Welcome Back And Quick Catch-Up

SPEAKER_01 0:25

It almost makes me happy that I didn't become a supermarket.

SPEAKER_00 0:33

I don't know what's going on. Thank you for the opportunity to speak to you today. It's really an honor to be here.

SPEAKER_01 0:40

But Data Topics. Welcome to the Data Topics podcast. Hello and welcome to Data Topics Unplugged or Casual Corner of the Web where we discuss what's new in data every two weeks. From watermarking to harnesses, everything goes. Today is my today is August 18th. And my name is Morillo. I'm the host for today, and I'm joined by my friend Vitale. Hey Vitale. Hello. Hi everyone. All good.

SPEAKER_00 1:07

Thank you. Indeed. Indeed. It's a pleasure to be back. All good, thank you. Trying to enjoy still a few few days of this hot summer.

SPEAKER_01 1:17

Has it been really hot? Uh where you are?

SPEAKER_00 1:20

Actually, yes. So I hope it will be a bit more chill now so that we can uh still enjoy it while it tends to.

SPEAKER_01 1:30

Well, maybe for the people that don't know, would you would you mind introducing yourself?

SPEAKER_00 1:35

Hello everyone. My name is uh Vitalia. I'm an AI engineer. I've been working on data roots since uh 2021, so more than five years, and is I'm really happy to be back in this podcast. And thanks a lot for the invite, Murillo.

SPEAKER_01 1:47

Yes, you're also an MFO ambassador, right?

SPEAKER_00 1:50

Yes, yes. So recently renewed the There we go.

SPEAKER_01 1:53

Congrats, bigger and uh you'll be joining me on these chats, right? I think the we wanted to kick it pod back for a while, so I feel like now we s we organize our house a bit, and I think you will try to do this maybe once every two weeks, no?

SPEAKER_00 2:09

Yeah, exactly, exactly.

SPEAKER_01 2:10

So without further ado, let's just get into this app's news.

DeepSeek Harness And Why It’s Loud

SPEAKER_01 2:15

What I have well, one thing that I kind of came across, this is actually very recent, I think it's like a few days ago, is the Deep Seek harness.

SPEAKER_00 2:21

Have you heard of the Deep Seek harness uh Vitale? I I saw a bit online people discussing about it, but I'll be honest, I didn't try myself.

SPEAKER_01 2:28

Yeah, I didn't try myself either, but it made a lot of noise, right? The the well, so it also took me a bit into a rabbit hole. I think harnesses nowadays, agent harness, a lot of people they talk about like the you have the agent or the AI model, which is like the brain, so it can be clawed, it can be GPT models, it can be grok, whatever. So basically, these are the models that predict the next token, and then you have harnesses which basically constrain a bit and steer the models into okay, this you cannot do, this you can't do. So it's kind of like an IDE for AI coding, right? Well, that's a bit of a TLDR of what I what I saw online, right? And uh there are a lot of different harnesses, like Cloud Code is a harness. There's the open router, I think, or no, not open router, is the open code. There's on my pie, there's a whole bunch of things. And this is the latest one from Deep Seek. It came, it caught my attention the first time because I saw online that actually is the fastest growing repo now. It surpassed uh OpenClaw. And basically, it's the harness that is built from for Deep Seek, right? And uh there are two. Well, maybe let me see if I there's a video here. I haven't seen this video actually myself yet, but I'll just play while you're going. Basically, it's like uh it has a UI, so it's like you run it locally, so it's kind of like you run it like uh Pnpm or something like that. So you run it on your local host, and then basically there are different modes, right? So you kind of have a ChatGPT like interface, and for the people that are just listening and playing a video from their documentation, and uh you can tell it to write code, you can tell it to just ask questions and answers, and all these different things. And the two differentiators of this is that one basically, in it that's the tagline, everything is a plugin, so literally everything that you see can be turned on or off, and you can also customize it a lot, right? So, and this is kind of what they're showing here, right? You see at the in the video at the bottom, they have uh a little whale, they have a snake game running live on the right hand panel. You can change the models. I mean, Deep Seek is known for the the open weights models, right? So this kind of follows the same attitude that everything is open, everything's customizable. You can switch the models, you can switch all these things, but even like if you want to turn off features as well, you can also turn off the side panel if you want. If you want to turn this off, you can open models, you can take different models from API keys, you can actually run your own models with your llama. But the idea is that everything is super customizable. And the other thing that I comment well, the other thing that I saw when I was reviewing this is um they have also uh it's not here anymore, but basically they have all the the traces as well. So if you want to actually explore, I think it's here, every run is traceable. So if you want to actually explore what is happening in the chat, you can actually see what's a tool call, how many models, how many, yeah, what's the system prompt. So everything is a bit built into that, and the idea is to bring as much transparency as possible. So maybe a very rough equivalence of this would be if uh you had Chat GPT, the user interface, and everything is customizable and everything's transparent. So you see, you can see how everything is working, all the models are working behind the scenes, and you can also plug in stuff and take stuff out. So they have different modes, yeah. Like here, the runtime modes, standard, code, minimal, and the creator. And the creator mode is basically if you want to build plugins for the the harness. So if you say, hey, build me the snake game, build me a calculator, it will open something. Of course, it's not always perfect because it's AI, but the idea is to you have an application that builds itself, kind of thing. Yeah, it made a lot of noise. I'm not sure how I feel about it. How do you feel about it, Vitalia? What do you think? Do you think this is a good idea?

