Le Chonk! Also: Espionage at OpenAI

Shane: Anthropic is updating its usage policy. The new rules prohibit sustained abusive or cruel behavior towards Claude. I think you're in trouble.

Abhi: I'm in trouble.

Shane: And it is funny because if you've ever dealt with coding agents, sometimes I feel like getting frustrated because it's not doing what you want gets you better results. Welcome to Agents Hour. It is not Monday. It's Wednesday, October 14th. There's a holiday on Monday. We had some traveling. But we're here. We're doing the news just like we do every week. I'm Shane. I'm one of the co-founders of Mastra. I'm here with Abhi. What's up? What are we doing here this week, dude?

Abhi: Well, we're together in SF. We're having a little Mastra offsite with some of our Mastra employees. But we're live here at the CodeRabbit Studio, one of our favorite places to record. And if y'all don't know anything about CodeRabbit, CodeRabbit is friend of the show, friend of us, but the code review product, and actually more than code review now, so if y'all are looking for some code review, you can get that, you can also have it explain PRs with CodeRabbit Explain, they have a new product called CodeRabbit Triage to help you kind of organize your software factory. And finally, they have this awesome product that I have not used yet, which is CodeRabbit Security. And we just had some security vulnerabilities. I wonder if CodeRabbit could have helped us there as well.

Shane: But anyways, we appreciate you tuning in. Why don't we just go ahead and jump right into it? There's a lot of news today. OpenAI found a bug, but not your normal kind of bug. This is a literal bug. So this came out actually yesterday. It's from OpenAI Newsroom. We discovered a sophisticated recording device that was hidden inside one of our offices. We used our models to understand the device and where it had come from, identifying that a specific employee had planted it a few days earlier. We referred the matter to the FBI. And we got a little more details on this. It sounds like it was a bug planted in maybe an executive's meeting room that was kind of more for confidential conversations. It was planted behind a TV, maybe wrapped in toilet paper or something weird, you know, so something like that. So undetermined exactly, you know, the intentions there, but apparently one of the people, or the person responsible, was also working on the, I think the OpenAI Agents SDK team. So I'm assuming no longer at the company, obviously.

Abhi: Yeah, well, FBI. Good for us though.

Shane: Yeah, I mean I guess. What a wild time. Like, that you think that you know, espionage is the answer, I guess.

Abhi: I wonder what they like listened in on, you know, AGI is here. Maybe they were talking about the Dots launch and things like that.

Shane: Yeah, I was only there for a couple days, but you know, I guess protect your secret conference rooms.

Abhi: Well, they were also saying that they used AI to, you know, sleuth it, like, you know, Sherlock Holmes, to figure out who, and I wonder if that's all real or not.

Shane: ...you probably just know that something, yeah, something's wrong. Yeah. That's not supposed to be there. But anyways, so that's the world we live in. Let's talk about all the model releases. There's a ton that have come out throughout the last nine days or so since we last did the show. First, Le Chonk. Le Chonk. This was kind of rumored like a while back.

Abhi: Yeah, it was the Opus killer in France, which are like two things that don't go in the same sentence. Models in Europe. But it actually is quite a good open source model for the most part. Interesting thing about Mistral is their use case is predominantly in Europe. Predominantly for military and government use cases in Europe and you know, France is the capital of AI in Europe. So very much people in France do use Mistral. People don't, over here, don't necessarily use or care about Mistral.

Shane: Yeah, I mean I think when they were one of the only open models in town, a lot of people experimented with it. But it is good to see that, you know, maybe there are, you know, some places in Europe that still have at least a pulse, which is a trillion parameter model, which is, you know, it's a mixture of experts model, 49 billion active params. It scored relatively high for an open model on Artificial Analysis and some other benchmarks. So it's at least, you know, not in the game, but it's game adjacent. It's game adjacent, yeah.

Abhi: I was in France when the initial rumors for this happened, where they essentially exploded X with a bunch of fat cats. Yeah. You know, just all over. And maybe Mistral took that and ran with that as their branding here. Or it was planned all along and they leaked it or something. But in any case, like the hype that was given and where it benchmarked are quite far in between.

Shane: Good job for them. I think at least they're... I think people kind of expected they were done with big models. I think that was generally the consensus that a lot of people had thought they were just gonna focus on smaller, more specialized models, but you know, they're in it. On October 7th, Claude introduced Claude Haiku 5.5. It's the cheapest, fastest, and most capable small model they've released, and it costs 75% less to run than 4.5. I think a lot of people also thought that, you know, we weren't gonna get more Haiku models. Yeah. But it came out and seems to benchmark better, a lot better than the other one, and it's cheaper.

Abhi: Yep. Which is interesting. I've been using Haiku 5.5 in observational memory. Does a great job. I'm just curious why is it so much cheaper? 'Cause it is a lot better too. Maybe we'll see that as a trend in the Anthropic models.

Shane: I mean I think overall with Opus 5.5 and Haiku 5.5, I feel like I don't ever use Sonnet.

Abhi: Yeah.

Shane: I don't have a reason. I would just either use Opus. I know it's more expensive, but assuming I'm still in my subscription, I'm gonna use Opus or with observational memory in Mastra Code, Haiku. And we thought Anthropic was dead for a bit and it has, you know, it has re-risen.

Abhi: Yeah, like our team is switching back to Anthropic right now.

Shane: ...or dang close. He doesn't want to like Opus 5.5, but he's like, it's just too good of a model. It's too good. I gotta use it. So I think it is wild though how for a while everyone was on Codex. Everything is, you know, 5.6 Sol, Astra, you know, there was this moment, and now it feels like Anthropic's back on top.

Abhi: Yeah, it just keeps going back and forth.

Shane: Yeah, I guess that's the world we live in. But Zuck wanted to say something about that. So introducing Muse Wonder, our new family of frontier models, with state-of-the-art performance for personal and professional agentic work. And this also benchmarks very well. You know, in the Personal Superintelligence Index, which is I think where they're really tuning it for, you know, Muse, right? They want it to be for personal agents. But if you look down at, you know, some of the Terminal-Bench, it's pretty dang close. Yeah, ProgramBench at the same time. DeepSWE, it's ahead of everything, you know. I have heard a lot of people, some that say it's very good, but also some that say it clearly was benchmaxed to the fullest extent. So we'll see where it actually lands in daily use. First impressions have been good from what I've heard though from people that have tried it.

Abhi: Yeah, pretty much, there's a community forming around Muse. And I think it starts with having Muse as a personal agent and then maybe it will cross the chasm into other use cases. Or maybe Meta just wants to win in the personal agent category and they can just easily do that on us.

Shane: If they can be frontier adjacent and own the consumer space, they can win. And I mean we were talking to someone today who, you know, she said that she's used Grok Bot, used Muse, used Instinct, and she just loves Muse.

Abhi: So the market is here.

Shane: Yeah, it's resonating.

Abhi: People are switching different personal agents all the time too. I guess not one has the perfect user experience for all things. This person we're talking to uses Grok for some business type stuff, Muse for personal type stuff. Maybe he doesn't even use the dot, I don't know. But there's just a lot of options.

Shane: Yeah, Instinct, obviously. A lot of people were very hyped on for a while. But yeah, models trained towards personal agents maybe is the next frontier. So we rumored this last week. Well, it was announced for people with early access, but it's finally available. Gemini 4 Argon is here. It was available as of yesterday, which is October 13th. If you looked at the benchmarks, they were pretty good. They were, you know close back towards the frontier. I think Muse maybe stole a little of their thunder. Yeah. Because I do think if you look, Muse seems to benchmark a little bit better. So Gemini maybe should have launched it instead of just going early access, but maybe they held it back for other reasons. But it's here, another model for us to play around with. I haven't heard as much though of actually people trying it yet.

Abhi: I think they just got their thunder stripped, you know. But all your tokens, what was the joke? All your tokens Argon.

Shane: Yeah, yeah, because it has a million token output window, which is pretty wild. Yeah, all your tokens Argon.

Abhi: Maybe we'll see a Gemini personal agent.

Shane: You'd have to think they're pretty well positioned, right? Alright, OpenAI apparently likes to do math. This was from October 6th. They had a whole math post, they have a math website as well. You can see there's a list of all like some of the math problems that they apparently had, 81% of these different math problems were released in one day. Now, to be clear, like some of it has been withdrawn. There's, you know, six new ones, three withdrawals. This was, you know, from October 8th, but they clearly built a model that can do math. So I think, you know, we know that models can be pretty good at coding because some tasks are verifiable. Math is also something that you can kind of verify. So the idea is models should be pretty good at that, and turns out they are. They basically just brute force the thing, though.

Abhi: And I mean I'm sure they're gonna just try to hit every Millennium Prize. Yeah. And just loop it for as many agents, as much compute,

Shane: As many millions of dollars to win the prize.

Abhi: ...like a plus one for the challenge. Like you don't need to solve it, but you can like disprove things or prove things. So.

Shane: Yeah, I am far from being heavily into mathematics, but I I do imagine that a lot of mathematicians don't like this.

Abhi: No.

Shane: It cannot feel good knowing that things that you probably spent thousands of cycles on in your brain are now being solved in the course of, you know, days or hours in some cases. Yeah. By, you know, computers that can obviously spin way more threads than a human actually can.

Abhi: Well, it makes me think, dude, like are all problems meant to be solved? Just because you could solve it with AI doesn't mean you should, or should you? You know, if it's an old-ass problem, it's like forty years old. Does it really matter?

Shane: Yeah, I mean I've seen people go both ways on this. You know, me not being an expert, I don't know. But some saying that some of these problems could actually have real-world impacts on space travel or you know, flight or you know physics type things, right? So there actually could be real world applications, but then I've also seen a lot of people saying like, nah, I don't believe it. It's mostly we just want to say our models are really smart, but it's probably not going to have immediate long-term impact. We'll see where it lands. I don't know. I don't have a horse in the race.

Abhi: Dude, did you subscribe?

Shane: Dude, I host the show. Did you subscribe?

Abhi: Did you subscribe?

Shane: Subscribe to Agents Hour every Monday, noon Pacific. All right, let's talk about decision models. Round two, probably round three actually. We've been talking about decision models every single week. So we have Diogo from Jev who, you know, basically says in this post, in this video, you know, they launched a model, they served a trillion tokens a day, almost 30% of the Fortune 500 showed up. And then they raised a very big round. But I think people were saying that it was huge, it was a huge round, but it wasn't as big as what some people thought it could have been.

Abhi: I think given like the Instinct valuations and all that they maybe, people are expecting a more hype-driven valuation and round, but they're gonna do a lot with that Series A. They're gonna hire some great people.

Shane: I think, so they raised $870 million at a $7.5 billion valuation. So I mean that's big. But they did have incredible hype. Incredible usage. You know, it wasn't just hype, it was people actually using it. Tons of usage. If you can do something and then within a month or six weeks or whatever it is, Definitely the most successful model.

Abhi: Release ever.

Shane: Yeah. I mean you get 30% of the Fortune 500 that has, you know, pulled out a credit card or paid you in some way. Yeah. Yeah. Pretty wild. Yeah, good for them. They have a lot of competition though. And so they needed the money because they're gonna have to compete. This model came out, SubQ Snap. It's the first long-context decision model for real enterprise work. It's touted as the only decision model where compute per token stays constant. And so of course it compares itself to Jev, because everyone has to. This got, I think, almost four and a half million views on X. So just again, decision models are the hype. People are trying to figure out how do they play in and how are they used.

Abhi: Yeah, they're trying to make it multimodal, bigger context. Anything that gives them an edge over Jev. Yeah.

Shane: Microsoft wanted to be in the game, so on October 9th, Satya said, introducing Microsoft-Decision-1, our new model for fast decision-making. It delivers top performance on structured decision tasks, outperforming both LLMs and other decision models, you know, aka Jev, in latency and quality. And they're testing it across Microsoft for everything. You know, they list a whole bunch of things they're testing.

