Parameter Update: 2025-13

Between the long list of stuff I covered last week and the vagueposting by Altman promising loads of content for next week, today's update almost feels tiny by comparison. That being said, allow me to prefix the news updates for a second with two random content recommendations:

1. Anyone who liked the interpretability work Anthropic shared over last month - listen to Shailee Jain (UC San Francisco) explain how these ideas might help us understand the human brain - cool stuff!
2. Ever since my Bachelor's Thesis, I've had a soft spot for graphs and intuitive explanations for model behaviors. As such, I also really enjoyed Petar Veličković's (DeepMind) talk on interpreting shortcomings of LLMs through the perspective of GNNs.

OpenAI

Memory in ChatGPT

After the very mid launch of Memory in ChatGPT a few months ago (which just had the model explicitly remember facts about you - something I immediately turned off), OpenAI has more-or-less openly talked about wanting to improve it. This week, we finally got the new and improved version, which is less transparent about its implementation but supposedly much better as it can reference all your old chats.

As one might have expected, this immediately led to the next viral twitter moment as everyone rushed to get themselves roasted based on their past interactions (sidenote: I really like "5 hour long LLM tutorial" Karpathy more than "vibe coding engagement bait" Karpathy). I would have liked to try doing that myself, but as it stands, access to the new memory version is limited to Pro users and also not live yet in the EU (sad!).

Later in the week, as people got over the excitement, some people bothered to actually check how well the tool remembered details from past interactions and were left slightly disappointed, so we'll see if the current hype cycle dies before I even got to try it.

On a strategic level, I like the idea that this is really just a precursor to "Log in with OpenAI", allowing us to bring our personal context with us to other apps. I see you, Sam!

... in other news

During a recent TED talk, Altman reiterated the "500 million weekly active users" number they already shared a few weeks ago - which wouldn't have been noteworthy if moderator Chris Anderson hadn't used it as a jumping off point to ask about usage "doubling in just a few weeks" - a tidbit Altman apparently shared backstage just before the talk.

The vagueposting continued into this week, with Altman teasing the release of o3, o4-mini and what might be called GPT-4.1.

Google

Firebase Studio

Announced during Google Cloud Next, Firebase Studio seems like wholesale rebrand of the entire thing into a crossover of Lovable/Replit/Cursor/...

Unfortunately, this announcement included sunsetting standalone Project IDX, which I always liked to use in a pinch. I've not yet tried how much better/worse Firebase Studio is, so let's see!

Ironwood TPU

In what might be the most significant announcement of the bunch, Google also announced their newest generation TPU, named Ironwood.

While I'm awaiting more specs (which we may or may not get), this feels like a good place to reiterate that this is Google's primary advantage in the AI arena right now. They are the only major competitor that doesn't rely on Nvidia as part of it's stack, and this announcement underscores to me that they realize just how important this is.

Agent2Agent

After publicly embracing MCP just last week, this week also saw the announcement of their own (but open) Agent2Agent protocol, meant to be an orthogonal addition. I'm interested to see how much adoption this garners, as the target audience does feel a bit different (also, lol).

Amazon

Flying somewhat between {my|the} radar, Amazon has also been cooking some seemingly pretty cool models.

The first one is Nova Sonic, a Voice-to-Voice model for immediate speech interactions. Also supports function calling.

The second one, Nova Act, is an action model for controlling web browsers. The announcement already happened two weeks ago, but I am only coming around to taking a look at it now and honestly, am mostly including it here as a throwback to them hiring away most of Adept around a year ago.

Other stuff

Meta

If you wondered how far the Llama4 rabbit hole goes - ex-Meta employees are (supposedly!) proudly claiming not to have worked on Llama4 and the model has now dropped below DeepSeek 2.5 (!) in LMArena.

DeepSeek

After their bonkers Open Source Week last month, DeepSeek is now working on merging portions of their infra stack back into the Open Source projects they originate from. Honestly, still no idea why they would be doing this (other than actual Altruism, which seems unlikely in the current environment?) but I'll gladly call out good things when I see them, so hats off!

Figure

Very much a developing story, but it seems that Figure CEO Brett Adcock may have been seriously exaggerating the current state of their partnership with BMW. Quoting an internal Slack message here: "Always doubt a humanoid demo when the clips of operation are <1min" - let's see how this turns out!