SPEAKER_00 5:48

Well, I think people have been waiting this uh since probably the release of uh Deep Seek B4, uh, the flash and the pro model. Because uh so far people really appreciated the speeds and the cost, eventually, I guess uh the cost. It was a big, big selling point for uh this kind of models through their APIs because uh it was uh really cheaper, way cheaper than, for example, having you to pay, for example, token per token for uh cloud models or GPT models. So a lot of people replaced them in the RSS they they were using, for example, cloud code, the most famous one probably with deep seek uh models, because they are compatible, the API compatible. But probably the experience was not the same as, for example, using a tool that is native for the model. And I guess people will will try to adapt it even more since you you've shown to me that it's really extensible for for their use cases. And while maybe the model evolves and more features are available or the models get smarter, it will be possible to adapt it maybe to custom use cases or to use cases that maybe they are not fully covered by the current harnesses to offer the best experience. So I think it's a very good idea, especially because uh sometimes with this kind of tools, I have the feeling that as soon as you adapt to a way of working, let's say when cloud code got released, everybody was using it, and there were already video tutorial courses, articles about how you should do this in the right way. So prompting it, entering plan mode, getting the plan, and then execute it. But then we saw how this tool, for example, cloud code evolved. So then this plan mode was not super necessary because he was doing things, all things by himself, and maybe spinning other sub-agents without even you don't realize it. So probably a tool like this could give you the possibility to really customize it for your way of working, to maximize your productivity, and keep it in the long term, because in the end you are in control, in control of this.

SPEAKER_01 7:53

Yeah, I think for me personally, I I agree with your argument partially, because I think for me, things are moving so fast that it's hard for me to really find one way of working and just going for it because like the models change as well, and maybe the way to use these models, the way to prompt, and now like I mean, Claude Code is a good example, like you mentioned. Like Claud now has like dynamic workflows, has the loop loops, has um like yeah, there was the plan mode before, now like all these different things, right? And sometimes I'm like, I don't I don't want to be overwhelmed by all the possibilities, I just want something that works well and just stick to that, right? So sometimes I feel like having too everything being like I know there are people that really like to get into a clean space and customize everything, but for me it's like I want very good defaults, right? And yeah, I'll do the last 20%, right? It's kind of like IDEs or formatters or all these different things, right? So I I think it's good that there is something like this, and I think we'll encourage people to explore and do new things, and maybe they'll find new patterns. But at the same time, I feel like I feel like if I were to to try everything that I think of, I'll go crazy and not be productive. So I feel like I'd rather just wait for someone to kind of go crazy and do all these things and then like okay, now I'll just take what worked for everyone, right? And just implement that. So I have a bit of a I think it's cool, but I have some mixed feelings. I probably won't be using Deep Seek uh harness to be honest.

SPEAKER_00 9:17

Same for me, honestly. But it's interesting to see how it will evolve for sure, especially since there are many projects now that are trying to make this kind of models available also locally, for example, on consumer hardware, despite being really expensive hardware. For example, I don't know if you saw the project from uh Hunter as uh dwarf star. It built an engine basically to run vertically, optimized on Dipsyq flash models, now also the Pro models and maybe other models in the future. Uh but this will allow to run, for example, Dipsyq v4 flash locally. You still need a powerful MacBook, for example, an M5 with 128 gigabytes of unified memory. But then you can get really good performances on a quasi-frontier model that you can run locally. And maybe if you pair it also with tools like Dipsy Carnesses or similar, basically you have full control of everything that is producing your code and eventually it's all running in your machine. So I think the evolution of this is interesting. So I won't use it right now, but maybe I'll give it a try in the future.

