Gemini 3, Walled Gardens, and the Editor Arms Race

Tim Williams: Hello and welcome to a brand new episode of Rubber Duck Radio I am your host Tim Williams lead software developer and alongside me again is the venerable Paul Mason Paul Mason: Hey Tim, glad to be here again. Do you mind if I kick this one off? Tim Williams: Go ahead. Paul Mason: Let's talk about Gemini 3's launch this week. Have you been flooded like I have with an insane number of viral social media marketing campaigns. Tim Williams: Thank god it wasn't just me. It has been insufferable. Paul Mason: No, and it's the same thing over and over. Very transparent. They begin with high praises of Gemini 3, then go into some light use case where it trounced the other models. Paul Mason: Then they present the same overwhelming benchmark numbers from the same set of graphs that every other influencer is using. Tim Williams: I know, the similarities between all of them are what makes me suspicious. Typically when we've seen new models come out, there is some viral traffic, but rarely are all of the influencers aligned on the same use cases and benchmarks. Paul Mason: It smacks of an Alphabet viral marketing campaign to me. But that's neither here nor there, have you played with the model yet? Tim Williams: Extensively. But before I throw out my opinion, I'd love to hear yours. Paul Mason: Fair enough. I'll try to keep it balanced and unbiased from the bad taste in my mouth I get from the viral marketing campaign. Let's start with the positives. Paul Mason: One strength I noted is that is very strong for building Greenfield POC type applications. It's almost as if they fine tuned the model to focus on that use case. Paul Mason: It's able to plan, build and debug as it goes along. Mind you, these are fairly basic applications, but it does appear to be as strong or stronger than Claude Sonnet 4.5. Paul Mason: One place where it failed, and this is probably more on the Cursor side, is that it can't create 'todo' list items. Which is strange because even most of the 'budget' models can do that. Paul Mason: I just chalk it up to the model's fine tuning towards generating the interactive, or magazine style output that we see in some of the influencer reviews. Tim Williams: I noticed that. It seems to have a very decent ability at spatial awareness and design chops. Paul Mason: Right, that's the other thing I wanted to mention. It seems to be pretty strong in UI generation compared to the other models. In both markdown style and layouts and HTML with React and Vue. Tim Williams: That's been one of my gripes with the leading models all along. Their design chops are nearly nonexistent. And as we've said before their CSS generation is atrocious. Paul Mason: The CSS code Gemini produced wasn't a lot better than the other leading models in my opinion. But the overall design aesthetics do seem to be better. Tim Williams: Agreed. It must have better training on thing's like Google's material design systems and internal design style guides that companies like Anthropic and OpenAI won't have access to. Paul Mason: Something like that probably. Anyway, here's where it seems to fall apart. When you're planning a big feature to an existing system it does great at creating a first pass as the feature. You go back and forth a few times with the model to refine the plan, then switch to agent mode and begin to build out the plan. Paul Mason: Everything goes pretty smoothly until you hit about 75% of the model's context window then it starts to unravel. The model seems to forget important parts of the plan. Tim Williams: Did you notice that it doesn't create bullet points in cursor to check off as it goes along? Paul Mason: I did, but I just assumed that was a Cursor issue. Tim Williams: It's probably some of both. I bet Google built this thing in tight coupling with their new code tool 'Antigravity.' Paul Mason: That would make sense and explain some of the weirdness in using it in Cursor. Tim Williams: My thoughts exactly. I ran into the same issues though, the model seems to completely lose its train of thought before you even hit the token limit. Paul Mason: We could speculate all day long, but what we might be seeing is the model makers starting to construct walled gardens. If that's true, that kind of sucks for developers at large. Tim Williams: Definitely. I suspect that Cursor is doing the same thing. Have you played with the 'Composer 1' model much? Paul Mason: Not a ton, what's the deal with that? Tim Williams: It is amazing at interacting with Cursor's tools. It is fast, it always creates proper to-do lists in plan mode. It has a cool feature during plan mode where it shows you a multi choice list of questions and answers one by one that you can click on. Paul Mason: That's funny. It's a return to traditional UI for models. If you step back and look at it, we're training them to use tools to present UIs that we're used to using. Tim Williams: It makes sense because you don't want to be presented with a huge list of questions in plan mode then have to type out a list of responses as you scan back over it. Paul Mason: Yeah, I think you're right then. I think they might be slowly starting to make walled gardens to keep developers from jumping ship every time a new model is introduced. Tim Williams: The problem with that strategy is, both you and I have had subpar experiences with Gemini 3 and it may actually be a great model. But we certainly can't verify that from our experience. Tim Williams: I feel like the model makers are doing a disservice to themselves if this is their plan, it's going to put a bad taste in the mouth of developers using these tools. Tim Williams: I want to