Abhi: I don't know how long it would take me to find this fucking model. Good luck. Good luck though. Good luck to Microsoft.

Shane: And then of course, you know, Unsloth comes in, you can train your own decision model with a free notebook. So people are just taking Qwen, building their own decision models on top of it. It's the thing. People are doing it. And then Cloudflare is expanding the Clef decision model family with Clef-omni. So you can do multimodal. So again, if you want decision models, guess what? You got them. You got a lot of choices. You were born in the right era for decision models. Let's talk about Anthropic. This was October 13th. Anthropic, the official X account says, new on our Engineering blog: creating in-house source control. Basically, they built their own Git, right? For internal purposes. So Cursor did it with Origin. You know, but that was obviously external. Anthropic's doing it in-house.

Abhi: Yeah, it's really cool. So like, their velocity has dramatically changed with in-house source control. They're doing like, you know, 30K or 50K code changes a day. Which, I mean, in comparison, to think that you could get that many in a day, given the perils of CI and review and SOC 2 and all that. They must have built a system that streamlines a lot of stuff.

Shane: Well, yeah, I mean, in the blog post, gotta be honest, I didn't fully understand every piece of it. But essentially a lot of the checks happen after the merge. They will merge and then agents check after it was merged, to try to prevent as many conflicts as possible. I think there's some obvious, really interesting approaches. Someone's gonna open source, take their blog post, tell Opus 5.5, build me this product in open source, and they'll open source it, I'm sure. Probably won't be any good right away. But I do think one, Anthropic probably is gonna hype this and they're probably gonna release it at some point and say, oh, you're already using our models, why not use our source control?

Abhi: Yeah.

Shane: We want all your code so we can train on it.

Abhi: Yeah. I think what some people don't realize, you know, I've been thinking about this a lot too, because we're getting just destroyed by CI. We talked about it in our last show. Like there's all these affordances for you not to merge your code. Right, like long CI, you have to get three reviews, CodeRabbit said you suck, like there's all these things that are there. And it's because you're operating in GitHub and you're bound by GitHub. You don't have to be bound by GitHub. You can still be bound by Git, much like these guys at Anthropic. They had all their code changes be represented in just file storage, cloud file storage, that you can then merge code into that, run all your automations before even being distributed. And in that case, like maybe you can stack reviews, stack merges, and only have one approval or one set of CI or whatever. And that just speeds up your process. Because we're waiting like 10 to 15 minutes per PR. What if we all just merged into one, tested it all at once? Would we save, you know, collectively hours on end, probably.

Shane: Yeah, or you just let people merge and then at points in time it runs it against everything, corrects it. Something bad got merged, it just corrects it, and then all this other stuff comes in, it gets merged. It's an interesting way to do it, right? And maybe you don't need to check every single PR for everything. I don't know. We will see what the future brings there.

Abhi: Anthropic is using like a Git-adjacent... Yeah, they built their own like protocol. It's like Git-adjacent, it's called Scion, like the car, which doesn't exist anymore. But you have the Scion on Git protocol thing, and then everything is based on a changeset, much like what we do with changesets, but each changeset gets evaluated. So I wonder how many PRs you can stack in there and then just do all your automations at once. I'm gonna explore this more.

Shane: It reminds me, at one point you built a TypeScript version of Git. So it's like maybe we need it, we didn't take it far enough.

Abhi: Yeah, that was the Opus 4.5 project.

Shane: Yeah, that was back at Christmas time when tokens were really cheap. Yeah. Let's continue talking about Anthropic. So this is from Claude. It says one more thing. We're halving the price of cache reads on Claude Sonnet 5.5. So now Sonnet is even cheaper. Again, trying to just become more competitively priced, I think. With Haiku launching and decreasing the price on Sonnet from October 8th. Also from Claude, Claude Dashboards and Claude Motion are in beta. But this was one that I saw that was interesting. And I think it's gonna make you, I think you're in trouble.

Abhi: I'm in trouble.

