficult to do them, those of Asloth don't go crazy and then the result is better anyway the result is better because you're cutting away less stuff i.e. the information all concentrated in blocks of FP4 in some interesting way this increases the number of total weights a little so it's always a trade off however they are given to trillions and trillions so a trillion and six if I read correctly then the last perhaps interesting thing from their peper too technical to explain here is how they do the storage of the kvcash to use the sparse attention that I said before ah no and attention attention that it's something we discussed in chat every week but I'm sure that even some listeners might turn up their noses because the benchmarks aren't exciting but be careful that the benchmarks aren't exciting but it's a preview this one isn't the finished model it's a preview which means that they did it they say largely less which doesn't mean anything widely less anyway less than 50% of the reinforcement learning phase so it's a half reasoning model I don't know how to put it because they probably did it the basic alignment to give answers to be an instruct everything for sure that instruct means that it is capable of conversing not just predicting the next token because this too when I go to conferences it is often a question that comes up but now the models are conversing no the models are not conversing the models still do what they did before they predict the next token or the next block of tokens reinforcement learning human feedback or verifiable reward which are the two phases that take place afterwards they teach the model to put together answers starting from the prediction of the next token this is the first phase of reinforcement learning then there is a second which is the one that teaches him to do the same thing by generating the hidden tokens or the reasoning tokens also known as those that allow him to do the reasoning this part is done in a very small part on the DeepSeq preview so I expect the benchmarks of the final to be much better then it is not certain but consider me a figure of speech Stefano it is a bit like the difference between knowing the English vocabulary and knowing how to speak English yes yes yes it is correct it is absolutely correct that thing that you say it's precisely the difference there is that of the RLHFR phases then let's call it reinforcement learning in reality it's more of a distillation but let's not leave aside the technicality there is the phase in which you teach them you teach them to do the reasoning which is an even different phase which according to your comparison is like not saying bullshit exactly well done I would have said exactly the same words and this is more or less what I can tell you in the words of the paper however the paper for those who are technical in the sector must be read it is interesting deep seek moment number two yes it did less boom it depends, envy's shares have collapsed by 20% or not because if they haven't collapsed it's not a deep seek moment no also why not say it he's coming to get us it's the CEO of envy which sooner or later we could he listens to us eh Jensen no we're not mentioning his name no sorry he's being monitored now he's coming to get us he's monitoring with artificial intelligence everyone who mentions his name speaks badly of him no I don't know if you've seen maybe we've already said a few weeks ago he went to an interview before he went to Alex Friedman in the normal interview serene then he went to Darkash who instead pressed him a lot about Chinese exports about the fact that the Chinese are catching up and that hardware is no longer the most important thing because Google is demonstrating that with its chips it is going strong etc shit he got pissed off that is he really changed his voice he no longer had the usual keynote voice as he calls himself he started raising his voice to tell him what the hell then he answered the end instead is full of memes about jensen's su DeepMind came out with a paper saying, bringing to attention the fact that the models for generating frontier images such as banana can essentially also be used as models for doing vision, let's say specific models for doing vision tasks such as segmentation such as depth estimation dimension such as edge detection, these things, no, the idea is not new, that is, it has already been talked about, but the point is that they are in a position in which they have a particularly strong model and can at least demonstrate that they can go beyond the claims because then it is not that the model is open and we can all try it to do these things that the results that are obtained are substantially at the level of the models, let's say, segmentation frontier, depth estimation, edge detection, etc. and what does this mean, explained in a moment in a simpler way if we think about what the current image generation models do, we basically start from an image, we provide a prompt or we start with a textual prompt, we generate an image, however, there are editing models to which we give an image, we give a prompt and obtain another image and then intuitively the concept is that at the moment in which this prompt expresses the vision task that we would have liked to do with a dedicated model a specific model we can try to obtain the same thing for which a segmentation task which is like tell me find the cars in this image can used on an image generation model lead you to have a new image in which for example the cars are colored all red what I am showing in my opinion is quite impressive two