Skip to content

Podcast

SED News- OpenClaw Goes Viral, Mistral’s Compute Play, and the Agent Arms Race

Software Engineering Daily

Source ↗ ← All highlights
  • OpenClaw Exposed Agent Power And Peril
    • OpenClaw revealed the power and risks of local, self-hosted autonomous agents that access files, apps, and API keys.
    • Peter Steinberger shipped an open-source agent quickly, sparking wide security concern and mass adoption interest before joining OpenAI. Transcript: Gregor Vand This has been obviously quite a whirlwind month for the founder, Peter Steinberger, an Austrian guy who had put together what is now known as OpenClaw. It was ClaudeBot. Sean Falconer Yeah, ClaudeBot, but not like Anthropic’s Claude. Gregor Vand No, but even so, Claude or Anthropic took issue with that, which is then it changed to MaltBot, I think, and then it’s OpenClaw, like with a sort of like lobstery icon around that. But yeah, I guess the headline is that he’s actually now joined OpenAI. So you had Anthropic telling him cease and desist on the name, and then you have OpenAI saying, hey, come join us. So yeah, what did you make of that, Sean? Sean Falconer Yeah, I mean, for anybody that maybe really has had their head in the sand for the last few weeks, this really blew up, I think, in the last probably like month or so. But in terms of like what OpenClaw is, it’s an open source, like self-hosted autonomous AI agent that runs on your local machine that you allow it to interact directly with like your Files, your applications, web services, whatever things that you want to grant it access to. You give it your API keys and so forth. And it acts like a 24-7 personal assistant uses LLMs and skills to kind of automate these different tasks for managing your calendar, sending messages, writing code. It has hooks into being able to react, you know, have conversations over WhatsApp or teleport or whatever kind of messaging interface you want. So you could be on your phone and go tell OpenClaw to send an email or to go off and manage your calendar. I remember years ago when I was just a young engineer out there, there was a guy that I worked with that kind of proposed this idea of using email as like a transport mechanism where you Could send in almost like a bash command over email. And then you would have like something that would listen to that and invoke the bash command. So it’s like, oh, I want to know what files are on my computer at home. And you would just send it over email. And the idea was everyone knows how to use email, so let’s just do that. And I was like, that’s a cool idea, but this seems extremely dangerous to do. So we never, ever get anything with it. But this is basically the agent version of that with crazy superpowers. So at the same time, I think enough people obviously found value in it to actually install it and use it. And I think it created so much buzz. And I think part of it’s almost like you can’t turn away from an accident or something like that you’re watching, like a train wreck. It’s doing a lot of things that everybody says not to do when it comes to AI or all the scary things. Gregor Vand Just basically open this up to your local file system, basically. Sean Falconer Yeah. So of course, like security people freaked out about it. I’m sure like given your background, it was probably really scary. And then there was also, I think some things where some of the API keys, people were able to like sniff out other people’s API keys. And then on top of that, they launched Maltbook, which was a social media platform created specifically for AI agents, like interact and post and comment. Essentially they use the language of like humans can only observe and that i think also plays into a lot of people’s like fear around everything that’s going on with ai and there’s a ton Gregor Vand Of stuff that caught the attention of everybody in the space basically there were sort of like articles on there like my human has asked me to write this like yeah So it’s sort of like watching A bunch of agents, computers, talk to each other. It’s kind of what that was meant to be. And as you say, it was almost like knowing that that’s going to really stir the pot a bit to have that kind of stuff. Yeah. Sean Falconer Yeah. And then even like Peter, who had had a successful company in the past and basically came out of retirement because of everything that was going in AI and eventually led to this was he Came out and he would say like I pushed the prod without like reviewing the code which again is like another thing that people have a lot of fear over so I think all these things in combination Really created a lot of fascination and discussion around it but I think the outcome is like incredible I certainly didn’t see this that he would sell this to open AI and now be at OpenAI. And between, I think, what he’s getting cash-wise and stock, is he the first sort of one-person billion-dollar company? Gregor Vand Yeah, that’s interesting. We sort of ignored that topic for a while, but yeah, this could be that. Yeah, the push to prods without looking at it. Well, I can believe that because I should probably say, disclosure, I work at Superbase. He used Superbase for the backend. Oh, right, yes. And did not turn on row level security. So as usual, Superbase got the rap for that. Why can we see a database, blah, blah, blah. And as our usual line on that is, turn on RLS. It’s very simple. Just turn it on. So yeah, pushing to prod without having reviewed anything yes that’s i guess what happens unfortunately but i mean great use of technology i mean i think we’re going to get into why this Over something like clod or like clod code etc but yeah i think this has been a very interesting wake up for like how fast things can still move. (Time 0:05:39)
  • Enable Row Level Security Before Production
    • Turn on row-level security (RLS) and validate prod deployments before exposing databases.