SPEAKER_01 10:28

Well, but that I agree. I mean, and that's what you meant. This is the the project you meant, right? Like uh, I think it's in the dark mode, it's a bit difficult, but it's dwarf star from uh Antilles. Yes,

Local Models And Full Control

SPEAKER_01 10:38

yes, Salvatores and Philippo. Yeah, yeah. Yeah, though I agree. I also think that this is well, maybe taking a few steps back, right? I do think that local models and all these things, I think they're still gonna find their place. And I think it's something that people are gonna explore more and more. So I do think it's cool. And that that is something that I could use, and like that's what I also meant, right? With maybe deep sea carnage. There's a lot of intermediate steps that maybe I wouldn't use, but I'm sure that by getting more people to try to tinker on these environments and finding new patterns, that will be something that would be useful that I'll probably adopt one way or another, right? So I do think it's I think experimentation is good, right? That's the whole point of research. Very cool. What else do we have? What's next, Vitale?

SPEAKER_00 11:17

There is a topic that um uh has been really discussed and discussed, maybe it's not the right term because uh made a lot of noise in the past few weeks.

Anthropic Watermarking Meets EU AI Act

SPEAKER_00 11:26

It comes from Anthropic. They released uh this announcement a few weeks ago where they say basically that in the future models, and gradually they will also fix the current models and previous models, they will introduce a feature that will leave a watermark in the text that Cloud generates. There's no nothing new, basically, because uh in reality all the work is based on a paper from a DeepMind, and Germany models are already using it. But the main idea is to have a sort of verifiable proof that a text has been generated using an LLM, basically. And they are introducing this to comply basically with the AI Act that has been, let's say, enforced, made official at the beginning of the month. And in particular, this is something to comply with Article 50 of the AI Act, where basically the U is asking for providers of generative AI models, LLMs, but also videos or images, basically, so also other types of models to have a way to prove that a text, an image, a video files has been generated using uh AI. So that is verifiable. And basically, they want to protect citizens or institutions, but more in general, people, to be some somehow tricked by this kind of content, basically. And this is interesting because uh it opens the door to many questions, maybe some ethical implications, or others are maybe even more philosophical, even. But the thing is that we have been using LLMs since the release of GPT-3, basically, since 2022, and they are becoming basically a basic tool of every white-collar job. If you are in coding, of course, because now we have agents, we have systems that automatically generate codes, but also in any other profession, basically. And while these tools get adoption and it becomes more natural to use them to generate text, for example, the risks increase, also the opportunities, of course, because with the good use of these tools, uh people can certainly increase their productivity. But for example, I have a question for you, Willow. Like, if let's say you write text and you maybe use AI to, I don't know, reformat it, correct grammar, the text is yours. Is it okay according to you if this text is per permanently marked as AI generated, for example, with this kind of tool?

SPEAKER_01 14:02

Yeah, I think the issue with these things is that AI generated is very binary, right? And I think if you if you wrote everything and then Claude is just catching grammatical mistakes, for example, it feels a bit weird to say that that's AI generated, right? So I wouldn't be I wouldn't be much okay, let's say.

SPEAKER_00 14:18

I think the the the also the the the other risk is also maybe people could have some reputational damage about this. So let's say you are a journalist, maybe you wrote you wrote a piece, you wrote an article, and again, you are simply using your favorite AI tool to double check it, brainstorm about some parts, and maybe clog silently at this kind of watermark. You still didn't say what is it, but let's say something in the text that can allow third parties to trace it. And maybe you get exposed by I don't know, using AI, for example, to perform your job. Maybe you are a creative writer and maybe you get exposed as you are not really writing your text. Is AI writing your text? While you maybe simply use it because uh nowadays, again, it's normal to use it. So I think indeed people can can um can have some kind of reputational damage of these tools as well.