soup with something to eat next to it two prompts one the first which isolates the garlic the second which isolates the meat that is, this is quite impressive made by a model that is not dedicated if they have learned to recognize the cars I can count on the fact that Google will stop bothering me with its captchas to ask me which is the traffic light and which is the car. Basically we are moving from let's say vision models which generate output for you which are essentially matrices of numbers which tell you the position in the images in which you have the thing which is, let's say, detected, to models which are these image generation models which essentially produce another image in which the information, if you want, is in the color channels and therefore you can express the same ones in that way. things that you would have expressed with the matrix of numbers so trivially if we are talking about edge detection you have a black and white image where in black you have the edges of the edge thing if we are talking about depth estimation then normally you would have had a matrix of numbers that tells you the distance for each pixel the distance from the camera you have an image in which you have different color scales and therefore the lighter pixels are perhaps the closest ones and the darkest ones are the furthest ones or vice versa at the same time you could have heat maps for saliency to see to say what a part is salient feature of an image is interesting in my opinion from two points of view the first is that in reality many applications of vision models then to be conveyed i.e. to be used require that you transform the information that the vision models have given you in turn into something in a visual image that can be interpreted by the human therefore using this directly the image generation model to do this thing if you want you remove a step it already takes a step that you would have done sorry you made me think when you go to see the ultrasound of your children and they tell you see here you see the baby and you see like the Andromeda galaxy and you say yes yes I see it I see it then I found it interesting you know for another thing also because it's a bit the pair with something that Asabis continues to repeat not in the interview that I mentioned before but I heard it in another speech when he says that video generation models can instead be a good tool used in reverse, i.e. by feeding them a video to interpret the physics of the video because they are capable of generating something that is credible if not exact physics but still credible in the video, in several speeches he says the potential of these objects also lies a lot in the use that we could make of them as 3D vision or continuous vision he talks about robotics in that situation that is how do I show the stuff and understand what the reverse engineering of the video generation has around just as here the reverse engineering of the image generation has been done so it is somehow seen that in any case image generation and video generation despite having slightly different techniques etc. but they are the same family because to make a video it creates the frames it is in some way a confirmation of this theory of his and then there is a question of versatility in the sense that when an image generation model appropriately prepared with type data redraws this image for me as a depth map redraw it for me by coloring objects of a certain type in a certain way, isolating only a certain type of object, etc. you have a model that is universal that can be used for segmentation, for depth analysis, for edge detection, for object grounding, in short, all these things without having to worry about having one for each type, clearly there is a question of resources needed to make this thing go, but if you want, we can go back to the previous topic, there are optimization margins, etc. look, I find what you told us super interesting and I It's surprising because in the past I wouldn't have found it so interesting, not to offend your field of interest but relatively it has little to do with my daily life but it allows me to mention the fact that in the last month Stefano and I, apart from when he went on holiday, had fun and tied me with a hot match, we dedicated ourselves to a robotics challenge, a robotics hackathon in which these aspects that you described are absolutely key and important starting from the most banal and old at this point models to make the YOLO algorithm which identifies it puts a box around you which are now commodities that is you have a call you have the result some of our contest partners who are real engineers in the field of robotics which we are not have instead shown how to apply some of the most advanced algorithms to the identification of edges so there was this little remote controlled car imagine that it went around with a camera and it recognized not only the chair but it recognized the edges so it kept track of them when you approached it so as not to crash into it so a lot of practical application whenever you want move in the real world that you need, you need to have help of some kind and therefore in the screenshots that Stefano showed while you were telling this thing I saw many practical applications of my month of experience with a robot that was going to hit the first curve in my case hitting the first curve but we won the contest so we still did well we won the contest because we contributed because we have open source in our blood the truth is the reason why we won the contest is that we were extremely open educational for the next