    • Gregor Vand noted OpenClaw used Supabase without RLS which exposed data — enabling RLS would have prevented common leaks. Transcript: Gregor Vand Well, I can believe that because I should probably say, disclosure, I work at Superbase. He used Superbase for the backend. Oh, right, yes. And did not turn on row level security. So as usual, Superbase got the rap for that. Why can we see a database, blah, blah, blah. And as our usual line on that is, turn on RLS. It’s very simple. Just turn it on. (Time 0:09:52)
  • Ads And Enterprise Funding The Model Arms Race
    • Chat platforms will monetize via ads and B2B to fund ever-growing model and infrastructure costs.
    • OpenAI is testing ads and charging high CPMs as token/inference spending pressures push new revenue models. Transcript: Gregor Vand So it’s something that we kind of touched on, I think, like a few SED news ago, like how is this all going to converge? Taking eyeball share away from google especially and like then you have perplexity doing their browser thing which i’ve not heard a lot about recently so who knows where that’s gone Again that was like a whole thing about eyeball share and i think it’s fair to say like chat gpt is still the most widely used i mean i don’t think in maybe our circles i feel like claude i Sean Falconer Think in the b2c space is probably the case but i don’t know the stats for sure yeah exactly but yeah i feel like claude maybe is seen as more the professional’s choice especially in tech These days but certainly gemini from google is on the yes yeah and then it’s just there if you have the google products yeah but i don’t often hear people going oh chat gpt blew me away especially Gregor Vand Not in like the work context. So yeah, it seems more like that’s just the one that most people have in their head. So that’s kind of where they go to. And yeah, they now seem to be rolling out ads within the platform. And I think the TLDR of why is just money. I mean, of course it’s money, but it’s like they are literally not running out of money exactly. But we’ve also covered this that they’ve got all these commitments and there were some funny things going on in the last couple of weeks where nvidia is not going to invest something Into open ai that they originally said they were basically and that then had a knock on effect to oracle because oracle was going to be supplying the inference for that and where this Is going is that oracle basically made a statement saying this doesn’t affect us. And the fact that they even made a statement means it probably does affect them. It’s basically the commentary there. So why am I talking about this? Because money at OpenAI is a huge question mark. And it seems like ads is where they’re going to plug the gap. Sean Falconer I mean, I think this is something that we probably predicted at some point, but not to pat ourselves on the back too much. I think it’s kind of like an obvious thing that probably a lot of people were predicting for at least one of these chat LM services to eventually explore. And if you’re in the B2C space, like classically, how do you monetize consumer, especially when the value that you’re providing is really about like eyeball capture. It can be a game, it can be social network, it can be search. All those things have been monetized largely through serving ads. So it makes sense to serve ads here. Now, the big thing is, and I think we also talked about this, is like, does this signal the end of the heyday of LLM Power Chat? Because just like we’re in sort of this utopia world with video streaming at one point when you could essentially go to Netflix years ago and Netflix had everything. You could just stream it. It was very cheap. And then all the providers of that IP eventually woke up and they’re like, you know, why are we giving our IP to Netflix? We should go and create our own service. And then you ended up with now everybody, like you got Disney, you got Peacock, like everybody has a service. And to cover the breadth of it, you end up with like 10 of these things. And basically we’re back to like cable television, except it’s like on demand. So we’re certainly out of the heyday of video streaming. And I wonder if this is the start of that for, at least from like a consumer’s perspective is like, how do you get value out of the consumer and i don’t think charging consumers like a monthly Fee is probably something that necessarily works it’s like how do you control essentially the rate limit on that at a value of ten dollars a month or whatever and token cost is going down Certainly a speed of tokens is in infrastructure costs are going down but at the time, they also need to generate money to invest in new infrastructure to train the next generation model, Which is very, very expensive to do. (Time 0:12:20)
  • Alibaba Used Free Goods To Kickstart Chatbot Commerce
    • Alibaba used $430M in Lunar New Year incentives to drive 120M orders through its Quen chatbot over six days.