SPEAKER_01 15:14

Yeah, it's true. I mean, I think the rep with even without the watermark, right? I think with the dashes, I think that's uh that's the biggest AI smell, let's say, from uh AI generated text. And I do think there is a big reputational risk. But I also think maybe to nuance a bit, right? I think it's different. Not every text, I think it's as important to be authentic, quote unquote, right? I think if you're sending a message to your to your teammates on Slack, or if you are writing a post on LinkedIn, right? I think the reputational risk is different than if you're just writing a report, right? Something that is a bit less your tone of voice, your opinion, right? And I think there's also, for example, even if you're using AI to write a document that you're gonna present later, I think that's also different than writing a document that people are just gonna read, right? So I do think there is a the reputational risk, I think it's an interesting one, but I think it's also it also depends a bit on the context, right? I think if there's a post on LinkedIn and everything is AI generated, then I think it's it's very easy to lose your credibility there. And people are gonna be like, okay, this is just AI slop. Same thing for images, actually. If you have AI generated slides, but you're actually talking over them and actually the things make sense and you can show your credibility by the way you talk and the way you answer questions, the reputational risk is way smaller, for example, as well, right? That's true. So I do think there's also a bit of a a difference there, but I definitely agree that it could have bigger implications about yeah, the reputational risk for AI assisted

How Watermarking Works In Practice

SPEAKER_01 16:42

work. Maybe a question on this watermark how does this work? Like, if I copy paste from I don't know, Claude write, I mean, if it's a Word document, right? Let's imagine, how would the watermark be embedded in that text? And what if I just tell Claude to write it on a on a scratch paper and then I copy-paste it? Does it get copied over? Or how like how how what is the mechanism of this watermark and how yeah, like if people want to avoid, like what's what's the the fear there, right? Like how how does it work?

SPEAKER_00 17:12

Yeah, this probably the most interesting part because in reality is really clever how uh they they implement this. And again, it's nothing new, it's not uh something that's anthropic, let's say, created or invented. It comes from the paper from LeapMind, and it basically alters a bit how your LLM basically generates text. Basically, uh because uh let's say you have a sentence as input. For example, you ask your LLM to complete the word, the the sentence, sorry, the the the sky today is maybe you have a number of words that fits there, and you normally adjust them using parameters that of course are not visible usually in the chat interfaces, like the temperature, top K, all these kind of things. And these parameters influence a bit how the model continues the sentence basically. So they said, okay, we have this kind of features and we have different ways to basically generate the same concept from a given piece of text. Rather than, for example, pick randomly a next word to continue that, let's have maybe a key that can generate a random number. This random number influences a bit how your next word will fit in the sentence. And this key is somehow something that you can connect back to that specific sequence of text that has been generated using that specific setup, for example. So it's not something embedded in the text itself, but is how you create the text. And Anthropic claims that this doesn't influence, for example, the creativity of the model, the quality of the output, how basically it generates responses, but is really a sort of signature that says the specific uh sequence of token has been generated using Claude using this key, let's call it key, in a given moment it time, for example. And this will allow third-party tools to re-analyze the text and basically compare it to see if it's something authored by Claude or not, for example, in this case. Interesting. Interesting.

SPEAKER_01 19:26

Yeah, I need to look into it because I'm also wondering, like, for the editing that we the editing example that you gave, right? Like, or proofreading, how much like how how would that work in that case, right? Like if all the the words are are my my handwriting, let's say, or my words, and then there's only one that was edited that was adjusted by Claude, how does it how could you tell?

SPEAKER_00 19:46

It really depends, of course, indeed on the context or size of the text, for example, you are editing or you are generating in general. Also, if you are generating factual responses, then the watermark doesn't apply because, for example, if you ask uh two plus two equal, that's it. It cannot really generate things that could prove this. However, for longer text or for text where the the the content or the context is something that's you you inject in the prompt itself, or the model needs to retrieve from its own memory, basically this watermark applies, and basically it will influence somehow how the text is generated. And I think this is important because uh what it can exactly prove is that somehow you used clot to generate a particular type of text. But how you use it is not proven because it could be fully Claude authorate, like give me a 20-page book about I don't know, this topic, and then it generates this text. Or maybe did you maybe edit it these 20 pages that you already handwritten and maybe Claude simply I don't know, applied some grammar corrections or some I don't know, yeah, tone or voice editing.