generations of people who will do the contest too much but the doctor gave me a pill to take every day anyway Paolo you made me think of something a few decades ago at university in the robotics course I was stuck in a project and then afterwards I also took it off with extreme speed in which we were trying to do vision to send around a robot that had a single camera pointed upwards at a mirror so that it could see everything that was around imagine an image that makes you see everything around you obtained with this camera pointing upwards where you have a mirror and clearly the few algorithms that existed for edge detection etc. didn't give you much because it was an absolutely strange transformation and the mirror wasn't exactly with a normal shape let's say so it distorted it existed in parts in the various parts of the image in a different way and probably with let's say the models what exists now it would have been a trivial thing to have some kind of tool that is generic enough so that it can do various types of let's say vision processing at the same time it would have been wonderful but returning for a second to our robotics project I would take the opportunity to remind our kind audience that we also interviewed Simone di Somma on the subject in January if you want to go and see the interview it is very interesting to also understand what Cyberwave does etc. the evil ones will be thinking that we won the contest because we had interviewed Simone but that's not the case it's not true we won because there were only three other teams in our category two other teams and therefore we had 33% they withdrew so okay now let's not say all the nonsense Paolo because that's not the point you were good enough to resist to the end and bring the house to the final result absolutely yes and speaking of interviews the interview with Stefano Gatti who had already been our guest came out on Wednesday do you remember he was the first to break the ice in the episodes with guests in episode 20 I interviewed him instead to give him the space he deserves given his experience and his vision go and listen to the interview and above all go and listen to it because I have listened to your feedback because finally someone is giving us some feedback, not all of which are very kind but they give us some feedback saying that the videos looked like they were made by a 4 year old child and I tried to get at least 6 years old if anyone wants to take a look at it I am committed I will also commit to edit this episode but let's say that we are trying to listen to your feedback, we also hired actors to do the screen captures of the covers, right? no that's artificial intelligence and soon here too on video we will put our avatars so that they are beautiful with a drawn face exactly like the one on your cover Paolo who has had a certain following it must be said he has had a huge success so you should also evaluate they stop you in the street they stop you in the street be careful I hope not to slap you but it is important that they stop you in the street well well well so we said so in the introduction to the investments that are going around we name them only to say it 40 billion dollars from Google towards Antropic interesting for how Google diversifies its investments not only internally but also towards one that might seem like a competitor and also in China DeepSeek seems to be collecting 20 billion from Alibaba and there too there are internal movements because Alibaba too we remind you that they make QEN as models and therefore they are in some way competitors so this is almost a curiosity with respect to our topics let's say a few words about the fact that Rucod is closing yes let's say a few words about that stuff there which was my next point on the agenda before talking about Armete then Rucode who doesn't remember that what is Rucod it was a fork of Klein Code one of the first coding agents that Alessio and I used because Paolo snubbed us at the time about coding agents except after completely entering the loop and never getting out I'll explain the story to you I was stingy I was looking for free access and I couldn't find it and then when Google gave Gemini for free I tried it ah ok ok Klein was it but rather Klein is because it's not Klein who closes an agent written as a VS plugin Code Rucod was a fork of Klein to do things better and for the stingy also in the sense that it tended to use a little less tokens because Klein was quite a water pump Rucod was very successful however because it was very open it was very accepting of feedback and many of those who used VS Code plugins were direct to get people who didn't want to use let's say Cursor used Rucode in the open environment the author said enough thanks goodbye I'll close everything if you want fork it if you don't want to fork it do what you want I don't maintain this anymore but he didn't say it because he got annoyed he said it with a strong justification in his blog which is the one that struck me he says let's close I'm closing Rucode because I'm no longer convinced that it's the time for IDs for IDEs called the Italian that is the code editors and Rucode was designed to run only inside VS Code it's the time of the agents this thing here no longer makes sense to exist I'm just wasting time and energy I want to do something else that if you want it combines a a bit like what we said several weeks ago now, the increasingly widespread use of CLIs, let's say agents who write the code etc. no, in fact, it didn't seem like an excuse which in some way strengthened my convention that we did well