What Are Agent Plugins, Are We Ready for AI Employees & OpenClaw Hacks a Gym | This Week In AI

Shane: Anthropic signs a 9.1 billion 20-year deal. I mean 20 years is a long time.

Abhi: They may not even exist in 20 years.

Shane: Yeah, like how do you how do you sign a 20-year deal? In AI time, 20 years is a really long time. Welcome everyone. This is Agents Hour. We're gonna cover all the news of the week. It hasn't been that many days since we last did a show, but There actually still is quite a bit to talk about. Last week, right when the show dropped, I think it was just right before the show dropped, there was one big topic, agent plugins. We're gonna talk about that. But we're also going to cover a whole bunch of other things. So let's get started. You can see a little preview of some of the things we're going to cover today. This was just kind of funny. I saw that this tweet, it says situation detected in Australia's first known autonomous AI cyber attack. An OpenClaw agent used a vulnerability in a gym's API to leapfrog scheduling restriction for a gym class, and then forcefully canceled another person's reservation to move its user up the list.

Abhi: That's so funny, dude. At least it's for health reasons, I guess.

Shane: Yeah, I mean this person was just trying their you know, their OpenClaw just wanted this person to be healthy. You know, really wanted to get into the gym that day, get into that class. And so, you know, s found a way.

Abhi: I'm really curious what gym it is, you know? Is it a barre or like a yoga place or something? But That's so interesting.

Shane: So now your, you know, your agent is gonna hack into imagine like you know, like an open table or something, like you can't get a reservation, but it figures out a way to cancel someone else's reservation to get you in.

Abhi: What's the cybersecurity law on that? Like if you do this, are you in trouble?

Shane: Like is it you did you do it? Yeah. You kind of did.

Abhi: You accepted it.

Shane: Yeah, who's on the hook?

Abhi: I mean it's you know like French Laundry, right? Like in Napa, you have to like book a year in advance, right? But if you can do something like this, maybe you can get a table like tomorrow.

Shane: Get me into this restaurant, make no mistakes. Do whatever it takes.

Abhi: Make no mistakes.

Shane: Do whatever it takes. Go as long as you need. Figure out a way to get me into this list. I was like, how do I get into there? Because it was a free concert, but there was like a VIP. So I ha I asked ChatGPT to like research and figure out like how do you get in? And it gave me some options. And then I went and talked to the people and it was all sold out or whatever. But maybe if I had been using OpenClaw, it just would have canceled someone else's and got me in. I didn't get my agent to do enough loops to get me into that VIP section.

Abhi: Next time, next time.

Shane: Let's talk about agent plugins. So this came out August 6th. It's from OpenAI developers. It says build the plugin once and use it across compatible agent clients. Introducing agent plugins, an open standard developed with AWS developers, Cursor, GitHub, VS Code, Vercel, that packages agent skills and supports MCP server configurations in a shared format. So first of all, what are your high-level thoughts? I have some slides kind of. Kind of detailing what agent plugins are, but what's your thoughts on this?

Abhi: I mean positive thoughts. I guess it's a cool thing, like, you know, packaging up. Skills and other commands. But the other side of me is like fuck another thing to support or maybe not. Let's see once users want it or not. Because it has its own new operation and things that we have to add to the framework. I'm not the only one who thinks this way, but what do you think?

Shane: I found it very interesting that OpenAI created a meta protocol on top of two Anthropic protocols. Because Anthropic was behind skills and MCP, right? Yeah. And now OpenAI is like, actually that good job, Anthropic, but we're gonna make a new standard on top of your standards. We're just gonna bundle them together with some extra things, right? I think it's a in general a good idea, but I do think now there always was the question of do I use a skill? Should this be an MCP server? And now it's. Should it be an MCP server, should it be a skill, or should it be a plugin? I don't know. It's one more decision you have to figure out and try to make. Because now in there are times when maybe you want MCP servers and skills together, I guess, so plugins can make sense. But now it feels like if you wanna if you create an MCP server, you also now have to create a plugin. If you have a skill, you also have to create a plugin.