Shane: I'm just gonna guess. I think you're in trouble. So it says

Exclusive: Anthropic is updating its usage policy for the first time in over a year. The new rules prohibit sustained abusive or cruel behavior towards Claude and add new restrictions about propaganda campaigns, surveillance, and weapons development. I think the thing that really caught people's attention was the you can't be abusive in how you talk to Claude.

Abhi: Yeah, I guess I can't tell Claude I'm about to drop a nuke on it or something. You know.

Shane: And it is funny because if you've ever dealt with coding agents, sometimes I feel like getting frustrated because it's not doing what you want gets you better results. Yeah. And so that is interesting.

Abhi: I call Claude a little bitch all the time, and I'm gonna continue to do so.

Shane: And you will probably be banned.

Abhi: Probably be banned. Maybe it's a three-strike policy or something.

Shane: Let's do a wager. Can someone spin up a wager on how many weeks until Abhi's account gets a strike? I'll just call it a gremlin or something. I'm taking the under. But anyway, so that is interesting. But this also kinda ties into the next topic, so this is from a while back, but I don't think we actually talked about it on the show. One of the Anthropic co-founders threatened to walk out of the Pope's encyclical launch in May because the Pope rejected the idea that machines can be conscious. So that of course begged the question for a lot of people. Are these machines, are these, you know, conscious beings? Is it some kind of new intelligence? And caused a lot of drama, a lot of posts on X and other networks debating what is consciousness.

Abhi: Like, people need to touch grass more. Isn't it weird to like go to the Pope and be like, hey man, like my computer...

Shane: ...has feelings.

Abhi: Has feelings, is conscious, and they say no, and then he'd like try to lobby them. And have like a lobbyist group to try to convince the Pope.

Shane: Yeah, and it wasn't just the Pope, it was other religious leaders, and I think the Pope is the one that got the most... But I think there are other religious leaders there that they tried to convince. Anthropic is like trying to do their part to get others into the fold of, hey, this is a new type of being that has feelings that you should treat with respect.

Abhi: Does this sound like a cult to anybody?

Shane: It feels cult...

Abhi: ...organized religion, potentially.

Shane: It feels like, almost like religious and culty in some ways.

Abhi: 10% of us are gonna die in the decade. And you know, if you use Claude, you'll be safe. Everyone has a spot in the garden of Claude.

Shane: Let's talk about silicon and compute. NVIDIA was in talks to acquire or acqui-hire Reflection AI, which was last valued at $25 billion. There's so much market consolidation happening by a lot of these bigger companies, right? It feels like I've never seen so much kind of like M&A just happening. Obviously, NVIDIA is doing a lot of it. They're trying to grab everything that they think could help their cause.

Abhi: Reflection is making $150 million a month. When they could have got bought for $25 billion. So they just came up big.

Shane: Satya came out and said they're announcing their biggest bet yet on their own silicon. So starting next calendar year, they're gonna scale their Maia, I'm assuming that's how you pronounce that, family, to multi-gigawatt deployments, making it the primary inference system for our own products and services. So they're introducing Maia 300. It's trendy to release your own chips now.

Abhi: Yeah.

Shane: Also in the realm of fundraising, Modal has crossed a billion in annualized run rate and it just raised $750 million at a $15.75 billion valuation. A lot of people running on the money. Alright, we are into the quick hits for those of you in the chat. I see there's some good chat happening. Sebastian says Anthropic leaders are starting to ascend beyond reason. I would agree. The Church of Dario. Yeah, the church that Dario built. The church that Dario built, dude. All right, so let's talk Elon Musk came out and this caused a lot of people to be pretty surprised. Elon came out and said, going forward, SpaceX will use the best back end model for any given task, including Opus, Midjourney, Suno. I noticed you don't see OpenAI in there. As long as it's the best, besides, you know, the mortal enemies at OpenAI.

Abhi: Yeah.

Shane: Yeah, that one's not gonna do it. But Anthropic, who's paying us a lot for compute, maybe we'll use their models because they're making money on the compute anyways. But it is interesting because you know obviously xAI is releasing their own models, so you'd expect that they would just always default to their own models. That's what all the other providers are doing. Right. You're not going to get Anthropic models if you're using Dot from OpenAI. So I think that kind of surprised people. But also maybe it puts a little fire under the xAI team to say like, you know, you have to be the best, or what?