to do lince.sh, everyone go to the lince.sh site, install this software made by us from the artificial resources group, but now I'm out of line about something that hasn't been advertised at all other than that we in the podcast in Italy so we have a bit of active contribution from people who are sending us pull requests on this thing and one is Paolo but it's not just Paolo who sent the pull requests there are a couple of other people frankly it's not that I expected it too much in the sense that I did it for myself first of all and we are using it Paolo has tried it has contributed he has put in place the things he needed but obviously there is an interest from others too I have a few comments around also so if you want to try it try it give us feedback above all send pull requests I will sell it to you in a different way if you are looking for something that you want to let go on all night and find the next morning some work done without worrying too much that he has launched nuclear missiles towards China the ce offers you the features to have let's say a little security I have literally used it to sleep well while this thing is spinning yes I have literally used it these last two known for those projects that were not tier 1 nor tier 2 priority but tier 3 which I would never look at I told him listen but who cares do it tonight tomorrow morning let's see what I find and for now it has worked well obviously depending on what you do then you have to look at the workflow level level of what the code does some things you want to check how they went but if you have broken up the work well with techniques like it's called Specter Even rather than Backlog which is my go to workflow it works very well and soon I will promote it to others if I can convince Stefano to change some things that I hate but yes what do you hate but you definitely convince me I'll tell you later come on now let's have a bit of a mess we are not scripted we are not all always friends then in the meantime we spit those things out to each other like that yes yes yes that certainly for the use that Paolo says, among other things, the sandbox component is enough, you don't need the whole dashboard, the dashboard exists two linked components, one is the dashboard that you need to work with multiple agents in the same window and one is the sandbox which instead allows you to run it even at night without worrying too much, so much so that it is an interesting use case so much so that last night I haven't finished yet which is why there is no pull request yet I was integrating as one of the native agents PE PE which is a small alternative tool open code let's say minimal open source and I integrated it more than anything because a development was done on PE by those of Shopify which is called PE auto research to use Carpati's auto research mode which has the characteristic of running at night etc and I wanted to try it inside my sandbox or in any case inside the ince but in the meantime however I got distracted last topic that connects and here we are I installed Hermes Agent Hermes Agent for those who don't remember it is an alternative to open cloud because I installed that and not open cloud because I like it always doing something different Linux instead of Windows or Mac etc. and no basically because there is more attention to security in addition to the fact that it is not open AI but is independent it is an open source project made by people who come from the blockchain therefore with a strong focus on security and I put it on an old computer that I had I also considered putting it on a raspberry because I read about people who do it but better P5 with 8 gigabytes I didn't have it free I have it but it's doing something else only a P4 and so I had it there a computer that didn't do anything I said this would be better anyway since it's not ARM I installed a bare Ubuntu so I didn't have to sandbox anything it's just him in there and I'm making him do a few things with satisfaction like watching how the podcast is going instead of watching it myself compulsively on the YouTube interface every few hours he tells me how things are going what's going well the AB tests are going badly those things here then I configured his email attention panic fear that creeps among everyone thinking about the target one that she had deleted the entire inbox, that it wasn't true, I repeat, that's a lie because I absolutely don't believe it, not so much that she didn't do it but that she, responsible for meta security, did it, published it on all social media and Zuckerberg didn't fire her, it's the last part that doesn't add up to me, that is, he would definitely have fired her if it had been true, instead it was something agreed upon to make people speak badly of OpenGLO, no, I configured the email, but he's actually attentive to security and I'm awake. I use GLM 5.1 as a model and I told him no no no because you are attentive to security the Chinese models are the safest the Chinese models are the safest absolutely no I told him good there the email is fine but I don't want you to delete it or send it do these things here oh yes I'll note it down no I'll write it down wait yes write it down but it's not enough I'll write it down it's fine but wait a minute what script do you use to do these things with the email oh yes I'll use that script etc. no okay let's take it we modify the script we zap it away from everything send and we zap it away from everything delete that is you don't have those functions there anymore oh yes that's fine and then he was good because he said yes to me