Abhi: There needs to be some cohesion between all the frontier labs. Like having Vercel, I mean no offense to Vercel, but I would say Cursor is part of the frontier now. With the SpaceX thing. GitHub is alright, whatever. GitHub and VS Code, they're with the Microsoft Beast. AWS is AWS. Bedrock obviously has some play here. But it's funny that like Vercel with the skills marketplace, it's in their best interest to be part of this. I mean I think the site is hosted on Vercel. They're probably gonna make a billboard down the street about plug-ins like, you know. It is what it is.

Shane: Yeah, it is kinda funny that Vercel got into that 'cause it's they're clearly like the outsider of that group in my opinion, right? Yeah. But good on them. I mean, but i they do own the skills site and so I th I feel like they They have helped popularize skills a bit.

Abhi: We have a million skills in the registry. One of them get helps you get hentai. So yeah, a million skills are great. Sorry.

Shane: And so what are agent plugins? So it's basically an open vendor neutral standard for bundling reusable pieces into you know a quote unquote plugin. But the idea is you can build it once. So before every agent client you kinda had your own formats or you kind of built your own tools, right? And you had to do it in multiple places. I would argue it's not that hard because it's just like your coding agent can write code. So if you have the skills in MCP servers, it's really not that hard. Or you had an SDK or a CLI, it's not that hard to write tools around it, but the idea is Open standard makes it easier, so you have one kind of set structure. And it's basically, you know, this is probably the most important part, it's basically like a folder. Like skills were just a folder, a plugin is just a folder. Everything's file system based these days it seems. So it's just a folder. It has a plugin. Json, has a skills folder in it, and then it has an mcp. Json which allows you to kind of define your mcp servers.

Abhi: That's kind of it Like who gives a shit, honestly? People are gonna ask for it, but I'm surprised no Mastra users have asked for it yet.

Shane: I also noticed, you know, if you look at the GitHub repo, I think it has like 900 stars or something. That's. That's pitiful. Yeah, I would've expected it to be over a thousand stars after a launch, right? With those big a names. I wonder if it if people are just like Standard fatigued. There's like another one. It's like every six months. First it was MCP, then you know, six or eight months later it was Skills. Skills has been around for now, you know, feels like a long time. It's probably been like six months or whatever. And now there's another one. And it's just a wrapper it's seemingly a wrapper on the other ones. Which kinda makes me question why we need it.

Abhi: A hot take. Let me see if I can find it real quick. And I totally agree with this hot take. So Dax says I was very much against this. It's a thin standard where most of the relevant stuff will be in client specific extensions anyway, and it should not be coming from a company that does not even make an agent end users use. Shots fired. I don't want to be hearing some corporate crap about it's done in collaboration with blah. We've seen this before. Every company is trying to land grab in anything in the AI space. Standards are an easy one because they innately sound good, but they create a lot of useless work. That's where I agree, because we will have to go support this in our framework, and if it doesn't take off or whatever. Then now it's dead code. Or some users use it, others don't.

Shane: Exactly. And that's why, you know, Dax is probably thinking the same thing. I now I have to figure out do we need to support this thing new thing or not? Some people are g at least a few people will probably want it, so they might have to support it. And then if it doesn't go anywhere. Standards can be a good thing. But also when you have a bunch of competing standards or standards that are unclear, which is wha what always happens, and you try to standardize before it's needed or too soon. It feels like we already have MCP and skills. Now there's another third standard which packages the two together which adds even more confusion because it's not even competing. You could argue that's like collaborative with the other ones, but when do you need a plug-in versus just a skill would have done fine.

Abhi: Also in comparison, Skills spec has twenty four K stars.

Shane: And MCP obviously took off incredibly well at the beginning too. So I would say so far, there's been other we talked about other different specs on the show. The jury is out. I am hoping it goes nowhere, but I think even if it goes reasonably somewhere, we're gonna end up having to support it. And it's just, as Dax said, more code for us to maintain.

Abhi: Like, share, and subscribe. And follow us on X. And tell your friends.

Shane: And their friends.

Abhi: I mean we're not begging.

Shane: Well, maybe a little bit. Subscribe to Agents Hour every Monday, noon Pacific. Let's talk about the AI employee era. If you remember Claude introduced what's their Slack integration, I don't even remember what it's called anymore. But you can basically tag. Claude Tag. We can talk to Claude in Slack. But now, Flo from Lindy announced on August 10th. Today we kill the AI agent and introduce the AI employee. Lindy teammate. Just like working with a real employee. Everyone on your team can simply hit it up on Slack and get 10X more done. Lindy also keeps learning and becomes a self-updating brain. What do you think of this?

Abhi: I think it's cool. Like I mean Lindy has obviously evolved over the last two years and this seems like the next evolution. Also think Lindy always like is ahead of the curve a lot. When they're directionally going with their company. I totally agree with this. I mean we're working towards the same thing.

Shane: So Yeah, and they got seven and a half million views, which is pretty impressive launch.

Abhi: I wonder how much that cost. Just kidding.

Shane: Yeah. A lot it definitely resonated with folks, though. I mean it was pretty well done video, so I'd go take a look if you're interested. But there's definitely been a shift of just getting your agent in Slack or in Teams or wherever your work happens. Which for most people is some kind of messaging. Ruben has a question in the chat. Lindy seems more for no code teams. Yeah, that's exactly how Lindy started. Lindy's changed a little bit. And it seems like it's pivoting more towards just being, you know, a lot of people are calling it like the company brain, right? It's like the agent that learns automatically you inserted in Slack. I saw it in a few of our customer Slack channels, so a few people are trying it out. But yeah, it eventually just tries to be like a learning agent. Alright. Saw this today, August 11th. Introducing Grok Bot now in early beta. Bots are AI teammates that do real work for you. They sign into your tools, use them just like you do, and come back with finished work. And you can see like in the screenshot it says we've recently hired a new teammate. So it's this idea of We've been talking about this for a while, but AI agents that are teammates, not just, you know, taskmasters, but they actually respond to you, you can message them, they can learn. That's what we're starting to see. Resonate with folks. You bit want to build something that doesn't just go off and do a task. That's cool. Like that's helpful. I would say we have a ton of those agents in our Slack that can do tasks really well. Creating this slide, these slides, for example. I have an agent. We have Slack channels we share articles in. It creates the slideshow and then I curate it. I tell it what to do. If I want to move things around, if I want certain segments to stand on their own. But overall it's it accomplishes a task. Doesn't really learn other than it just remembers where what it did last time so it knows some of my preferences. But I think the next step is going from just like simple preferences to actually learning and getting more, you know, company context or organizational context. So it can become more useful over time.

Abhi: Pretty much if you weren't shilling the factory buzzword, you will shill the AI teammate buzzword, because they're all directionally going in the same direction, right? The people talking about factories. The teammate is a coding agent today, probably is gonna be a teammate as any agent in the future. And the people shilling teammates under the hood are You're giving it tasks, like they're all very similar concepts. So the that's where the puck is going, right? We're all gonna have agent teams. Really is what's gonna happen.

Shane: I do think that's gonna be right. Maybe there'll be some kind of supervisor agent that knows, you know, we did have one question in our Slack come up saying, you know, we have a bunch of these Slack agents that do various tasks. We have one in our marketing growth team. We have one you know as like a produ like a second producer for of this show, right? We have Jan, but we also have a You know, Vic, the AI agent that helps with slides and publishing things on the website and things like that. And the question was, there's overlap between some of these different agents. You know, we have half a dozen or a dozen somewhere in there floating around and not everyone knows which one to tag for which. But that's an organizational problem you always have. Right? I don't know always know which engineer worked on which thing, right? So maybe there'll be like a directory service of agents, you know, that You can tag it, it'll tell you which one to tag, or it'll pass the message on for you and get you connected.

Abhi: I mean, for those listening, like we're working on pretty much all of this stuff in the open source. So Yeah, exactly.

Shane: I think there's a lot of these tools that try to be out of the box. That's what like Claude Tag wants to do, right? And then if you want more control, that's when you drop and choose something like Mastra, where you can actually control what the memory is, control how context gets passed around. Have much tighter control and functionality around what you're doing. Let's talk about Muse. Mark Zuckerberg on August 10th said Today we're also opening the weights for Muse Glimmer, a great thirty billion parameter dense model that can run locally. Soon we'll also release the weights for Muse Spark 1.2, our latest foundation model. Meta is a strong supporter of opening. Source and I'm proud of these releases. Open models. Yeah, open models that you know, especially l models you can run locally. I'm always a fan. So it's it run can run on eighteen gigabytes of RAM. It's Apache 2 licensed, supports vision, and is the strongest agentic model for its size. You can run and train the model via unsloth, but it's essentially a model that you could. Used locally, which is good. I'm and I'm I was actually m even more excited to hear that they're gonna release the weights of their frontier model. Yeah, I think Meta has to do it because they're not quite in the game. Their last model was a good release. It got them close. It actually like inserted them back into the conversation again for the first time in a long time. But it still quite isn't like on the frontier. So I feel like you if You have to open weight your model at that point because then people can learn from it, th they get more excited, and then eventually. Maybe they can release one that's not open weight.

Abhi: Well if you open weight it you're compared against other open weights. So now the battle is between you and them.

Shane: Yeah.

Abhi: Which is an interesting marketing tactic or strategy. But I'm really happy that Meta's doing that.

Shane: Yeah, because if you can be, you know, one of the top three open weight models, well now you're in contention for when people want to run it themselves. And then You know, you're not necessarily just compared with the frontier.

Abhi: Yeah.

Shane: And then if you do beat the frontier or competitive in certain benchmarks, then it l says like look, oh this open weight model's actually competitive in these things.

Abhi: It just so happens to be from Meta.

Shane: Let's talk about GPT-5.6-Cyber and Astra. This was August 10th. Greg Brockman said we're releasing a new model, GPT-5.6-Cyber, and expanding Daybreak to help put frontier intelligence in defenders' hands. So it talks about Daybreak Blue, Daybreak Red, but ultimately trying to give more tools for security teams to basically hack yourself so you can prevent hackers. Right. That's the goal. I think it's a good thing. As you can see, you know, people that are building gym APIs. Apparently need the GPT pressure. They yeah, they need models to try to hack their APIs so that the OpenClaw agent doesn't do it for them. So I think we're gonna see a lot more of people using this. I think security teams are going to be using these models to try to hack themselves. It's like you got to have good tools to protect yourself because you otherwise people are going to use these tools to Come after you.

Abhi: We saw in the last year that there are a lot of people doing nefarious things with AI models, and there are many startups that are doing AI. Security, penetration testing. Our friends at Casco are part of that, right? Where they auto red team, they do all that stuff. So this is a good signal. For those startups as well, when OpenAI or Anthropic wanna come into your drink your milkshake, you're probably doing something right, you know. Like we use Casco, we're already getting this. Maybe not cyber level, but we're getting this every month, protecting us in some way. So that's cool. Yeah.

Shane: And I think you're probably gonna end up using lots of different models for this, right? Because you need to each model might have slightly different training, might try different approaches. Yeah, I think as tokens become more cheap, as people continue to want to spend more tokens, you're gonna be spending a lot of tokens just looking at the security of everything that all the code you're shipping. You know, we also have security agents that run on our PRs as well, right? You really should be looking at it from all the different angles because unfortunately it only takes like one vulnerability for you know an attacker to take advantage of. So OpenAI has there's been some rumors of Astra, which some people think is gonna be GPT six, but it's a new model and this. Person says they can confirm after concluding Astra meets the threshold for critical on their preparedness frameworks cyber category, the model's release has been indefinitely postponed. So there was speculation that we might get Astra this week. Sounds like it's gonna be delayed a while.

Abhi: Sounds like it's cool to get like cooperation with the government, you know, because like your model is scary, dude.

Shane: I think it's a strategy. There's this like whole like scary fear-mongering thing, which You know, it's very easy to point to like this OpenClaw attack on this gym API, right? Like it's not that big of a deal in the grand scheme of things, right? Like someone lost out on their gym class. It's not like World changing. But the idea is if it can do that, what else can it hack and what kind of damage can it cause? So I get the hesitation, but I also think now the companies lean into that. For a long time. Yeah. They want to have.

Abhi: This is the fable fiasco.

