Parameter Update: 2025-17

Even after dominating the news for the past two weeks with their GPT-4o sycophancy fiasco, it seems that OpenAI is hellbent on getting some more coverage in my timeline. Well, good for them then, that this week was a bit of a snooze otherwise.

OpenAI

Buying Windsurf

After trying to unsuccessfully purchase Cursor (which has been shopping around at a $10 Billion valuation), OpenAI has now settled for Windsurf. Now, I will admit I'm slightly confused by this move - not only did they pay an absolutely ridiculous $3 Billion for a VSCode fork instead of vibecoding their own, they also picked one absolutely no one I know actually uses (I am told this is because it's an enterprise product, but really... is it?). So while OpenAI still seems to be winning at the game overall (don't forget they bought into Cursor back in 2023), this seems like a radical move to pull right now.

While this announcement did come in line with some pretty neat new Windsurf features, it also came in the same week Cursor announced some much more exciting updates (background agent, anyone?) - as well as their new student plan, which is just "Free for a year" (definitely try it out if you haven't!).

Restructuring & Leadership Changes

While the Windsurf buyout was ongoing, it seems that the corporate restructuring attempt of turning OpenAI into a for-profit seems to have failed for now, with them instead settling for a Public Benefit Corporation.

Before you breath easy though, keep in mind this is also the week they brought Fidji Simo on board - best known for bringing ads to Instacart. Yikes.

Reinforcement Fine-Tuning

In more exciting news - the RFT API that I was so excited about late last year has finally made it's way to production, now with o4-mini instead of o1 - took them long enough!

Mistral: Medium 3

After actually impressing me with their Le Chat relaunch a few weeks ago, Mistral had pretty much disappeared from my timeline. Turns out they were cooking on their new Medium 3 model.

Unfortunately, comparing to Llama 4 does seem to tell most of the story here, as they continue to struggle to significantly outperform GPT-4o.

Google: Gemini 2.5 Pro Update

In an attempt to push their advantage with Gemini 2.5 Pro, Google has announced a special "I/O" variant (in preparation for their I/O event). This one pushes the envelope in all metrics on LMArena and is apparently SOTA at video understanding and coding specifically. Unfortunately, in my experience, it also has much worse vibes though - where's the old version gone, Google?

Stripe

Among many other things, Stripe (yes, the money company!) has announced that they trained their own transformer model at their annual (?) "Sessions" event. This one is for fraud detection, apparently increasing detection rate to 97%. That being said: