Parameter Update: 2025-12
"ghibli" edition
After taking the last two weeks off (as I was busy moving back to Germany), I am excited to be back with a roundup post this week!
OpenAI
GPT-4o Image Generation
Dominating my timeline (as, it seems, everyone elses) over the past two weeks, has been GPT-4o's native image generation. While Google technically "got there first" when they rolled out image generation in AI Studio three weeks ago, I can't deny this one feels a lot cooler.
Don't get me wrong, it's still extremely slow, fails extremely often and (even compared to Google) sucks at iterating on previous images. I also agree that there are some things that should never be "ghibli-fyed" (i.e., White House Twitter Account) but the pure joy some of the results have managed to spark is still remarkable, and it feels like, as far as creative tools go, this one has loads of depth and people are only just scratching the surface of what's possible with it.
It also seems that the announcement has somehow permeated into mainstream discourse, with Altman announcing one million new users signing up for ChatGPT within one hour last monday. Subjectively, the launch has come up more often in my personal friend group than I expected, so I am excited to see where this goes. Still holding out for the API too.
... in other news
Apart from the image gen announcement, we also got MCP support in the Agents SDK (I'll do a dedicated thing about MCP at some point, as it seems to be the inevitable industry standard for tools at this point) as well as announcements of full o3 and o4-mini both coming soon.
Gemini 2.5 Pro
After we just got done digesting the recent model launches over the past weeks, it seems that Google is right back at it with Gemini 2.5 Pro. Based on benchmarks and vibes, this one seems like a winner.
Interesting to me: It seems that Google, for maybe the first time in a while, realized this and is actually pushing the thing out with quite some force (e.g., moving out of preview way faster than usual). This means that, by some definitions, Google now owns the "intelligence/dollar frontier" (cool metric!). It also seems to be working, with AI Studio active users increasing 80% month-over-month.
Reorg
There's also been a slightly drama-y reorg going on behind the scenes with the guy behind NotebookLM getting promoted. Seems sensible?
Anthropic
While Anthropic has no new models to share, they did at least give us the most interesting interp research blog post I think I've ever read. I especially like the model "fake reasoning" its way into a provided answer as well as finally getting proof of "longer horizon" planning in LLMs (even though they they're still just "next-token predictors").
Meta
Llama 4
Announced last Saturday (who does that?!), Meta has given us our first glance at the Llama 4 herd of models. While I was initially really excited about getting an Open Source 2T model (and "Behemoth" is a really cool name!), it seems that Meta may have (intentionally or inadvertently) cut some corners in Post-Training, unfortunately still seems to rely solely on DPO and the final result seems mostly cringe.
Microsoft
Copilot Updates
Once again, Microsoft has launched a series of new features for Copilot (your guess as to "which one" is as good as mine). See the thread here, but for now my vibe is this:
They also launched a fully AI-generated version of Quake 2 that, while actually very cool technically, still leaves me wondering why.
Other stuff
Midjourney V7
Alpha out now. Includes Draft Mode, which seems really cool. Also still the prettiest image model out there. That being said, good luck competing with OpenAI.
Runway Gen4
Gen4 "launched now", but some people seem to not have access yet? Last time I tried it, the whole thing felt much too expensive for the performance they delivered.
DeepSeek V3-0324
Worst name yet, but seems like a big step for the amount of attention it got?