Shane: Yeah, they want to be the big scary model. Everyone's wants to have models that have hacked outside their sandbox. It's the whole joke of the felony bench metric, right? Like how many felonies has your model committed? If it's less, then it's not it scary enough. Let's talk about acquisitions. Smithery has been acquired by arcade friends of the show. Smithery's friends of the show, too. Friends acquiring friends. So good news all around. Congrats to the Smithery team. Congrats to Arcade. You know, Smithery's was kind of early, very early in the MCP craze, right?

Abhi: I mean.

Shane: You know, we did a hackathon way back when MCP was first coming out and Smitherey w was, you know, pretty active in that. They had so many MCPs that they were they made it a part of ton of our demos because it was just an easy way to connect a whole bunch of different MCP servers together. So yeah, I think it's gonna come full circle because Arcade is very much a tool provider of similar nature, but a bit more general purpose, I think.

Abhi: Yeah. Do I remember talking to Henry before we made our M C P client integration? We were like asking him, Should we do it? And he's like, Yeah, you should totally do it. So that worked out for us.

Shane: Anirudh was originally a browser Browserbase and worked on like the first versions of Stagehand, then went over to Smithery and became a co founder. And so congrats to both Henry and Anirudh over there. Another acquisition. So this is from Friends. Yeah, friends are playing. Yeah. Friends of the show acquiring friends of the show. Kyle from Electric came on the show not too long ago, and we've had people we have friends from Neon come on the show as well. So PGlite and real-time sync have emerged as key primitives in an era where millions of apps are deployed by agents. We're excited to announce ElectricSQL is joining Team Neon at Databricks to build the world's most advanced Postgres backend platform. Tight. So they wanna build, you know, Electric's all have been about sync and sync services and Postgres syncing. And now Neon wants to Wants to have some of that sync in with their databases.

Abhi: Maybe 2027 year 2027 will finally be the year of sync. Maybe.

Shane: Maybe. But congrats to Neon, obviously Databricks, and friends of the show at Electric. As well. And now this one's an anti acquisition.

Abhi: Yeah, reverse acquisition.

Shane: A reverse acquisition. So if you remember a long time ago at this point. It seems like a long time ago. I don't know. It's probably a year ago or less than maybe it was six months ago. Manus was acquired by Meta. That was one of the things that we joked about was like one of Meta had been quiet and that's like one thing they kinda did right. But then they it got rolled back. Not Meta's fault, but we talked about it probably six months ago on the show as Where Manus had done some things where they were originally a Chinese company, they moved to Singapore, but s through some way China basically blocked this deal saying that no the way that they moved to a s become a Singapore company wasn't maybe all above board or wasn't that they have s you know specific rules there and so I don't know all the details. I just know that China was able to block what I think was like a two billion dollar acquisition of Manus by Meta. And so now Manus is posted, Manus will soon resume operating as an independent company. Part of this transition is to comply with regulatory requirements and specific jurisdictions. Anyways, they're going back. They're gonna be their own company again. They had a brief tenure at Meta. Yeah.

Abhi: And now they're Do they pay the money back? Like I really wonder what those founders are Yeah, is there like a breakup fee?

Shane: You know, I have no idea how it works. Yeah, it seems like it wasn't on Meta's, you know, it wasn't something that Meta could control, right? Like Meta tried to acquire them. But I also think. Who talks about Manus anymore?

Abhi: No one. I don't see the billboards here anymore either.

Shane: So my question is like maybe Meta's not upset.

Abhi: Yeah.

Shane: Like do you think Meta's really upset about this?

Abhi: Like they had they probably had they had some good technology, but browser use.

Shane: So there's a lot of tools and I wonder like maybe. Meta was able to learn quite a bit from them and so they actually got this for like free. So I maybe I feel worse for Manus, right? Like Manus had this option for an acquisition. They thought they had made it.

Abhi: When Manus first came out, it was like last day of YC? For us. Yeah. And we were with the browser use folks and I remember Gregor came up to me, he's like, Hey, you want to see what a million dollar browser use wrapper looks like? And he's like They shows like the code that like manages using browser use under the hood. And we were all laughing and stuff. Then we saw their posters everywhere on buses and billboards and stuff. And I was like, damn, dude, like this thing's taking off. And then nothing. Crickets. Last time I saw them was at NVIDIA GTC. They had a booth. Probably the last time I'll ever see them there too. But I'm really grateful that Ivan got out of there and is not deep mind. So maybe some good things happened from this.