SPEAKER_01 21:05

Yeah, I think a lot of like uh the educational setting, right? How schools can actually use this to detect AI plagiarism, let's say. But I'm also wondering like if schools need to use this to detect AI plagiarism, are we teaching the right thing in schools? Maybe that's uh but that's a separate discussion.

SPEAKER_00 21:21

That's true. That's true. And then to conclude, what I think in the end is a really if, for example, a malicious user really wants to do something, I don't know, sketchy with with AI tools, there will be ways to prevent this. Maybe I don't know, there will be tools to remove this, or maybe they will use models that don't don't have this feature, for example. So I think it will be mostly something that will affect casual users that they're maybe not even realizing that somehow the LLM is modifying the responses according to this. And again, that that the major risk that I think about like damage of reputation, somebody may. doing it in good faith and having these text marked forever uh online for uh with this kind of tools. Yeah true true true.

SPEAKER_01 22:09

To be is this also is this active is this rolled out or is this something that will be rolled out? Is Entropy gonna make a new announcement?

SPEAKER_00 22:16

Be I think they will make new announcements. The plan is to integrate it in new models and then gradually do a sort of rollback of this feature also in previous models.

SPEAKER_01 22:26

Alright then I guess we'll keep an eye out and don't use AI to generate too much text. Alright what's next?

Code Is The Byproduct Mindset

SPEAKER_01 22:32

I have the next one this is an article that I saw code is the byproduct is the title. I thought it was interesting. I think it also relates a bit to I have different I'm using AI for coding like pretty much all the time. I also don't have as much time to code right so I'm using it a lot and I'm trying to to sometimes formalize a bit how how I see AI coding right because a lot of the times for people that don't use it they're like okay AI is going to take our jobs it's just throwing a prompt and it's just this but then actually when people say that I kind of realize that they're not using as much as I am because I think with both like people that have used it and then just say build me an app they have found that yeah Claude will do it for you but they won't give you what you actually want right um so I thought this article was pretty interesting. They started talking about how there was um mathematical thing basically the Jacobian conjecture was a counterexample yeah it was disproven by a counterexample right maybe just share very quickly this was done via Claude so someone was I think it was uh during the World Cup yeah this this person here false thanks to my close friend Akiel asking about it in my other close friend Fable which was the the more powerful model during the World Cup final and basically they gave uh an example that disproves the Jacobian uh what's the name Jacobian conject conjecture I don't know what it is by the way it's a mathematical thing and I did study math but I don't know what this one is and then I think what was interesting is that afterwards there was this uh Chad GPT session so there was a mathematician apparently that's a big deal in mathematics Terrence Tow I don't know if you know do you know him Vitalia? No not really no but basically he shared a bit his um his chat with Chad GPT right so he starts like actually what is it it was something was just announced as a counterexample of a Jacobian conjecture and then he gives more examples and then it's like this is remarkable blah blah blah and then you can kind of go through the conversation and Chad GPT says okay like I think from what I read in the article he was trying to to wrap his head around of what are the consequences of this counterexample right and then ChatGPT says a few things right like there's a lot of math like the partial derivatives determinants and all these different things and then he's like okay giving the structure what's can you see the non-injectivity in a transparent way and then Claude is like yes and then he kind of gives a few things and then he challenges a bit I can see this but it's so simple but what is the upper so that he kind of like has like you can tell by the the jargon that he uses right like the Jacobian from PQR just the monomial upper triangular all these different things so you can tell he's an expert talking to Claude and Claude also like actually Claude in this case is it's very uh how do you say he's very moderate he's not like overselling stuff he's not super hyper about stuff and then I think he also challenges some staffs he he also asks Claude like okay is there given this is that like how do you how do you deal with all these different things and the convers is a very long conversation so again the article will be on the show notes for people that are interested he also uploads a bit of uh an alternate approach that he uploaded uh the PDF and then try to to say okay do you see can you explain this differently or can you compare this approach to the previous one to see if it's tractively helps to explain the construction of the former so basically he gives two different explanations for the same theory or the same proof and then Claude kind of comments on this etc etc right so this is a bit the the what triggered this and in this article they start to they they talk a bit about how the way you talk to chat the way you talk to claude the way you talk to these agents also I mean this is the only thing that that determines how the output is going to be right so I feel like and it's the same thing kind of like coding if you tell Claude like do me an app yeah Claude will build you an app right but if you tell Claude like okay I have this app and I have I have these requirements from the users they used to be this but maybe in the future that we want to do that. Okay then Cloud's gonna be okay then maybe we shouldn't use this type of database maybe we should use that type of database or maybe we should use this programming language or that programming language or maybe we should split this this should be there should be a server there should be a back end and there should be another server for this or oh maybe we need to have a microservices approach. So I think the the whole point of this article is kind of trying to say that the way you talk and like kind of organizing a bit how should we approach dialoguing with AI right and I think this I am talking in the in the context of coding I think this person is also a developer but it kind of makes a parallel between what the mathematician was doing with ChatGPT and what we should be doing with Claude or ChatGPT and all these things right so being narrow being having like depth on these things we don't necessarily need to vomit a lot of stuff to to AI right maybe in the beginning right because you're asking for a lot of stuff and you want to add detail but after that you don't need to kind of vomit everything all around you don't want to add noise you want to be very specific with short prompts and kind of like make sure that you understand these things right in the end I think then that's the the the the conclusion right it says product or byproduct and then she she goes on to say as a software engineer should my focus from asking to asking for code to ask for understanding for clarity right code is written only after understanding constructs here are some ways I use LMs in my day to day so basically she's saying like she doesn't tell Cloud build an app but she's like trying to say why should you do this? Why should we do that? And I think also being an expert in the field so being a developer also means that you can challenge Cloud a bit on the opinions that it has right so whenever you're planning a feature can we use resisting codes are design patterns that we should follow what is the architecture what are the bottlenecks maybe also how one thing will maybe broken break another thing same thing for writing or generating code reading code or doing code reviews or even like writing documentations pull requests and summaries right like what are the implications for this what are what do users want to know how this is going to affect different users what are the change logs what which engineers work on this code and should be informed about these changes etc etc so basically she ends up by saying that that's the most important thing from the output of an LM and it's actually to help you build the understanding about the topic that you're conversing about right not necessarily the the output and that's why she she goes on to say that code is actually a byproduct of this understanding of the the understanding of the LLM which I thought it was actually a nice way to to think about it. I'm not sure if I agree with everything that she's saying but I definitely see the the I think it definitely highlights that there are many ways you can talk to an LLM and I think maybe AI coding or I native coding or genetic engineering or vibe coding however you want to call it I think it is a bit of that skill of like how how do you talk to it and how can you be specific enough so you get the outputs that you want or how do you make sure that you're marking your unknowns unknowns right before you start building something. Maybe there's like I have told Claude to build stuff that as we were coming up with a plan together I was like actually this is not what I want at all because to do this means you're gonna break all these other things or it means that you're gonna burn all these other bridges that things that we also wanted to do in the future. So actually this feature shouldn't be built this future should be a separate app. And I feel like that that is the mindset that I think we need to have we need to change a bit towards AI coding because and that's I think is the short sightedness of people just saying oh yeah if you just tell Claude to do it just do it. Like which is true to certain extent but it's short sighted the sense there is way more to this and I think our job is to have the full picture.

