Grok Bot Doesn’t Run on Grok Anymore?

Elon Musk posts a message on X and says, By the way, Grokbot now uses Opus 5.5 as default for everything. He wants to light a fire under his own team, saying, Hey, our models are not good enough. We have to use Anthropic's. Do better. That is insane that Grokbot does not use Grok under the hood. Tomorrow was invented in the last 30 days. It's the Last 30 Days podcast with Matt Van Horn. What our builders tell their stories of the future being born. What's real? What's next? Who's breaking through? The hottest minds in AI are talking straight to you. It's the Last 30 Days podcast. Hit it tomorrow, today. Matt Van Horn on the mic, leading the way. Build, ship, ship, learn, create. Tomorrow was invented in the last 30 days. All right, and welcome to Last 30 Days. Let's talk about the news and what's been going on in the last 30 days. R.E.A. by your friend Morluto. Tell us about Morluto and R.E.A. I've heard everything about reverse engineer anything this week. It's been unbelievably popular. It's been the most explosive repo since OpenCL. So went from— You know, within, I think it's 12 days from almost zero to 255,000 stars on GitHub. Like, those numbers are bananas, and it's really just striking a chord. And I think what's interesting is at the same time, while this tool is amazing and I'm a contributor to it and I'm a big fan, what's fascinating is the models have gotten so good that they can also reverse engineer anything. Like, if I go back to the printing press, which is ancient now when we started that, I want to say that was in like May that we launched the printing press. Like, that was about reverse engineering APIs. And so now it's graduated to full applications. You know, it's an MCP wrapper. The skeptics keep pointing that out. Better Stack put it plainly: it's actually just an MCP that connects your agents to popular reverse engineering tools like Hopper, Ghidra, and IDA with some skills on top. The same post notes the number one is a wrapper around number five, and the wrapper pulled 38x more stars than the tool underneath. Nate Herk, who was an early printing press user, testing, I don't think this is anything really that new. He did say that pointing an agent at one specific feature to copy is more targeted than earlier vibe coding. I mean, I think this is all happening at once. I think the technology is here that you can reverse engineer things, that the models are getting so good, and if you have skills that are expert in looking for the right things, that's extremely valuable. And I think that's a lot of the magic that REA is doing. And everything is open source now is the cultural framing. Better Stack connects REA to ArcCraft, a set of Rust rebuilds of Adobe-style apps, someone remade Photoshop, Illustrator, etc. Practitioners say it's easier but not easy. The most useful thread on Locallama, the top comment says, The stories you hear about how easy reverse engineering is now with AI are absolutely exaggerated. It's doable, not easy. The one that I'm trying to do right now, so my six-year-old got a mod retro for his birthday, which plays original Game Boy cartridges. So you can't, like, install software on it. Literally, you have to, like— Put a cartridge in. And so he actually, he's playing my childhood Game Boy games, which is kind of fun. But anyway, my favorite childhood game ever is a game called Super Nintendo, Final Fantasy III, which is more advanced than a Game Boy. Game Boy is like older technology compared to Super Nintendo. And so I'm using Reverse Engineer Anything to take Final Fantasy III and to de-feature it to make it fit on a four megabyte cartridge, whereas the original game is much more advanced than that. Anyway, it's built, over the last three days I've been building it. It's built 60% of the game so far. I've played the first like 15 minutes. It actually works. It like took the soundtrack and like lowered the res to whatever the MIDI defaults are on it. So it's actually working. And I have Claude Code running the skill, and I have Claude Code using Luna Router to do most of the work. So Luna Router is taking all the tokens, and it seems to be working. So why do you think people are losing their mind over this? Because a lot of this is not new, right? You could have six months ago, you know, pointed your agent at a repo and had it copy any app on the internet. So what do you think is different this time? Well, so a repo implies it's open source. This can kind of look at everything and strip it apart. And so I think that's a lot of the magic. And just LLMs have gotten very, very good at doing that. And so it's unlocking a whole new world. And so kind of the freakout, the panic is software dead? And I think the answer is no. We're seeing this with the big labs where every couple weeks a new feature comes out. Let's go with personal agents, right? Remember back in the day when Grokbot launched? It's like, oh my goodness, this is crazy. And then like Instinct launches. And then Muse launches, and then Dot launches, and then