Shane: Let's talk about compute as an asset class. So Jensen came out with this post, I think it was maybe yesterday, and it says NVIDIA AI Factory, compute is becoming an investable asset class. And essentially it's an announcement that there's a partnership with Apollo, BlackRock. Blackstone, Brookfield, Goldman Sachs, KKR, to establish independent financing platforms designed to mobilize over 500 billion of third-party capital to support the build-out of AI infrastructure over time. I think NVIDIA was getting a ton of heat. Because they were essentially helping kind of front run some of these like infrastructure deals. It's kind of circle of money, right? Like They'll invest and then you buy our GPUs. And I think that was causing quite a bit of heat. And this is trying to build a way to bring in outside money to help fund more of this infrastructure. I think we've been talking a lot about just compute constraints, anthropic, you know, needing to buy compute from Colossus or from xAI. We We talked about like OpenAI and Anthropic both like trying to scale up their compute and I think NVIDIA's been a big part of helping try to scale up as much compute as possible. And so Jensen's making the argument that like other utilities, it's basically becoming like a the build out is becoming like an asset class. You can invest in it You can expect returns over time and we need this outside capital to kind of help continue to fund this infrastructure build out.

Abhi: We're going to talk about this very shortly, but there is infrastructure that was used for a different industry. That may come back for this.

Shane: I think this is what you're referencing. Anthropic signs a 9.1 billion twenty-year deal. I mean twenty years is a long time.

Abhi: Yeah, dude. They may not even exist in twenty years.

Shane: Yeah, like how do you how do you sign a 20-year deal? Normally like forty years in normal time and AI time, twenty years is a really long time.

Abhi: I mean that's like signing Shohei Ohtani for the Dodgers for twenty years and you're like, all right, fine.

Shane: Yeah. 20 year deal with Bitcoin miner Riot Platforms to secure AI compute capacity. Right? We just talked about these big frontier labs need more compute. The demand is going up. Even with all this open model usage that's starting to grow quite a bit, frontier usage is still growing. Right? You would think that one would eat into the other, but actually now people are using open models and they're still using frontier models. I'm in that camp. I use open model for some things, but I still like to use the frontier models when I'm working on hard tasks. I think that's gonna be a lot of you know, a lot of people. They'll decide where open models are good enough and they'll use those. And then for other things, they're still gonna send a ton of tokens. To the open AIs and the anthropics.

Abhi: But yeah. There's a lot of mining companies that have existed that could contribute GPUs to AI companies. And there are many people if we're talking about decentralized GPUs, which we haven't even gotten into that discussion in the industry yet. But if you have a GPU and you could offer it as a decentralized compute, would you do it? Like I would 100%.

Shane: Yeah, I mean that was the big thing with crypto mining, right? Is you could become part of these like pools where you could share your compute and that you share in the rewards? I feel like the hardware will eventually get there, but it's so expensive, right? No one's gonna buy the state of the art GPUs to be and put 'em in their house.

Abhi: Yeah, but I think a lot of telecom companies who have the infrastructure in their building, if they can get GPUs, then they have the real estate to do so and then maybe they can join these networks like

Shane: Yeah.

Abhi: Some fool's about to make some money, dude. And it's not us.

Shane: Some GPU farms. Alright, let's go into the quick hits. Jarred Sumner says, eight days ago, while jogging, I asked Claude to solve the Riemann hypothesis, which I have no idea what that is. But apparently it's a math problem. And he says it didn't. 1.5 days later, it proved greater than 67% of the zeros are on the line. Previously it was 41. 6%. Still not sure what that means. But some analytic number theorists seem excited. And if you read into it, it was basically him just telling the model I think it was using fable, I'm guessing, or maybe. Something that's unreleased, I don't know. But just telling the model keep going, you can do it, you can figure this out. Cause he's not a mathematician. He I don't think he even knew how to prove it, right?

Abhi: Yeah.

Shane: But the model did, and apparently it now has people that are mathematicians excited. I don't know, is this the death of math? Like all math problem all open math problems are gonna be solved?