be able to pick the best tool for me, then pick the best model to use in that service. I don't want to be walled into a specific tool just to get access to a model's intelligence. Paul Mason: Same. Tim Williams: So let's talk about what happens if this really is the strategy. If the model makers are building walled gardens, and the tool makers are building walled gardens, and we're all just supposed to pick a side and stay there like it's 2010 all over again. Paul Mason: Yeah, the problem is we've seen this movie. And every single time a company tries to box developers in "for their own good," the ecosystem rots from the inside. You can lock me in, but you can't stop me from noticing the mildew on the walls. Tim Williams: Exactly. Think back to Facebook's platform era. Remember when everything was going to be written with their APIs and their little custom UI layer, and they wanted developers to build entire businesses on top of it? Paul Mason: Until they changed the rules mid-flight, shut down half the APIs, and everyone lost their businesses. A great time. Ten out of ten for platform stability. Tim Williams: Or Apple before they loosened the reins on web tech. For years the App Store walled garden slowed down entire categories of innovation because developers had to wait for Apple to bless a feature before anyone could use it. And what did that get us? Five years of "no, you can't have proper web push notifications, Tim, because we don't feel like it." Paul Mason: Same thing with Google Inbox. Amazing app. Incredible UI. Everyone loved it. And then , poof , gone. That's what happens when the whole ecosystem is trapped behind a single company's strategic whim. And now they want us to hand them not just our apps, but our agents? Tim Williams: That's my fear here. If Gemini 3 only works properly inside Antigravity, and Composer 1 only works properly inside Cursor, and the other tool makers follow suit, then developers lose the one advantage we've always had , choosing the best combination of model and tool for the job. Paul Mason: And ironically, the quality of the models will get worse, too. Because you don't get the brutal feedback loop of "hey, something's broken" when everything is insulated by the tool. Tim Williams: Yeah, the walled garden becomes a padded room. And the users , meaning developers , get infantilized. Less control, fewer escape hatches, and a dependence on whatever UX decisions the platform teams cooked up that quarter. Paul Mason: The funniest part is that walled gardens never die quickly. They slowly decay. It's like using Flash sites in 2012 , you can smell the death on the air, but the corpse just keeps twitching. Tim Williams: So if this is the strategy, long term it just means fragmentation. People will spin up open alternatives. Devs will migrate back to tools that don't impose those constraints. And the companies that forced the lock-in will start hemorrhaging credibility. Paul Mason: Developer trust is a one-strike game. You break it, we don't come back. We fork something and keep going. Tim Williams: And the irony is that the companies doing this think it's about retention. But every time someone locks down their ecosystem, they forget that developers treat walls as puzzles. They always find a door. Or they just build a new house and leave you alone in yours. Paul Mason: Exactly. If you want retention, give us freedom. If you want a walled garden, watch us leave. Tim Williams: Alright, let's get a little comedic with this because I swear these walled gardens are starting to feel less like developer platforms and more like those HOA-controlled neighborhoods where you get fined for having the wrong shade of beige on your front door. Paul Mason: Exactly. Gemini's over there like, "Uh, sorry, you can't run that model outside Antigravity because the grass height in your prompt exceeds community guidelines." Meanwhile Cursor shows up in a little golf cart to tell you your plan file is three lines too long and violates subsection C of the neighborhood charter. Tim Williams: Oh, and don't forget the monthly HOA meeting where some platform PM stands up with a PowerPoint saying, "We heard your feedback, and we're pleased to announce… absolutely no meaningful improvements. But here's a new decorative fountain." Paul Mason: The fountain is the worst. It's always something useless like: "We redesigned the settings menu again!" Thank you, Steve, but nobody asked for a different shade of blue. I want my model to stop hallucinating API endpoints. Tim Williams: And of course there's always that one neighbor , usually a model , who keeps promising to fix their issues but somehow never does. It's like, "Hey Gemini, your CSS generation broke the neighborhood fence again." And Gemini's like, "We're aware of the issue and actively looking into it." Meanwhile the fence has collapsed on the dog. Paul Mason: The dog being my entire build. Tim Williams: Exactly. And then the HOA wonders why half the neighborhood suddenly moves out in the middle of the night and starts a commune in the woods using open-source weights and NVIDIA GPUs. Paul Mason: Yeah, they're out there in the wilderness building their own little off-grid utopia. No PMs, no forced UI updates, no subscription-only lawn watering schedule. Just raw models running free like deer. Tim Williams: And then someone shows up like, "Hey, do you guys have fire codes?" And the open-source community just shrugs and says, "We have GitHub issues. Same thing." Paul Mason: Honestly, I'll take that any day over an HOA telling me I can't use my model's full context window without written approval from the architecture committee. Tim Williams: That's the future if these