Pho launches, and then Wabi launches, right? And what's, like, again, everyone was working on these similar things, but, like, you can learn from each other's creations so quickly, and it's so easy to write software now that I think it's not about the death of software. I think it's about the death of building software once and forgetting about it and that being good enough. I think that's what it's the death of. Whereas it's just now a crazy race to keep innovating. It's like globalization at scale, where, like, you know, something happens in one country, like, oh, cool, like, let's go with, like, Pringles. Almost every country in the world has Pringles, right? But, like, go back 50 years, they didn't have Pringles there. They had maybe potato chips in most places, right? But it's just like, okay, now we know how to produce a Pringle supply chain across the world. It's kind of like that. I think we're just going to see much faster, better software. Is reverse engineer anything the Napster of today's generation? Because there's rights that you're sort of taking apart and decompiling when you use REA. So where do you think that the IP goes when you have REA available? I think it doesn't matter. Like, I think that because anyone can build anything, like, it's, again, it becomes a problem for companies that stop innovating, that are just trying to live off of the old way of doing things. And so I think that's the problem. And so I think there's going to be lawsuits, there's going to be things like that. But in this new world where anyone can see anything, here's something that's interesting. A lot of people don't know this. So my last company was a hardware company. People know that. But a lot of people don't know that. Hardware, mechanical design is inherently open source because you can literally, like, so like when we started June, we didn't know how to make an oven or a toaster oven. So it was the first thing we did. We bought a bunch of toaster ovens and we ripped them apart, and we saw how they were made. And you can see everything, right? There's no, like, secret encrypted part in mechanical engineering for a product like a countertop oven. And so hardware is inherently open source because once you can physically have one, you can see how anything was done and be inspired by it. And so this is kind of starting to be true for software, which used to be closed. You couldn't see it. You could just kind of see what's happening on the outside, but not what's happening on the inside. Now we're starting to be able to see what's happening on the inside. So software is becoming a bit like hardware in that sense, which is really fascinating. Awesome. Let's switch gears and look at ChatGPT dots versus Grokbot. So I know you're a big Grokbot user, as am I. ChatGPT launched dots, so you get one singular persistent agent versus Grokbot, where you can have many agents. I think that the sentiment is a little bit mixed. Is this really a me-too type of product? ChatGPT had to launch just to be competitive, or is there something innovative here? It feels early, so I'm a massive ChatGPT fan. So kind of ChatGPT is my go-to for anything personal. Like if I'm on the go, I want to see something, like there's a freeway was closed that's causing lots of drama for my wife and I this morning, planning our day. And so screenshotted the freeway closed, sent it to ChatGPT. What the heck is going on here? Use the thinking model to do the best research and get back to me so we know when this problem disappears. Right? It's still my default. And so ChatGPT has some of the best models in the world, and they have one of the best teams in the world. So the team there is so impressive. By the way, it's funny because Grokbot is a bunch of bots and— You called it Dot, but it's actually Dot, which is just one dot. But I think it's early. I think that they've got great models and a great team, and so I'm very excited for it. It's not a big part of my workflows right now. I have not set up a lot of automations like I have within Grokbot, which is still my favorite. But it's one to watch. And again, similar to Muse, rushed it out very fast because they wanted to compete with Instinct. I believe Dot rushed their product out very quickly, but I think it will get better very, very fast. My biggest complaint, and I've been loud about this on X, is it needs Tailscale support. I really, really need Tailscale support from a security perspective. And once it has that, I will be very excited. But I have it working in Agent Tin Can with the browser extension, so you can use Dot within Agent Tin Can. It's just not as good of an experience if they let me install Tailscale on my cloud machine. So when looking head to head, ChatGPT Dot is best for workflows. For existing ChatGPT Pro users, it's sort of like your main concierge agent within ChatGPT and Codex, and you can use everything there, versus Grokbot, where it's best for solo operators running a multi-agent business, essentially. So I think there's more to see. We saw the launch of Hark. Hark is