    • The campaign gave free bubble tea and goods to nudge users into making their first chatbot-mediated food orders. Transcript: Sean Falconer We’re going with a lot of this stuff. People already use these tools to go and ask for advice on purchases that they’re making, and then they go to some other place to go and make that purchase. But it doesn’t take that much foresight to figure out, okay, well, if we just connect the lines a little bit here. And there’s been investments from Google in terms of standards around doing secure purchases through agents and so forth. So I think that’s not that far away where this is sort of the default. Gregor Vand Yeah, this is an interesting one. Because yeah, as this campaign was around, yeah, ordering food. So if you sort of play that through, okay, a user maybe hasn’t realized or just hasn’t tried to order food through the chatbot. And so yeah, they offer this incentive. And then once the user’s done it once, they’re like, oh, this is possible. You assume that maybe half at least are like, oh, this is actually nicer than going through my food apps. (Time 0:25:30)
  • Model Providers Acquire Infra To Build Moats
    • Model providers are buying infra to build end-to-end moats that raise switching costs.
    • Mistral acquired Koyeb to add inference and app hosting, letting models spin up containers and databases for customers. Transcript: Gregor Vand So the French model provider, but you know, has Mistral Compute as well. But they’ve bought a company, I think you pronounce it Koyeb. So K-O And basically Koyeb provides inference. And I think there’s kind of two sides to the story here. One is that Koyeb provides inference and can like bolster the compute side of the Mistral models compute. But then Koyeb also provides spinning up like databases or like containerized apps and this kind of thing. As I understand it behind the scenes, that’s actually Neon. So this is not Koyab’s own infra providing that. But I think it looks like the Mistral still sees huge value in that, even if it’s based upon like you’re two steps away from who’s actually providing the infra but we’re seeing this a lot I would say like we’re seeing a lot now where and this may be a topic for a movie next time Sean but where do people like Lovable go in the future if we’re seeing people like Mistral probably Move in this direction where from the model you could spin up infra and like i.e. End-to create an app i think we’re going to be seeing a little bit of a step shift this year across all the players i think actually yeah yeah i mean i think that it seems like all the model Sean Falconer Providers are making steps to create more of a moat around their services by owning more of sort of the end to end pipeline like making more of these things just sort of part of the model Experience. Because if you just have a model, then it’s kind of like, well, the switching cost is low, I can go with the best model or the one that’s like the best balancing cost for whatever my workload Happens to be. But if suddenly the model also has all these other things that own, maybe holds on to important contextual data for my business or for my personally, or it makes it really easy for me to Just like spin up agents and run them in cloud infrastructure. And I don’t have to go use some other application to do that. Then why go somewhere else, right? You’re creating essentially one experience that kind of serves your end to end purposes. Yeah, exactly. Gregor Vand Yeah. So very kind of interesting. It’s probably the first, I guess, pure acquisition we’ve seen in this space so far. A model provider outright buying the infra provider, if you like. I mean, Lovable, we reported on Lovable buying, the name of ASEMI now, but Lovable being Swedish, they also bought a Swedish, Molnet, that’s the company. They bought Malnet, who smaller outfit, three guys there, whereas Koyab, I think, is 13 and just has like clearly bigger customers and more sort of advanced offerings. So less of an acqui-hire here, I think, and more of a like full rollout team and infrastructure, which is super interesting. (Time 0:27:06)
  • Agentic Engineering Compresses Idea To POC
    • Agentic engineering compresses idea-to-POC timelines by combining models with tool use and multi-agent orchestration.