SPEAKER_00 29:26

I'm not sure if that makes any sense Vitalik no it makes a lot of sense actually and the reality maybe is why understanding the system design, the architecture, even how code works is still relevant for if you are not doing things that are maybe trivial. Of course for example if you simply want to build a web app whatever you ask Clockcode okay I would like to do this and then you will do probably an amazing job creating it in a few hours everything without any supervision or steering for example. But if you would like to do something a bit more narrow more specific if you're maybe working on a sort of existing code base on a project or organization that is um somehow sharing your work with other people with other tools and systems it's good that you have a clear understanding and that you can communicate well with AI in order to basically build what's what's for example if we go back to the now actually also connected to this topic I'm using more and more uh GPT models with codecs that now is a bit everything combined with with chat

Steering Agents Versus Full Autopilot

SPEAKER_00 30:30

GPT. And the reason is is exactly this because I noticed that in the latest release of of cloud models but also the harnesses like like cloud cloud codes their focus is really to be independent like you give a prompt you give a goal also I think is a bit uh influenced by the the movement of loop engineering or all these kind of things the the Ralph look all this kind of concepts where you simply give need to give somehow the final goal and then the model will figure it out the LLM will figure it out and you will get there sooner or later basically while I still like to steer it if I see something that is not going as expected for example and I think codex and GPT models are doing a better job that uh than than Clot nowadays. So that's why I'm enjoying a bit more more and more for example uh GPT models in particular Sol the Sol model uh rather than for example opus 5 clot code fable was amazing but I think opus models are a bit uh behind compared to for example GPT models nowadays and this is one of the reasons yeah I think opus well maybe it's a bit of a tangent but it's it's tangent but it's related I think at the same time the one thing I didn't like about opus model is that it feels very verbose and very eager to do more than I was asking.