companies keep going down the walled garden path. We're all gonna be living in a gated community with a very pretty fountain and absolutely no freedom. Paul Mason: Meanwhile the rest of us will be in the woods, building a cabin with a 4090 and a dream. Tim Williams: And the thing is, that commune in the woods? It's not even hypothetical anymore. It's real. It's called OpenCode. It literally exists. It's an actual open-source alternative to Cursor being built out in the open, no HOA, no PM telling you your lawn can only be React-green. Paul Mason: Right, it's the folks at SST , same team that's been pushing the "developers should own their tools" philosophy for years. It's like they got sick of living in the gated community, packed their bags, hiked into the forest, and said, "Alright, we're building our own town. Tools are optional, debugging's free, and chores rotate." Tim Williams: And if you look at their GitHub , which is already surprisingly active for something so early , you can totally see the commune energy. Everyone's contributing, everyone's discussing, no one's trying to lock down the garden and charge admission. They're like, "Here are the blueprints. Borrow them. Fork them. Build a treehouse. We don't care." Paul Mason: Meanwhile, back in the HOA, someone's getting a warning letter because their prompt indentation violates architectural guidelines. Tim Williams: Exactly. OpenCode's over there saying, "Do whatever you want. Build your own editor. Connect your own agents. Self-host. Modify the UI. Go nuts." It's basically the open-source Burning Man for developers. Paul Mason: The real question then is , is OpenCode the best open-source path forward? Because there are a few contenders now. You've got OpenCode, you've got Zed going open-core, you've got VS Code extensions tying into local models, even JetBrains is flirting with the idea of opening more of their agent interface. Tim Williams: And don't forget Emacs, which has basically been a rogue commune since the 1970s. Emacs users are like, "Oh, you guys are just discovering freedom? That's cute." Paul Mason: Yeah, like, "Welcome to the woods, kids. We've been living off the land for decades. Here's a Lisp hatchet." Tim Williams: So let's actually debate it. If you had to pick the best open-source alternative today , right now, 2025 , what's your vote? Paul Mason: Okay, so my pick is OpenCode. And it's not even because it's finished , it's because the architecture is right. They're building with the assumption that devs want to customize everything. They're not betting on one model, or one tool, or one vision. They're betting on flexibility. And historically, that's what wins long term. Tim Williams: Fair. My pick is actually VS Code plus local models. Hear me out. It's not as philosophically pure as OpenCode. But it's got the ecosystem, it's got the extensions, it's got the tooling, and it's got the weight of the developer world behind it. You give VS Code a couple more cycles of agent integration and a few more community-driven model wrappers, and it's basically unstoppable. Paul Mason: That's the safe, pragmatic choice. OpenCode is the idealistic one. Yours is like, "Let's build a commune but next to a Costco." Tim Williams: I mean… yeah. Have you tried buying GPUs in a commune? Good luck. Someone's gonna trade a 3090 for a bag of lentils. Paul Mason: The real winner is probably going to be whichever project nails the "bring-your-own-model, bring-your-own-tools, bring-your-own-agents" pipeline without making developers want to stab themselves. Tim Williams: And whichever one resists the temptation to turn into a walled garden later. The moment an HOA creeps in , it's over. Paul Mason: The thing is, the open source tools have no incentive to become walled gardens. Conversely they're incentivized to keep their tools as interoperable as possible, don't you think? Tim Williams: Yes and no. They still want loyalty in the Open Source community, because loyalty helps them push their product forward. Even open source communities have an imperative to spread and survive. Paul Mason: Yeah, I guess it's just their motivation that's different. They don't have shareholders breathing down their neck every quarter and bashing them with the quarterly profit numbers. Tim Williams: Right. Paul Mason: They do have a need to survive though, I see that. Tim Williams: Alright, let's pivot a little. Because with all this talk about open-source communes and walled gardens, I've been thinking a lot about the editor landscape. We've got Cursor, Windsurf, JetBrains AI, Zed, VS Code with a hundred different agent extensions, OpenCode, Replit, literally everybody is building their own version of "the future of coding," and none of them agree on what that actually means. Paul Mason: Yeah, it's like walking into a craft fair where every booth claims they've invented the one true bowl. "Oh this one auto-stirs while you eat, this one uses machine learning to predict what soup you want, this one comes with Kubernetes support for no reason." Tim Williams: And meanwhile Cursor is pushing updates so fast that I swear I restart it more than I restart my Mac. Half the time I'm in the middle of something, I switch spaces, look back, and there's that little popup: "Cursor needs to restart to update." I'm like, "Buddy, I just relaunched you thirty minutes ago." Paul Mason: It's brutal for anyone who organizes their work by full screens on macOS. You finally arrange everything like a little productivity mandala, and then Cursor sneezes and your layout gets turned into modern art. Tim Williams: Yeah, Cursor updates so often it's practically a subscription to chaos. You're paying for the privilege of being surprised at random intervals. Paul Mason: But here's the thing, this is where the ecosystem gets weird. Because you'd think with everybody racing forward this quickly, they'd be converging toward some shared idea of what the "agent-powered editor" is supposed to be. But I don't think they are. I think they're all diverging like crazy, each one trying to figure out its own superpower. Tim Williams: I actually disagree. I think they're converging , just not intentionally. The market is pushing them toward the same center. Cursor was first to nail the "agent in your file system" loop. Windsurf is clearly reverse engineering that workflow. JetBrains is trying to graft agents into the world's heaviest IDE. Zed is trying to turn collaboration into an agentic superpower. Replit is shipping "agent-everywhere" like Oprah giving away cars. OpenCode is building the open-spec version of Cursor. Everyone's sprinting in the same direction whether they admit it or not. Paul Mason: I don't know. I still see divergence. For example, Cursor's superpower is the agent tightly integrated with your local repo. Windsurf's identity is "we have the cleanest UI and the smoothest editing experience." JetBrains is leaning on their deep static analysis, giving the agent more context than God. Zed is betting on multiplayer coding. VS Code is trying to be the universal adapter for local models. OpenCode is trying to become the sovereign editor where you can build your own agent pipeline. These are totally different bets. Tim Williams: Yeah, but at the UX level they're starting to look more and more alike. You've got plan mode, edit mode, scratchpads, multi-turn agents, file diffs, workspace context, inline apply buttons… they're all being pulled toward the same gravitational center. We're basically reliving the "every phone became a rectangle with apps" moment. Paul Mason: Maybe. But think about it , this space is so early that everyone is still searching for their Pokémon-type specialization. Cursor: file operations. JetBrains: static analysis. Zed: multiplayer. Windsurf: quality-of-life and polish. VS Code: ecosystem gravity. OpenCode: freedom-to-the-people. Each one is hoping their trait becomes the golden ticket. Tim Williams: And the question becomes: do we actually want them to converge, or do we want divergence so we get actual innovation? Paul Mason: I want divergence. I want the weird experiments. I want the editor equivalent of a cyberpunk garage where someone adds a plasma rifle to Notepad++. Let people get freaky with it. Tim Williams: And I want convergence. I am tired of learning a new interface every time a team invents "the future of coding." I want one predictable workflow that works everywhere. No more having to remember whether "apply changes" is a button, a slash command, or a chat reaction emoji. Paul Mason: That's fair. But convergence also means stagnation. If everyone settles on one paradigm, that's when the HOA slips in again and starts telling you what shade of beige your editor themes should be. Tim Williams: So what's the compromise then? Paul Mason: Probably this: shared standards, divergent implementations. Like web browsers. They're all rectangles that show websites, but one is fast, one is private, one steals your RAM, and one isn't updated because your uncle refuses to stop using it. Tim Williams: That might actually be where OpenCode pushes things. A spec for model-tool-editor interaction, and then let everyone build their own flavor of agentic coding on top of it. Paul Mason: And if they do it right, maybe the daily forced updates will stop being "chaos grenades" and start being something developers can actually rely on. Tim Williams: I wouldn't bet on that, but hey , a man can dream. Tim Williams: Alright, I think that's a good place to wrap this one up. We covered a lot today , Gemini 3's influencer circus, the creeping rise of walled gardens, the HOA metaphor that got way too real way too fast, and the commune in the woods powered by OpenCode and a stack of dusty GPUs. Paul Mason: And we dove into the editor arms race , who's sprinting, who's tripping, who's updating every twelve minutes for no reason , and whether all these tools are crashing into the same shape or diverging like a litter of hyperactive puppies. Tim Williams: My money's still on convergence. Tools gravitate toward each other. Developers pick workflows that make sense. And eventually the market pushes everything toward a shared center. Paul Mason: And my money's still on divergence. Chaos creates innovation. Somebody's going to stumble onto a superpower the rest of us didn't see coming, and that's going to keep this whole space weird , in a good way. Tim Williams: Either way, we're entering a fascinating era. AI-powered editors aren't just "editors" anymore. They're operating systems for development. And depending on who wins , the HOAs or the communes , the next decade of coding is either going to feel liberating or claustrophobic. Paul Mason: But whichever way it goes, you know we'll be here every week complaining about it, reverse-engineering it, and occasionally praising it through gritted teeth when it actually works. Tim Williams: That's Rubber Duck Radio in a nutshell right there. Paul Mason: Exactly. Thanks for hanging out with us. If you enjoyed this episode, leave a review, send us your dev horror stories, or just shout at us on social media about how we're wrong. That's half the fun. Tim Williams: Until next time , keep coding, keep experimenting, and remember: if your editor restarts itself twice in one hour, you are legally allowed to scream into the void. Paul Mason: It's in the HOA bylaws. Tim Williams: See you next week.