    • Tools like Claude Code, Codex and frameworks like Gastown let agents run tasks, access files, and collaborate in parallel. Transcript: Gregor Vand So obviously OpenClaw has kind of brought things to a huge head on that one but you know if we look at anthropics like opus 4.6 and you know we’ve got gpt codecs like basically it’s even Coming back into mainstream media like that coding is being disrupted which i found quite amusing because it’s like why why suddenly are all these mainstream media the reference i Was i was going to bring out was just my dad literally messaged me and said like have you seen this and i’m like why this is so why is my dad who retired not in tech why is he suddenly reading Things that are telling him that like coding is being disrupted you know i always look for these kind of like bellwethers or like canary in the coal mine kind of thing like what is the thing That’s suddenly making people wake up to this? And it does seem that agentic coding is why. Basically, it’s showing that it’s not just ask code to be produced and code comes back, but suddenly all these other tasks are being completed in the background that is suddenly taking Over developers’ workflow completely, basically. Sean Falconer Yeah, I mean, think that there’s probably a couple things that have gone into this but definitely feels like in the last four months what the developer experience is has changed drastically And i think that’s it’s gotten to the point where the tools plus the models are so good that even the naysayers are starting to realize that this is like undeniable what you can do with These things. Because when you were doing kind of like one shot, like you sent a prompt to a model and you’re like, hey, generate this function for me or take this function and help me like optimize this Stuff. Like that’s helpful. But I think you’re seeing, you know, 20, 30 percent, maybe efficiency improvements there or something like that. Whereas now with the agentic engineering using things like cloud code or using cursor running codex 5.3 from OpenAI or whatever it is, you’re able to take stuff that would typically Might take you hours to do and do it one shot prompt and they’re just able to do it it’s it’s really incredible having uh you know spent some time using these tools myself in the last few Weeks like i think it’s it’s it purely there is some you know really a transformational shift that’s happening and then on top of that so you know we went from sort of co-pilots in assistance Where it’s really like auto complete you, you know, you’re, you’re kind of one shot to now energetic engineering. And I think the evolution of that is now like these, like Ralph Ligam loops and Gastown and sort of multi-agent orchestration. And there’s some debate over like how ready those are, but people are really on the bleeding edge are doing that. And to use like Steve Yegg’s language, it’s like agents really take chat and basically put chat in the loop and then these things like ralph wiggum gastown and so forth are putting like Kind of agents of the loop where you have multiple agents running and you see these people you know talking about their setups online where they’re they got like six mac minis running And like on their uh desk all the time so they can spin up more instances of quad code so they can always have an agent running. And so I think that the combination of the model and the tools getting to the point where they’re really good at managing contacts and probably the explosion of the number of MCP servers That are available and things becoming more like easier for agents to sort of navigate. Clearly, there’s to me there’s just this transformational shift that’s happening and i think that puts the entire industry in a place where there’s probably some uncomfortable questions Gregor Vand We need to be kind of asking ourselves so exactly i mean just to kind of i guess semi-summarize to this point the fact that models can use tools at all is already sort of quite a big step change And and use them reliably yeah exactly and you and use them reliably and i mean it still depends kind of on your interface to that like i’ve been pretty impressed with actually notion Ai recently because it just has the the tooling and the access to you know files and slack and all this kind of stuff but so models i mean when we say models can use tools you know the service That is built upon a model effectively can use the tools you then got you know multi-agent orchestration so the fact you can sub-agents working in parallel and like as you were kind of Alluding to like we’ve got these frameworks