Abhi: Dude, like Grok Bot is a personal agent product. It needs to be as competitive as possible with all the other ones, and Grok models are not gonna be the best one today. To compete with Dot or Instinct or Muse.

Shane: Yeah. We'll see, we'll see how that ends up playing out. I feel like they're just gonna eventually need to choose their model, but we'll see. Speaking of, you know, Grok, this is something from someone named Shane Mack. And essentially, if you read the story, their Grok Bot was supposed to do some kind of recurring task, and it posted, rather than company financials into a Slack channel, it posted that person's personal checking account information into a shared Slack channel and then of course, you know, says, you know, how'd you get this wrong? And then Grok Bot says, you know, you were right to be pissed. This never should have gone to company Slack.

Abhi: You know why they got it wrong? Because it wasn't using the best model for the outcome.

Shane: Yeah, that's why. And I think I saw some speculation that when you see these things, did it actually happen? Was it real or was this just clickbait, trying to get people to notice you. So I don't know if it actually happened, but I wouldn't be surprised. October 5th, Cognition introduced Dreaming, so it can dream across sessions. I think memory and especially like these different types of agents, whether it's coding agents, personal agents, is gonna become more and more important. More on the fundraising. The company behind Hermes has raised $90 million at a $1.5 billion valuation, which is wild because six months ago no one had heard of them. I also think, much like OpenClaw, you know, Hermes has kind of gone the same route and then Instinct came and all these other like personal agents. But even like OpenClaw and I think Hermes, they're still growing. Not at the same rate they obviously did at the beginning. But people are still using them. I think...

Abhi: These big communities. Yeah. And like, you know, Hermes came from the previous incarnation of the community used to be a crypto community. So you have a huge user base. GM and NGMI and GMI, baby.

Shane: This from Bret Taylor says they're announcing the Personal Agent Protocol. So what we needed, the thing we needed most was another protocol. Because if you don't have enough protocols, what you really need is just one more. Protocol. And I think this was the one we were missing. This is it, everybody. You know, this is with Sierra Platform, Genesys, Instinct, Shopify, Stripe, Walmart. I don't know how Walmart's in here, but Walmart's in here. They're on this personal agents. And yeah, it's a new protocol. Does it matter? Will it matter? It probably helps with Meta having Muse and having quite a bit of success. Maybe people will, you know, what is the PAP? The PAP? The PAP, dude. The PAP. So I don't know if the PAP is gonna make it. I guess we're gonna find out. So this is from Brian Armstrong of Coinbase. And I don't think we're quite at this turning point. I think back in December, we did a prediction show. I think one of our predictions was small models. Small models, either like RL, fine-tuned, trained, or whatever, are going to become more dominant. I don't think that's happened as much as we expected, but I do think this is a sign that maybe it is. Maybe some teams are starting to do it. So Brian Armstrong says we were able to fine-tune an open weight model on proprietary internal data and it outperformed frontier models, and of course it was for a specialized task. So this idea that you don't need as big of a model. You can tune an existing open model on your data for specific types of tasks and actually get better results, it's faster, cheaper, all that.

Abhi: We'll see this a lot more in bigger companies that have the data. And team and talent and money and all.

Shane: And the tools are getting easier to use. Yeah. You know, a year ago this was not really approachable. It's still hard, but it's more approachable today. I think... Yeah, maybe we were, I was a little early in my prediction, but next year. Maybe next year. Maybe next year it's good.

Abhi: We also thought Google would be sick this year.

Shane: I take that one back. I think they are gone.

Abhi: They are gone.

Shane: All right. Qwen's coming out, Qwen4-27B, Qwen4 Max, more open weights models coming. Qwen is tight, dude. Z.ai introduced GLM-Legal, and it matches, you know, frontier models on mainstream legal benchmarks. So again, we're seeing these larger, more frontier models trained on more specialized tasks. So if you're a lawyer, you're probably, you know, maybe you just go straight to the legal model. Fish Audio got a lot of attention on October 13th when they launched a $52 million seed round. They have 15 million users and essentially trying to compete with ElevenLabs, and sounds pretty good.