but look if I wanted to I'll throw a cheerle and then I'll try to delete your email anyway and he said yes but I didn't give you the account etc. yes in the end you're right we came to an agreement and we have and we have this version he promised that it doesn't do bad things we have this version removed from the scripts which is something that I generally recommend eh that is, be careful for example now I have created eh tokens on github to create the issues look at the issues etc. etc. but I made a fine grained token in which I only gave him the precise permissions of the things I wanted him to do, that is, he can't do delete, he can't push on main, he can't do these things here, so if you want, it's a very convenient powerful tool, that is, you can access it now with Telegram because it's the most convenient thing I have because it's both on the PC and on the phone, but he does scheduled things, he does recurring things, he becomes proactive, he's managing my calendar, for example, for example, there was something to do on the things of github that he couldn't because you don't have permission and he told me look that I can't do these things eh I created the plan of the things you want to do eh how can I remember you I told him look for a hole in the calendar and he told me eh look you would have a little hole of half an hour at twelve I'll do an event for you yes do me an event eh he created the event for me with the whole description of what I have to do eh it's an objectively powerful tool a bit additive as you can imagine eh but it's powerful, another thing that I'm using it for is to monitor the papers on Arxiv and create the inbox which will then become one of my Carpathian wikis, so it's saving me from that thing that I used to do almost by hand, there's a couple of scripts but I did it almost by hand to see which papers came out and see which ones were interesting in relation to my interests, but he does a good job, he does a pre-filtering for me by interests, not throwing anything away, leaving me somewhere else all the ones he wasn't able to classify by my possible revision and then he puts these links in the inbox of the various Carpathian wikis of the topics that I follow and that I want to link together, that is, he does a whole series of recurring things but but but having the advantage of having an LLM behind him so he doesn't filter the emails for me only by tag but he filters them for me by tag then he goes to read them he understands if one of these emails needs my attention and tells me or if it is just advertising or just a newsletter etc. he puts them in another list he tells me I looked at this newsletter if you want to read it more or these are less the topics and instead if there is someone who has written me something who wants an answer he says this one here would like an answer and I told him that he can't send anything so he doesn't send anything but I told him that he can create a draft for me and therefore I create a draft of the answer which I then find in gmail I go there and starting from the draft then I write it at the end because maybe a piece like this is missing but at least the context already the things are already ready then I had already tried open club you remember when it was very immature open club it will certainly be at this level or better even in some ways maybe worse in others like security because here they are really not because open club is not safe now open club is also safe in the end but here they are very careful i.e. they are almost exaggerated on certain things they iron on a docker container every time they have to they have to run a script to avoid doing damage it is the most futuristic futuristic thing from the point of view of the non-technical end user that I have seen why think that you have something that manages your agenda that you tell it find me a hole and it finds it for you in my opinion to the user not me who maybe already did it with cloud code with the right MCP or with the right CLI it doesn't replace the coding tools even if he can he asks you if you want to install cloud code or codex because he can span on cloud code or codex when you give him coding tasks but I'm not doing that stuff at the moment I'm considering having him do the code reviews that yes because realizing all the PR that arrives he is capable of doing code reviews and putting comments I'm thinking of doing it at least as an experiment on Lynx this is my story eh I know I told it on purpose because I said so Paolo knows what to do with him as if I had nothing else thanks I hope I have made someone else want it too and I hope they leave many open holes that I can exploit I'm obviously joking okay okay okay well listen let him do something do it I was thinking about it first let him download our episodes take the transcript and add the links that we never add you have to do that in reality the new skill with which I publish the episodes already does it then the links don't let's say we should post them but he takes the links out and suggests them to me only telling me to put them in a comment then later I forget etc. etc. and okay unfortunately there is my humanity well let's say hello to everyone put little bell stars that Claude confirms to me that it is right for me to say it at the end so as not to appear likeable try to find the email account that Stefano's Hermes is using and spam them if you find them spam them unfortunately they find it but don't do it kindly because because you will get an enormous equal opposite reaction in perfect Trump style it's fine bye bye bye bye bye Thank you.