like gas town now where like you know you have kind of like a lead agent effectively. And I love Gastown’s way of like a mayor and you’ve got, which other people do they have? They have like, well, the other mayor, the town, rigs, crew members. I mean, I think it’s really cool. Just a great way to like add a way for sort of people to kind of get their heads around what’s going on. And then something I’ve noticed just without any thinking of this previously, but, idea of context compaction where AI can actually summarize, basically it’s on memory, and then Can keep working on things without hitting limits. We’re beyond the stage, at least I feel on any paid plan now, we’re on the stage now, where you don’t really, you very seldom hit like, oh, I’ve run the course of my conversation. It’s like, you can always add to a chat that you’ve had previously, but it’s able to do things with that. And then, you know, you start off with say a new chat within, in my case, Claude. And I did get a bit of a like, oh, that’s scary. Like moment when it’s like, I’m asking something and it’s like, and so based on this project to do with X that you’re already working on like i was like oh no it knows it it has all the context About me so i mean it was good but it was just a slightly strange moment when i was like ah it does reference back to these other things i’ve been working on okay understood yeah i think like Sean Falconer One of the big things that we get out of this is and maybe we’re not a hundred percent there but I feel like it’s certainly going in that way, is we’re really compressing this time from idea, From idea to at least POC. And we’ll get to the point where it’s like, you’re really compressing the timeline from idea to production as well, where essentially the cost of experimenting, it becomes almost Zero, other than the token cost. But like I could just tell Claude to go do something and, you know, one shot or maybe a couple of shots, like it’s actually able to like accomplish that thing. And I think the interesting thing from my perspective is if we are really taking this thing that, you know, historically has been an expensive part of being able to create like a product Is the time it takes to write the actual code. And you could press that cycle. Now, some of that cost shifts to like testing and test harness and stuff. And I think over time, maybe that will also start to go away as the LMs are actually able to take over a lot of that work. (Time 0:30:10)
  • Connect Models To Local Tools But Preserve Context
    • Use tools that reliably connect models to local files and services while managing context limits.
    • Gregor praised Notion AI and Claude for preserving chat memory and safely combining file, Slack, and tool access. Transcript: Gregor Vand Use them reliably yeah exactly and you and use them reliably and i mean it still depends kind of on your interface to that like i’ve been pretty impressed with actually notion ai recently Because it just has the the tooling and the access to you know files and slack and all this kind of stuff but so models i mean when we say models can use tools you know the service that is built Upon a model effectively can use the tools you then got you know multi-agent orchestration so the fact you can sub-agents working in parallel and like as you were kind of alluding to Like we’ve got these frameworks like gas town now where like you know you have kind of like a lead agent effectively. And I love Gastown’s way of like a mayor and you’ve got, which other people do they have? They have like, well, the other mayor, the town, rigs, crew members. I mean, I think it’s really cool. Just a great way to like add a way for sort of people to kind of get their heads around what’s going on. And then something I’ve noticed just without any thinking of this previously, but, idea of context compaction where AI can actually summarize, basically it’s on memory, and then Can keep working on things without hitting limits. We’re beyond the stage, at least I feel on any paid plan now, we’re on the stage now, where you don’t really, you very seldom hit like, oh, I’ve run the course of my conversation. It’s like, you can always add to a chat that you’ve had previously, but it’s able to do things with that. And then, you know, you start off with say a new chat within, in my case, Claude. And I did get a bit of a like, oh, that’s scary. (Time 0:33:57)
  • Engineering Roles Shift Toward Judgment And Orchestration
    • The role of engineers will shift from manual coding to supervision, judgment, and orchestration as agents handle more implementation.