the latest personal agent, and what I wanted to do is compare it to Instinct. So, Matt, do you know a little bit about Hark? Have you used it? I have. I got early access, which was nice, and it was pretty smart. It's funny. It actually, I got really excited because I have all the proactive agent sending me stuff. Oh, would you like to do this? Oh, would you like to do this? And most of the time it's like, No, shut up. But Hark was like, Hey, you have $250 in SeatGeek. And I never use SeekGeek, but I was like, Oh yeah, I think I sold a ticket there once. And it's like, Do you want me to get it? I'm like, Wow, thank you, Hark. You found something I hadn't seen, and you're offered to go get it. Go get it. And then the TLDR is a few minutes later says, Sorry, I got bot blocked. I can't get your money for you. So I did not collect that money. I gave up on that task. But I was really excited for a second. I have this—it's not Hark-specific. This is a generalized problem in the industry. And so, yeah, but a lot of people love it. Like Alex Finn loves Hark. I sent him a message and I was like, Dude, are you getting paid by Hark? He's like, No, it's a great product. And he just, again, he gets excited about everything shiny and new. But for good reason, people do like it. And it's funny because I'm overwhelmed by agents and kind of well set up in my setup, but I'm impressed how much love it's getting right now, considering everything going on. What I want to know is, you have the explosion of the personal agent, Instinct I know lives on SMS, for example, only. Are all of these personal agents using the same underlying LLM models, or do they have some special harness or wrapper that they're using? Here's what I know. So the rumor is Instinct is running a tuned, custom-tuned Chinese model, and it's very good. I'd say my favorite thing about Instinct is how it will go through walls and doesn't like saying no to you. Whereas I feel like Muse and Dot love to say no to me. They're like, Sorry, I can't do that. Sorry, I'm not comfortable doing that. Sorry, you'll have to sign that for—I can't sign documents for you. Whereas Instinct's like, I will sign all the things. Just tell me what to do. I'm hungry to please. And I'm just like, You're my type of friend, Instinct. And so Hark, I do not know the answer. My guess is it's similar. Something really fascinating that Grokbot did the other day. So Grokbot was running an unannounced Grok model that was decently fast but not very intelligent. And Grok has intelligent models, but this is just like the default one. And so I'd actually set up a bunch of hacks to, like—because Grokbot can talk to Cursor, and Cursor has access to all the Anthropic models. So I would actually have set up certain automations within Grokbot where I'm like, Hey, please send this to Fable Five. And so I'd set up a lot of automation for that because the Grok model was a little dumb for my liking for everyday work. And then what's amazing and crazy from a cost perspective, people are not giving this that much love, is Elon Musk, just out of the blue, because Elon just, you can just do things, he posts a message on X and says, By the way, Grokbot now uses Opus 5.5 as default for everything. And it's just like, holy crap, like that is so expensive. They haven't raised prices. Like, he just, he wants—I think he was doing two things. One, he wants everyone to have a great experience in Grokbot by running a very smart model. The second piece is he wants to light a fire under his own team, saying, Hey, our models are not good enough. We have to use Anthropic to do better. Like, that feels like a very Elon thing to do. That is insane that Grokbot does not use Grok under the hood. And I made a, like, IQ test, and Grokbot did not perform that well. The first time I ran it, and the second time I ran it, it was way up there because it started running Opus. So that's what actually inspired me to run the second one. Wow. So bottom line, between Hark and Instinct, choose Hark if you want a personal agent today without the invite codes of Instinct, and you care about not having your data trained on. Interesting. Keeping expectations low on speed. Versus choose Instinct if you can get the invite, which, thank you, Matt, you gave me the invite for Instinct, and you want the most capable, most proactive agent right now, and that can make phone calls and chases deadlines and super proactive on SMS. So Fish Audio announced a big fundraising round. They released their Drama 3 model. So this is an 11 Labs competitor, so think like voice AI. So the new TTS king is in town, Fish Drama 3. Local Llama was impressed but skeptical. The top replys in the thread poked at the launch claims. On the Nvidia H100, add, you know, your local H100, ha ha ha. Never ask a TTS company where did they get their data from, said of the demo voice, The first girl sounds absolutely wasted. I love Reddit, by the way. The first girl sounds absolutely wasted. The Uncle Roger launch video spread widely. I believe it had 11 million views. People