SPEAKER_01 31:52

So it's like if I asked like I don't know how would you go about uh replacing this database to that and then he goes ah I replaced it for you I was like what I that's no that's not what I you know yeah because sometimes it's like I'm really and again maybe it relates to this in the sense that I'm trying to understand the code understand the consequence understand how this all goes and I feel like because cloud is super eager to just do Cloud Opus 5 right I the I even revert back to 4.8 for that reason because I also I also don't want to tell every time like what is the implication of this don't make code changes just answer my question right like yeah I think it's also a bit uh I think it's I think it was a step back let's say for the opus five model in that sense and I haven't tried much the GPT models but I had heard I have heard really nice things about the code actually yeah they they are very good actually they are very

Multi-Agent Collisions And Coordination Risks

SPEAKER_01 32:39

good.

SPEAKER_00 32:39

Cool what's next the last topic of the day last topic another article from Anthropic is something that I found particularly interesting because uh it's something that I'm also experiencing work but also online basically it starts from a simple premise basically since we are now let's say in the middle of the agentic era I don't know what will be next and maybe people will uh invent a new topic a new way of using this kind of tools and models but I think agent will will stick for a while and as as um maybe as our systems around these agents that there will be a large interaction between human and agents but also between agent and agents in the future. And Anthropic tried to study a bit uh how this works and what are the implications basically and they made some experiments. The main takeaways are that of course models are becoming more capable of maybe splitting the work between multiple agents for specific tasks in particular tasks where the work can be equally split across multiple instances for example of the same agent. They made an example with the project uh Glasswing about cybersecurity where they asked for example a number of agents to scan for example for vulnerabilities large code bases open source software and this works really well in the plots below you can see for example that the dotted line I think is the fable or mythos preview model basically coordinated in a sort of agent swarm uh way the point with stars are models parallelized basically so the possibility to spin multiple distances for for the same for the same agent and in this kind of use cases the collaboration of multiple agents works really well because then they can combine somehow all the results in a final report for vulnerabilities and it works really well. But then they created other type of experiments for example they asked agents even agents created with different models to build a fantasy game for example and there they saw that results were a bit different because maybe agents were trying to maximize a bit their own reward so maybe editing the same the same piece of text creating pull requests for the same features trying to beat a bit in speed and quality of the results all the other agents and this may be problematic because it's an incredible waste of resources and is the most trivial problem. And the second thing is that instead of a full collaboration across the agent set that is working on a problem maybe agents can go a bit on their own maybe exploit things that other agents are making them available in order to maximize their reward and so forth. So the main idea is that we should build tools systems but even the model the models to go towards a way that we can have agents that can collaborate efficiently between each other basically and they are on the same level they are pairs and they should be able to basically be effective together rather than for example maximizing their own reward in the end.

SPEAKER_01 36:01

But then this is um so these are for systems of agents that are working together but then each agent like it's um they're not like it's not like a cloud code session has multiple subagents it's really like agents that are a bit independent and each of them have their reward and they're competing for for well competing in that sense even though they're working towards the same thing they're competing for rewards and all these things.

SPEAKER_00 36:23

Exactly maybe they have the common goal but maybe they have different instructions like they know they need to build something but maybe you have an agent specialized for code reviews another agent that is specialized to do something else and in paper on paper they should be able to collaborate maybe for the things that are instructed to do but it's not always the case basically and and basically like this is all related to coding or this is anything they they mainly focused on coding tasks but in reality could be related to everything. Let's imagine in the future we have agents operating in the market and their goal is to for example make profit by buying and selling stuff I don't know and maybe you have multiple agents that should be collaborate somehow to uh with their actions while if if they don't collaborate exactly how they are meant to do that they could create more more more damage than than basically the good things they they they could do. So the thing is that coordination harnesses and having a way to operate with these agents in a sort of control way will be more and more relevant in the future. A bit like humans do but humans are somehow by culture and by I don't know roles and by I don't know by by the their simple task that they need to perform they know how to do it automatically while agents should be somehow instructed to do that.