    • Sean and Gregor argue interpersonal skills, architecture judgment, and product taste become the primary human value. Transcript: Sean Falconer You know, what does that do to an organization? Like, how do we have to think about like, what is the modern organizational structure? We’ve been kind of using roughly the same kind of organizational structure for engineering, product teams, TPMs, and stuff like that for probably 20 years. In particular, like language or framework, and you can experiment more freely. Do we need the same sort of ratio of PM to engineers that we’ve had previously? Do the lines between different technical roles start to blur drastically? Do you have front end and back end? Do you have people who are expertise? I think in many ways, like the deep expertise in particular technologies starts to lose some value when you can rely on the agent to do that. So then it’s like, oh, well, you need sort of more broad expertise to understand architecturally how these things fit together. There’s probably other things that come to play in terms of what the value of a person starts to transform into being, like the taste, how many ideas they can come up with. Like, what is that new choke point for a company if it’s not sort of the laborious task of writing the code? Gregor Vand Yeah, exactly. And yeah, I think you mentioned like, in passing around, are we seeing the end of the 10X engineer, basically? You know, the 10X asshole, effectively. Sean Falconer Yeah, I mean, we’ve all probably had, you know, worked on teams or been in companies where there’s someone who’s kind of like the jerky engineer or jerky expert, but people tolerate It because they have deep expertise in something. Like maybe they’ve been there for a long time and they’re the only person who understands like the full breadth of this code base or they really know the specific type of technology or Something and people kind of put up with that. If the models are really good at that you know does the the 10x asshole essentially does that role have a place in the modern ai forward company versus other types of skills that are maybe Higher value like interpersonal skills driving alignment you know sort of the human component of it and i mean i think looking at the actual tooling that you know has has changed we obviously Gregor Vand Talked quite a lot at the beginning about OpenClaw. But part of the reason it’s really then also jolted everyone. First of all, it’s open source, you know, and it’s not tied to one of the models specifically. And, you know, it’s come out of like, as we call it, effectively one guy. Okay, very talented guy. But, you know, one has has been able to kind of put this thing together and then i think it’s for whatever reason it’s shown a lot of people what they can actually do with agents that for Some reason like claude and open ai weren’t able to fully show or maybe had sort of lulled people into a sense of well no claude and chat gpt are for x and that’s you know, nothing to do with Agents, which is obviously they’re trying to shift the narrative on that. But we’ve only kind of maybe seen that narrative shift very, very recently with Opus and Codex, basically. Is that kind of what you see as well? Sean Falconer I think that’s fair. And I think there’s also been this, now this shift that’s happened where sort of the CLI and the desktop is becoming like the housing using for the agent and then the agent has the ability To kind of reach out to these different systems whether it’s like local systems or could be things on the web and doing you know search or talking over mtp or whatever happens to be and I think that’s an interesting shift i don’t know that that’s something that would have been obvious and will it stay or not? I don’t know, but it seems like people are almost like more comfortable running these things within their local environment. And part of that probably has to be probably related to the fact that so much of this is kind of tied to specifically to agentic engineering. And even though so many things exist on the web, a lot of engineers still prefer to run an IDE locally and have the code available to them locally. And I think Claude and some of these other agentic engineering tools have done a good job of mapping to that existing way of working that engineers are used to, where you can do these things Within your local environment while still giving them the power touch all these different systems. Gregor Vand So I think as you’re kind of touching on, in terms of where is engineering going, I mean, we almost touch on this at some point every month, I think. I feel we always have to because there are just always these leaps each month and something is being questioned around what is it to be a software engineer now. Yeah i mean i do see that the jobs that are sticking around are the ones where you know someone has say like broad knowledge over something and it does require you know human judgment interpersonal Skills you know i think the interpersonal skills is an interesting one because there’s still a lot that goes on in companies where quite frankly it is just people are not agreeing on Something and actually often it requires maybe a third person to come and kind of be the person to get the two experts to agree on how you’re going to get a path forwards and i don’t at least I don’t see almost like flipping a coin i don’t see that two experts are going to go oh we’ll just let open claw decide for us or something like you know because because then they’ll say Oh well how are you prompting or how are your agents set up and how so how how can we trust this so there’s always going to be so long as a human is in the loop somewhere which which they very Much are there’s always still going to be room for humans with engineering knowledge but i think yeah perhaps more on the the softer stuff which is just a higher level judgment or negotiation Effectively. Negotiation being actually a big one as well. Sean Falconer Yeah, I mean, I think that there’s engineering certainly isn’t just about writing code. It’s really about solving problems and operating software systems and potentially operating those things at massive scale. And I don’t think that just having an agent that can code for you necessarily takes away all those problems. I think the big question that people are sort of having right now is, do you need as many engineers doing this if you take sort of the coding piece of that off the plate? Can you have one engineer that’s now able to do what 10 people did before? I think that’s kind of the big sort of uncomfortable question that everybody in the world of technology is kind of wrestling with. And it’s just like, if you see past trends, like we had the industrial revolution that made the labor involved with things like farming, not as intense. So you didn’t need as many farmers. It’s not like farmers went away completely, but you could have essentially less of them doing more work. And that led to an explosion and growth of humans around the world and all kinds of other things. So there was like a ton of good things that came out of that. (Time 0:36:50)