Codex is the kind of product that deserves a definitive history, and @danshipper's written it! As someone who was personally Codex-pilled by Dan AND got to tag along with to the OpenAI offices for some of the reporting for this piece, it's so exciting to see this out in the world.

I'm working on a draft, and I wasn't happy with how it was coming together on a sentence level in ChatGPT, so I scooted on over to Antigravity to see what Argon had to say. It's happening: the frontier AI duopoly is cracking.

For multiplayer agents, it's really important to know what data the agent has access to, who else can access it, and what they can do with it. Here's a high-level overview of how we built permissions into @every agent to maximize agency, trust, and speed (and avoid catastrophe)

We built @Every Agent to work the way Every works, and that includes how we write. Every Agent comes with all 32 skills from Compound Writing built right in, so now you can draft, edit, and get eight different opinions on what you've written, all without leaving Slack! Happy birthday, Every Agent 🎂

Once again, interfaith commentary reading is hilarious. On Song of Songs: Catholic Study Bible: it doesn’t say they’re married Jewish Study Bible: it doesn’t say they’re married ESV Study Bible: they’re definitely, definitely married NLT Life Application: Don’t have sex until you’re married.

I'm a simple person: I see new a writing model, I test that new writing model. My initial read on Echo is that it's in a similar place to ChatGPT when I was having it write sonnets about poop in late 2022: mostly cute party tricks (LLM explainer in the style of Ecclesiastes! Emily Dickinson explains the taste of coffee!), not quite ready for primetime as a model for heavy-duty writing. It struggles in similar places that models from the frontier labs struggle: continuity of ideas from point to point, theory of mind for the reader, and judgment about what context the reader needs in order to be able to track the progression of an argument. There are flashes I quite like - it called my Career Coach folder a "special snowflake" in my setup, which is language I genuinely might have reached for myself. And I may lift the last sentence from this example for the draft I'm pulling together. But the moves from idea to idea are abrupt. It feels like it's dutifully checking off points on the outline I gave it rather than shaping those points into a cohesive whole. And I can't shake the sense that it's mimicking me rather than embodying my perspective - which may, to be fair, just be the uncanny bargain we make when we let models draft on our behalf. I'm *thrilled* to see new labs putting effort into training models for writing. I think they have a good shot at solving specific challenges models face with this specific kind of work that the frontier labs can't solve when they need their models to be all things for all people. Looking forward to tracking the progress of Echo, Deft, and whichever other models come along for this purpose. If you've got one you want me to check out, you know where to find me!