Rapidly over the past few months I’ve felt a shift that is akin to other moments I’ve felt during ai progress. ChatGPT was the first glimpse of what ai can do. Claude Code and Cursor drastically changed my day to day. o1 felt like the first model I could trust with certain difficult tasks. Opus 4.5 and gpt 5.2 became true work horses. And now I find it quite hard to describe what’s going on in this moment but I’ll attempt to below. The models have been incredibly smart for a while and recently they’ve become incredibly capable, but in my own work and life there were still so many gaps that I felt and friction that did not enable me to get the most out of the tools. These gaps are closing, and fast. There are four concrete reasons I feel this is the case and luckily for us, it appears these are across many different products and companies. 1. The models are reaching super intelligence. Hand GPT 3 to someone in 1990 and they would imagine us living in a technological utopia. Fable 5 is still the smartest model currently released and you can feel it when you wield it. It has skill, judgement, and taste that I’ve not seen before and these qualities are clearly emerging as models continue to scale. Everything we’ve said about models doing the execution and humans driving judgement and taste feels like it is dissolving and I feel that most viscerally with Fable. When working on a tough design, asking it to “make it better and more intuitive” yields often better results than I could have thought of. 2. Codex browser use. “At least I will needed to drive things outside the code” spoiler, wrong and it happened faster than I expected. The codex app in particular is the best example here. Not only is it good, it’s fast, can work across multiple tabs, and simply works. We’ve likely all fumbled with earlier computer use tools and not seen results or watched the models get stuck on trivial pages. That doesn’t happen anymore. 3. Amp code, a new one in my arsenal but one of my favorites. The big unlock was friction reduction. They spawn agents in “orbs” with their own computer to write code, test it, and provide you an isolated dev server to work with. Their team said it best in a recent episode of their podcast, Raising an Agent. I forget the quote but something along the lines of “running agents locally took friction to 2, maybe 1 percent. But it was still there. With cloud agents [orbs] the friction to start a new thread is 0 and that has an outsized effect on how much you can ship” I’ve also found isolated cloud agents to clear my mental clutter. I don’t have to track 20 changes across worktrees, I know when I open a thread, what I did, and just what I did is there. 4. Always on personal agents. We got a glimpse of this with OpenClaw but Grok Bot just takes the cake. Simple, easy, and abstracts all the tricky parts. I can simply hand over a few logins and have the same trust I would with a co worker to tackle tasks. I’ve already felt my usage and scope of task grow knowing how helpful it can be. We now live in the age that was only spoken of just a few months ago. You can command a fleet of agents with just your voice from your phone. (Realtime chat in Codex is an honorable mention here). The gaps are closing, the new success metrics will be based around how much you can output and what few gaps you can continue to fill in.

Slop is mostly solved imo. Agents now fail when they don't understand the bigger picture. Here's how I battle this issue, hoping you can just give this link to your agent and get results!

14 months ago I wrote a sentence I did not feel qualified to write. I was alone in a dark office at 4am banging my head off the desk trying to fix a bug while contemplating what my career was becoming. I had tried my hand at a few software gigs and spent time in the startup world, but I did not really know what I was doing or what to do next. I just knew I cared about AI more than almost anyone around me. Then I saw @businessbarista post about @tenex_labs . I applied, and in the application I wrote, "I truly believe I'm exactly who Tenex is looking for." Honestly, I had almost no evidence for that. 14 months later, I not only feel qualified to write that sentence. I know I was right. Tenex has been the best place I have ever worked because it has been the best place I have ever learned. I have built full stack apps, mobile apps, AI systems that scale, and deeply personalized software that would have sounded fake to me a few years ago. To my dismay, I have also learned far more about Google Cloud than I ever planned to know. But the real reason I love it here is simpler than the work. Before Tenex, I often felt like the nerd in the room. I wanted to talk about models, agents, tools, weird workflows, and where all of this was going long after everyone else had moved on. Here, that is everyone in the room. The pace is high, and these are the sharpest people I have ever had the opportunity to work with. But the thing that has mattered most to me is not just how smart everyone is. It is how generous they are with what they know. People explain the thing they just learned, pull you into the details, and ask the question that makes your idea better. You learn faster because the room expects more from you, but you also feel more like yourself because the thing you cared about alone is suddenly something everyone cares about together. Tenex didn't just make me better. It made me feel right for caring this much in the first place and helped me find my people. If you want to find your people, apply here. (my original application form still in our Notion below)

I think too many people see AI labs as just the models. The model matters, obviously. But the better question is who controls the stuff around it: compute, power, data, distribution, products people use, government relationships, and the ability to wire all of that together without lighting the company on fire. At some point the model starts to look like the UI for the utility underneath. So unfortunately, I have to talk about Factorio. If you’ve never played Factorio, it’s just Minecraft for people who think the fun part is supply chains. You crash on a planet, mine rocks, build machines to mine, build machines to build, and eventually build the factory that builds factories. The game is bottlenecks. You fix iron, then power breaks. You fix power, then throughput breaks. You fix throughput, then some tiny thing three screens away is starving the whole factory. That is the AI race now. So, rough map of the factory floor. 1/ OpenAI has the default app. ChatGPT was the first real AI product people used, and when someone says AI, a lot of them still mean ChatGPT. Monster advantage. The problem is the factory still depends on someone else’s infrastructure, particularly Microsoft. Stargate is the attempt to own more of the power bill, but it can also become partner spaghetti around one insane product. 2/ Anthropic has taste. Claude is still the model I trust most when the work requires taste instead of horsepower: intent understanding, steerability, the feeling that it wants to work with you. But trust does not magically buy power. The model serious teams trust can still get outscaled by someone with a wider resource base. Also, the Fable mess was a reminder that government is now part of the game, not adjacent to it. 3/ Google should be absolutely destroying this. They have TPUs, Search, YouTube, Android, Chrome, Gmail, Docs, Cloud, distribution to billions of humans, and more data than anyone wants to admit. If Gemini, Workspace, Search, Cloud, YouTube, and dev tools ever feed the same system, everyone should be terrified. If not, every keynote stays a little bit “wait, which product is this in again?” 4/ Meta, Meta, Meta. With Llama, Meta is trying to make its defaults everyone’s defaults. In open source, defaults are power. The play is probably smart: give everyone the ecosystem, keep the sharpest edge for yourself, and hope open source doesn’t commoditize the thing you’re trying to win. Very Zuck, and probably underrated. 5/ xAI and Elon are where this gets messy. The obvious take is “Grok is behind, therefore xAI is behind.” Maybe, if this is only a chatbot race. But if the AI race is Factorio, Grok might be the wrong thing to judge. X gives him messy human feedback and distribution, Tesla gives him cars, batteries, factories, robots, and a path to embodiment, SpaceX gives him satellites, defense relevance, and government relationships, Colossus gives him data-center execution, and xAI gives him the model. None of this automatically composes, and because it is Elon the factory might just be five burning factories in a trench coat. But if it does compose, the upside is not “Grok gets a little better.” It is that the weird pieces start feeding each other. That is not a chatbot company. That is a factory. Soooo just Factorio. The smartest model matters. It’s just not the whole game...just like Factorio