Three labs are clearly ahead of everyone else building language models right now, OpenAI, Anthropic, SpaceXAI, with Gemini close enough behind that it might not stay a three-way race for long.
The first reaction to that, every time one of them ships something new, is the same: how is anyone supposed to keep up? The second reaction, once that settles, is the one that actually sticks: how do I get this into my own workflow before the next one lands?
I've gotten to watch someone do that better than almost anyone I know. Through the ad network at beehiiv, I met Matt Swulinski (previously head of growth at Wispr Flow, now head of growth at Viktor) and he might be the clearest example I've seen of what a marketing engineer actually looks like. He's been on a bit of a podcast tour lately, and I've learned more from listening to him talk through his own process than from almost anything else on this topic recently (posted about it recently).
Watching someone stress-test a tool that well made me want to go find my own edges with it.
Mine turned out to be a green card application.
Back in 2023, I paid an immigration law firm somewhere between $3,000-$4,000 to handle my case. There was no version of that process where AI could have helped, the tools just weren't there yet. This year, mostly out of curiosity, I ran that same case back through Grok Bot. I gave it everything except a handful of documents I couldn't share for timing reasons, and told it explicitly to stay out of anything like extremely sensitive — just tell me where to pull it from, don't fill it in yourself. It filled out the rest of the form correctly.
Like 100% correctly.
I don't fully understand the security implications that comes with handing a model that much of your own paperwork, and I know people close to me who want nothing to do with any of this on principle. I can't tell them they're wrong to feel that way. I just know what it did with a case I couldn't have afforded to get wrong the first time — and it's absolutely mindblowing.