Parameter Update: 2025-43

Pretty slow week all around, so some time to catch up on the OpenAI deposition stuff (which is actually pretty interesting).

OpenAI

AWS Deal

In case anyone is keeping track of the magic money circle - OpenAI has pledged to buy $38 Billion of compute from AWS over the next 7 years. Interestingly, the announcement talks about "tens of millions of CPUs" and only "hundreds of thousands" of GPUs. As a sidenote: AWS had a total of US$107.6 billion in revenue last year, so this deal seems like it will contribute a meaningful percentage of that moving forward.

Interrupting Reasoning

A small quality-of-life improvement I expect to see a lot of other companies copy: OpenAI now allows you to add context to reasoning queries in ChatGPT without losing progress. As someone how regularly resets prompts to add context, this is a cool UX that I didn't see coming. Neat!

Emails

Already briefly covered this last week, but as there is a load of stuff in there, here are some more highlights from the recent discovery/deposition as part of the Elon Musk lawsuit:

• Ilya compiled a ~52-page memo alleging a "pattern of lying and manipulation" by Sam Altman (pitting execs against each other, conflicting info to shape outcomes) ahead of Altman's Nov 2023 removal from the board.
• Ilya on why the doc was compiled without Altman's involvement : "Had he become aware of these discussions, he would just find a way to make them disappear"
• Most of the screenshots in the doc were compiled by then-CEO Mira Murati.
• We also got a bunch of rather funny internal communication:

Google

Apple Intelligence Deal

According to rumours flying around right now, Apple is about to close a $1B/year deal with Google to use Gemini for the revamped Siri. This comes after months of Apple struggling to get anything off the ground and echos their existing agreement in Search, which has Google pay $20B/year to remain the default search on iPhone. While this feels like admitting defeat, Apple is historically not averse to running this type of partnership until it doesn't benefit them anymore - and they have deep pockets to keep it going for a while.

Nano Banana 2 Rumors

This is very speculative, but today my timeline was flush with some "leaked" Nano Banana 2 generations. I usually don't spend too much time on this type of stuff, but (1) the generated images look very good (like, step-function improvement from nano-banana) and (2) they also get flagged as originating from a Google AI model by SynthID. Also wild to see Google potentially leak an uncensored model in 2025.

Kimi K2 Thinking

Kimi.ai has been on a roll lately - but this is the one I was waiting for. When the base Kimi K2 dropped, it was already pretty great for a non-reasoning model. This week, we finally got the reasoning variant, scoring SOTA in at least some benchmarks and executing loads of tool calls. This is one area where I still prefer o3 to GPT-5, as the latter tends to be a bit lazy, so let's see if Kimi can replace it for good. Either way, the model is fully open source and in their own API is priced identically to the base K2 model (that being half the input und a quarter the output price of GPT-5-thinking). I also really appreciate the wholesome response by their. competitor Z.ai - imagine if OpenAI congratulated Google on their launches instead of vagueposting and rushing launches on the same day.

Meta: Generative Ads Model

While Meta still seems about as far away from building superintelligence as anyone else, this week we got a new look at what their AI teams are working on - ads!

Meta’s Generative Ads Model (GEM): The Central Brain Accelerating Ads Recommendation AI Innovation

We’re sharing details about Meta’s Generative Ads Recommendation Model (GEM), a new foundation model that delivers increased ad performance and advertiser ROI by enhancing other ads recommendation …

Engineering at MetaHuayu Li

Their new recommendation model was trained "at LLM scale" and achieves a +5% conversion rate in Instagram ads (which actually seems like a pretty big boost). I wasn't able to figure out if this is work coming out of FAIR, their superintelligence team or another team altogether (most of the listed authors just have "Meta" listed on LinkedIn?), but I can't imagine that anyone applied to Meta with dreamed of building this - and it stings after their recent round of layoffs.