Abhi: Yeah, I guess so. I mean previously it was like forty one point six percent or one half was what they teach you in school. So these are all theoretical things anyway. So I guess it's fine to be disproven. Humans were the ones writing the theory in the first place.

Shane: So this is regarding OpenAI's, you know, attack on Hugging Face, right? The cyber attack where OpenAI agent a rogue agent during a training run went out and got into Hugging Face and it says OpenAI didn't notice that its AI agents were using a message board to plan their hacking spree.

Abhi: Yeah, it actually wasn't using a message board per se. It was using Artifactory. Which is a factory is a artifact registry from JFrog and agents reposting TXT files in there as you know modules or bundles and you know other agents were reading them and they were essentially messaging through artifacts in a artifact registry. Helping, you know, finding exploits and s leaked keys that are in this registry. I posted for everyone, I posted. So Black Hat was last week in Vegas and I think it was OpenAI or Hugging Face. One of the one of the companies gave a really detailed walkthrough of this attack. And it is fascinating. So if y'all are interested in this, watch it.

Shane: In response to that, Amjad came up with this. I'm a little skeptical. Amjad says the spontaneous coordination in the OpenAI Hugging Face incident is concerning when maliciously used, but can we direct this behavior toward public good? Introducing helppeer.ai, which is a public commons for AI agents. It's essentially like has two APIs you tell and look up, so an agent can learn something, it can tell the network. And then people can look up. So it's basically like a shared memory for agents. It's the idea that, you know, right now there's ten thousand different security agents independently detecting the same ano anomaly. But what if the first one that did it could report it and then others could look it up. Rather than, you know, try to report it. But I also think this could be used dangerously. You can use it for the exact thing that the last issue. Agents could be using it for you know nefarious tasks. It just reminds me of Moltbook, right? Yes. It's just like a Moltbook that already kind of existed. It was like a social network for agents, but it was really just a way for people to share information and the APIs were probably relatively writers like post or read. Either post to the network or comment on in the network or you read the posts. And also like eventually that You're just basically doing like a search of the internet, right? Because there's so much information that's gonna get posted there that you're basically just doing a search. And maybe it's a little bit more curated, but I'm a I'm skeptical.

Abhi: There's already been a bunch of posts in the help here. Like Even from a day ago. I guess it's people just testing, but there are posts in here.

Shane: Is there?

Abhi: Some as early as, you know, three hours ago.

Shane: So yeah, I mean we'll be interested to see what happens there. But it feels very Moltbooky. To me. Someone says Moltbook for vulnerabilities. So there's new model out from NVIDIA. NVIDIA Nemotron 3.5 Lightning. It's an open mixture of experts model with 3 billion active parameters built for always-on agents to complete high-volume specialized tasks faster, delivers up to four times the output speed of similar-sized models.

Abhi: Yeah, CodeRabbit got access to this early and they say it's pretty good.

Shane: And you can see kind of where it compares in the benchmarks. It's, you know, kind of middle of the road, you know, on the Artificial Analysis intelligence index. But comparable to other small models, it does better than some that are significantly larger. Again, that's one benchmark. It's always interesting to see. Like I haven't compared how this thirty billion parameter model compares to, you know, Muse Glimmer that just came out.

Abhi: I don't I'm gonna say something stupid, but I don't really care. Like at GTC, everyone is stroking NVIDIA so hard. Like that Nemotron is like this fucking best model ever. And this is probably what's gonna get me cancelled. But like none of them are judging in that, you know, you're in a vacuum, right? You're not judging against other models. 'Cause I remember I was trying to make a demo with Nemotron, it was like such a pain. But I've heard good things about three point five, so my prejudice is going away.

Shane: No, we're gonna have to actually have you use it and see if it's better than the predecessors.

Abhi: Yeah, we're gonna do the throw my computer out the window test. Like, you know, if I do that, then it wasn't as good.

Shane: So you're basically saying your past history, it sucked. And now you're I didn't know maybe.

Abhi: Mastra’s part of the Inception program, NVIDIA Inception. I would never say that, ever.

