Parameter Update: 2025-42

OpenAI restructuring feels like kind of a big deal?

Cursor 2.0

While it seemed like the roof has been closing in on Cursor for a while now (with Codex/Claude Code on one end and Lovable/Replit/... on the other end), this week it feels like they started pushing back properly. With Cursor 2.0, they announced their own "Composer" model, which supposedly matches other SOTA coding models while being much faster (judgement's still out as I'm waiting for actual benchmarks), browser integration (which will make frontend work much nicer), voice input and other UI changes making Agent Mode much more front-and-centre. Not sure if this will help much - this week also marked the departure of one of the cofounders - but it's good to see they're still kicking.

Claude for Excel

Not sure how exactly they managed to make this compliant with Microsoft's Add-in guidelines, but Anthropic now allows you to use Claude directly inside Excel - with the added ability of having it manipulate sheets directly for you. This seems like a bit of a wider push, given the announcement also includes a set of new Skills (we covered the new format last week) and demos of MCP integrations. I'm very interested to see how well this end up working with more complex sheets, but the examples already have me plenty excited!

1X Neo Preorder Launch

We really haven't had a nice vaporware-adjacent AI product launch in a little bit, and this one at least seems pretty cool: 1X has announced that their first robot, Neo, is now available for preorder at either $500/month or $20,000 one-time. They plan on shipping in 2026, with the important caveat being that very few features will actually work autonomously at that point - for most things, you'll get to schedule an appointment where an employee will do the work through teleoperation. Depending on your perspective, this is either the ideal way to get this tech into hands as soon as possible (and collect valuable training data) or the stupidest way to get around immigration/outsourcing restrictions we’ve seen yet. Either way - good memes.

OpenAI

Recapitalization

After months of turmoil, OpenAI has completed their “restructuring“ into a separate non-profit and PBC (public benefit corporation, the same thing Anthropic already is). At first glance, this seems unproblematic (the non-profit remains in control of the PBC), but it also involves some slightly sketchy ownership terms, including a restructuring of the Microsoft deal and a removal of the prior "profit cap". For now, the latter is unlikely to matter much - OpenAI is still burning cash left and right - but moving forward this will make it easier for them to raise money (and conquer the world as they seem willing to do). Apart from the facts around the deal, the main thing that stood out to me here are the weird vibes surrounding the whole thing, with OpenAI team members tweeting as if the company just got out of a toxic relationship.

OpenAI vs Musk Depositions

Very much a developing story, but just today the Ilya's deposition transcripts were unsealed with only minor redactions, revealing loads of details about the internal history of OpenAI. So far, things are looking very juicy if you're into big-lab gossip - Altman's brief firing in 2023, a brief exploration of a merger with Anthropic (where Dario Amodei would have ended up CEO of both) and much more. Good read, looking forward to the inevitable social network-esque netflix adaptation in a couple of years.

gpt-oss-safeguard

Not too much talk on my timeline about this one, but OpenAI has released what might be the first reasoning model specifically tuned for safety/moderation. As jailbreaks get more sophisticated and false positives (subjetively) more common, this might be a powerful tool, but not sure how this one specifically stacks up.