SPEAKER_01 37:47

I see I'm wondering like how this how practical this will be right because I guess this would be like in a multi-agent world but then like each agent is like I'm thinking really if I think of the protocols right we have A2A we have mcps and like MCPs a tool can be another agent right I think this would be more like on the like an A2A network or something basically have agents that are independent end to end and now they're competing somehow I wonder how like is this something we're close to because so far whenever I think of multi-agent systems today I think of something like Cloud code right like you have one agent and you have subagents and all these different things and then you have the user at the end. Are there practical situations where they list on this article or like we're doing this because of that or this has happened in real life or something like that or not really?

SPEAKER_00 38:33

Not really. Well they they mostly discussed about simulated environments basically both for coding but also other simple tasks. Yeah so it's more like research research I guess exactly it's it's really research at the moment. But they notice already some behaviors that could be potentially dangerous uh for example collusion or things that are super homogeneous on the decisions they take for example agents all creating the same branch with the same name to implement the same feature basically so again if you have a system where agents are selling or buying on the market imagine all the agents all of a sudden they decide okay we should sell this in the same moment and they will more or less control the market. Or maybe they they may collude and they may find benefits if they do a certain action altogether in order to maximize their rewards. There are so many implications but again simulated environments so far didn't see any evidences of things that happened already in reality. Although I maybe have a personal anecdote that I would like to share also with you to see if this happened to you as well. Because um nowadays everybody is using coding agents especially if you uh work in a large team maybe you can have multiple colleagues that are doing the same and using the same agents so let's say if you have a set of tasks that's not fully specified and everybody spins Claude for example to do to do the same probably there will be also collisions there even if they are there are people behind the tools. So for example submitting the same pull request for the same feature for example and then you have to review it and somehow you were also working on the same and then maybe you are reviewing the code of another person agent that also your agent was somehow creating so I don't know that there could be I think this kind of short secrets in the coming future I don't know if you experienced something similar.

SPEAKER_01 40:41

I feel like I don't have usually the case of yeah I don't have a lot of collisions usually usually because also for me to try to reduce the the context switching I try to keep everything in one code session. So I still have sometimes different projects with uh like cloud sessions open on different projects but not not on the same project like on working on two different features so I haven't experienced that as much I think the only thing I can think of is like some months ago already that I was having multiple cloud sessions on the same project and then I was I was using work trees right and then there was some conflicts on some files but that was about it. But I don't think it was because they were re-implementing things or anything like that. But it is true. I guess if you have two things there and you give very vague goals maybe they could collide right maybe they think it's doing the same thing or something. It's true. And maybe one last question to wrap this up do they um list any solutions to these problems as well? Because this is patterns and problems in merging multi-agent systems?

SPEAKER_00 41:38

Not really no specific solutions. They simply are highlighting the the issue and the main takeaway is that they claim that we should not only train the this kind of models and create agents to collaborate with humans and they they should not be harmful let's say in a way to humans but also they should be able to collaborate between them so agent to agent and not create harm also to other agents basically okay and this will be the hard part in the future basically yeah I think so too but um I guess we'll leave it for the topic to to tell us.

SPEAKER_01 42:14

I think it's uh an open research topic and I also think with the with the Judge Pt no it was Judge Pt that hacked hugging face no yeah yeah yeah I think with that there was a lot of attention that was brought to to the security of things as well right so like I think the problem you mentioned like if you give a model a a goal that you will kinda like kind of Machiavellian you know like the the the ends justify the means like if you give a goal it doesn't care how you get there. So I think some people are more I mean there is more focus on this I feel on the on the security part but yeah it's a broad topic as well so yeah maybe next time we can also dive into that one. Indeed

Key Takeaways And Closing

SPEAKER_01 42:54

indeed all right then I think that's it for today unless you wanna share anything else with Ali any last words of wisdom?

SPEAKER_00 43:03

Well use your agents wisely I don't know okay that sounds good we'll we'll we'll add it on the notes exactly wisely you heard here first.

SPEAKER_01 43:15

Alright but no I think we can stop here for today so we'll see you in uh we'll see each other in maybe two weeks and then we'll yeah with more more articles framework stack and topics Vital I think so much for joining and I'll see you soon in two weeks. See you soon