talk about Fish Audio as a cheaper ElevenLabs. Fish Audio could do to voice AI what DeepSeek did to LMs. Hobby projects on one GPU to 21 million ARR with 22 people in a year. ElevenLabs should be nervous. A lot of the buzz is paid. Most of the large YouTube and short-form reach was sponsored or affiliate content, but it's supposed to be very, very good. Fish Audio just released Drama Tree, their most controllable voice model yet. So today, Uncle Roger here to see not only can this model sound like me, but can it also roast like me? Do you know who this Uncle Roger character is? I was not aware of Uncle Roger before watching this video, and I would love to know who this is. He's a YouTuber from Malaysia, from my quick search, but he's definitely a big deal. GTA 6 is coming. Are you going to play, Justin? I grew up on Vice City. That was my absolute favorite game ever, but I have not kept up with it. I would love to get reintroduced to GTA. The launch trailer came out today. It's everywhere. It's a really cool launch trailer. The top comment on YouTube: For my great-great-grandchildren, your grandpa was here, because Grand Theft Auto 6 is finally coming out next month. The trailer was also a music launch. It debuts Flower in a Hurricane from Justice featuring Tame Impala and SZA. Rockstar released it as a standalone single and named Justice as the game's original score composers. I have not listened to it. I mean, I watched the trailer, but I have not listened to the song yet. GTA 6 is in a happy meltdown on Reddit. Spinoff posts already up, a side-by-side with GTA's launch trailer using the same narrator, frame captures, and the raw footage. Not everyone is sold on the visuals. 1X user pushback on people complaining about GTA 6 being downgraded and looking blurry calls out the 30 frames per second double standard. A Bloomberg Inside Story new screenshots built hype before the trailer, and award eligibility and pre-launch reviews are confirmed. The cutoff is November 20th, so GTA 6 is eligible this December. And a leaker is holding the game hostage. A leaker named Cyberleak claims to have the full playable build. They threaten to release it unless Take-Two brings back physical discs and apologizes. So we'll see if that happens. That's the most interesting part of the story, that a leaker is holding the game hostage, which I'm not sure I understand. This is like an Edward Snowden moment, but he doesn't want money, it sounds like. He wants it on a physical disc, which makes no sense. Let's talk about Potato. Lauren Tan, the 2,500 PR-a-month engineer, is this month's Agent Workflow celebrity. Lauren Tan now works on Grokbot and Cursor at SpaceX AI. The story about merging thousands of PRs is the most-watched coding agent conversation of the window. She's the creator of Pystack and has shipped thousands of PRs a month at SpaceX. Tan says the whole project started from burnout. I took a month off because I was feeling kind of burnt out. I was really interested in the idea of how I extract my own abilities and give this to the agent. A lot of how I'm able to do this is through trust. If you set up your environment for your agents really, really well, you can leave them working unattended. Viewers loved it and asked about the cost. The top comment on the stream is a joke: The great video would be even better if each time a presenter took a sip out of their large water bottle, the bottles became comically larger and larger. The sharpest one asks about money. People who work for AI companies really need to get asked, like, What is your equivalent API billing monthly burn rate? Pystack is spreading past Cursor. Pystack is starting to explode. Grokbot is the day job, and today's X feed shows it. And I just want to say, like, I'm a Lauren fan. We've messaged a few times. I've never talked to her one-on-one, but we've had a few Slack messages and X messages, and I'm just very impressed by how she's become this, like, mega celebrity very, very quickly. And she hasn't even worked at Cursor one year, so she literally has not even, like— Even, like, vested her one year of Cursor stock, which is crazy because she's become such, like, the face of Grokbot, which is very impressive. And the other thing is that, like, her posts go super, super viral all the time. Elon Musk regularly is reposting her. She said, Hey bot, please pre-order the new Tesla Roadster for me immediately, thanks. So she posted that, which would automatically talk to her Grokbot, and then Elon Musk just reposted that and just said, Fire. So she is very good at the X game, and she is a very talented engineer and doing amazing things. I've been very impressed by Potato. And by the way, I thought it was pronounced Poteto because it's P-O-T-E-T-O, but she told me otherwise. This was fun. Thank you, Justin. Thank you, listeners. This has been last 30 days, and again, please leave a comment. Let us know what you think. We're trying out a new segment type of how we're doing this and separating this from the main interviews, so please, please give us thoughts.