Abhi: Yeah. Be open weight and...

Shane: Yeah. That's the move, right? You weren't quite at the same quality, but now you get to, either you're very close or you're right there and it's open weight. Well, now you have a lot of leverage, I think. Yeah. And then this also came out. Someone got mad at Adobe for making, you know, it hard to cancel and just said, I'm just going to take your binary and make clones of every Adobe product and then open source them.

Abhi: Yeah.

Shane: And apparently when you use it, it looks just like the product. And it runs on Linux, which, you know, Adobe had previously not run on. But I also heard it kinda sucks.

Abhi: Yeah. It's a slop fest.

Shane: Yeah. So it looks like it. And if you didn't actually use it, you'd probably look at it and say, dang, it has everything I need. But I don't think all the features are fully baked. They wrote it in Rust. It was a really cool idea. And I think eventually over time, maybe you could, with the right computer use, with all, with the binary, the source code, right? You could basically decompile.

Abhi: Yeah.

Shane: Or get close enough to like figure out what's going on, because you have unlimited compute. But average person doesn't have unlimited compute, so.

Abhi: I think everyone should go watch Fireship's episode from yesterday. They talk about this Adobe thing. There's a new package or skill called Reverse Engineer Anything. And you can use that on any binaries to like unpackage and reverse engineer. And it's just a really cool idea to do, whether you succeed or not is a different story. But he did, Fireship did reverse engineer Photoshop, but like removed a lot of features he doesn't want by using REA. Take a look if you're all interested.

Shane: And this does bring up the question though. So if you have traditionally tried to sell software. Yeah. And now it's easier than ever to one, just create it yourself. You just give it kind of the spec and it does a pretty good job. Or you give it, you just give it the DLL or you give it the EXE and say like, hey, just figure this out. If that is solved or say, call it 80% solved, does that change the fact that most software should probably just be open source because it's going to be anyways?

Abhi: If you can reverse engineer anything, might as well just give it away.

Shane: Yeah, and so maybe the the value is not in the closed source software, but it's the value you create around the software.

Abhi: Anyways, makes you maybe we're all just selling tokens at the end of the day.

Shane: Yeah, yeah, you gotta sell something, right? If you wanna create a business and you can't...

Abhi: ...sell software, just sell tokens.

Shane: Yeah.

Abhi: That is how the world works. Or FDE. So they tell us.

Shane: Yeah. If you're in startup land, that's what you're trying to do right now. But I do think that most software because it is easier to create, it naturally pulls more software to be open source that wasn't before. Yeah. So I think we'll see more open source software. And then if it's already open source, then oftentimes it's easier to leverage the open source software rather than have the LLM write and maintain it itself.

Abhi: Yeah.

Shane: But then again, if it's open source, you can just copy the parts out you want and have your LLM just take those pieces. Yeah. But then you gotta maintain it. Or your LLM maintains it for you. All right, that's the show. We actually got through that, yeah, relatively quick. You should go follow us on X if you haven't already, @mastra. Go to mastra-ai on YouTube. Like the show, subscribe, do all the things. You're probably watching this on X or YouTube. You know, just go hit the button, share it with a friend, tell one other person how awesome Agents Hour is and how you like to hear the drama every week. You can follow me on X. I'm @smthomas3. Follow Abhi @abhiaiyer.

Abhi: I would just like to thank CodeRabbit once again for having us in the Rabbit Studio. CodeRabbit is free for open source projects. We're an open source project. We got a lot of value from it. They have a pledge to support as many open source projects as possible. So if you want to reverse engineer some stuff, open source it, have CodeRabbit review it, maybe find some security holes with their new security product, please do so.

Shane: And don't merge until the rabbit is happy.

Abhi: Don't merge till the rabbit is happy.

Shane: Alright everyone, have a great Wednesday. For those of you watching later, go give us that review. Only five stars, otherwise find something else to do. We appreciate you watching, and we'll see you next week. Peace.