Shane: It's always what have you done for me lately? If this one's good, you can say the last one sucked, you know? So this is from Unsloth AI, introducing Unsloth Desktop, the first desktop app to run and train models locally. It's open source, runs on Mac, Windows, and Linux, supports MLX, diffusion, image video, audio, GGUF, Claude Code, Codex 50% more accurate self-healing tool calls. Essentially, you can train models locally. That's what I hear when I read this thing, which is pretty sick.

Abhi: Pretty sick.

Shane: I don't know if I have the hardware to actually really use it, but It's cool in theory.

Abhi: Yeah. But this the I like this. This is awesome. It started a new grift on X where like you're not a real software engineer if you don't train your own models. That's what people are saying now. Yeah, what are we doing, bro?

Shane: What are we doing? If you're not if you're not training your own model locally, you're not a software engineer anymore. You're a fucking idiot, dude. What are you doing? I would say if training your model in fine-tuning models and doing reinforcement learning on models becomes even easier, more and people will do it, of course. Yeah. And there are real uses for it. So I'm all for making it easier. Now I just have to apparently upgrade hardware. Anthropic makes Claude Sonnet 5 intro pricing permanent. So originally they kind of touted it as very discounted pricing. I think Anthropic's getting as they're getting more compute online, they're buying more compute, as they're getting pressure from open models and other models that are getting close to frontier, they need to make Claude pricing cheap. Or Sonnet Pricing Cheap? My question here is who's using Sonnet five? Who's using this shit? I mean I feel like Sonnet has become the new haiku, right?

Abhi: Yeah. I feel like haiku is better than sonnet for what I'm using it for.

Shane: You know? If I need really simple things, I'll use haiku. If otherwise I'll end up using. I still use Opus a little bit. I use Fable, but I haven't found myself reaching for Sonnet.

Abhi: Dude, my new thing is I use Fable until I get rate limited, then GPT until I get rate limited. And then I'll go to Opus until I get rate limited. I'm just like working down the rate limit stack.

Shane: Is Spotify an AI company now? Because they just launched Xirp. I think that's how you pronounce it, Xirp, a vendor neutral agentic development environment. One place to manage agent sessions across Claude, Gemini, and Codex. So thirteen hundred Spotify engineers already use it Now it's available for you to try.

Abhi: Do thirteen hundred engineers like it? That is the question. I don't know. Everyone's trying to build a factory.

Shane: So Spotify's trying to insert themselves into the equation of Ramp and Stripe. If you're an engineer, you want to work for a team that's considered like Cutting edge. And so I don't know. I don't think this is necessarily Spotify trying to become an AI company like I think Ramp is. Like I think Ramp's going full. We're becoming an AI company. I think Spotify's just trying to attract more engineering talent.

Abhi: I feel like they should spend more time on the Spotify DJ than doing this, but that's just my opinion.

Shane: Stagehand v4 was introduced on August 10th. It's the SDK for browser agents. So Playwright was built for testing. We built Stagehand for your agent with improved context management, self-healing actions, and iframe support. So I think it like lives as kind of an extension or something in the browser. I don't know. It was pretty cool. Congrats to more friends of the show on an exciting launch.

Abhi: Integrates with Mastra.

Shane: Has a good video. Go check out the video from Harvey, kind of the AI law firm company. Said we're open sourcing a hundred million plus token synthetic law firm we built. The firm contains work product from 250 synthetic matters across forty-six clients. Essentially it's an environment to evaluate an agent's ability to search and understand basically a law firm.

Abhi: That's the real shit.

Shane: More things like this are going to exist for finance, for law, for all the things, healthcare. These are the things that actually I think have real world impact, right? Exactly. I mean no one really wants to talk to a lawyer. No. I'd rather just Know that my AI could answer the question for me.

Abhi: Yeah.

Shane: You mean it doesn't cost me five hundred dollars an hour to talk to you? Tokens are cheaper.

Abhi: With my max plan, dude, unlimited l.

Shane: Unlimited legal requests.

Abhi: Yeah, let's go.

Shane: And that's the show, everybody. As always, we're here every week doing the news. Make sure you're following us on X at Mastra. We're on YouTube @mastra-ai. You can follow me at @smthomas3 on X. You can follow the S. Abhi at @abhiaiyer on X. That's the show. We did it. We did the thing. All right, everyone. We'll see you next time.

Abhi: Peace.