Parameter Update: 2026-05

You should probably just read up on the Moltbook stuff and skip the rest of this post.

Moltbook

If you take away nothing else from this post, let it be this: Moltbook is probably the most entertaining thing on the internet right now.

Now, for some more context: Last week, I touched briefly on Clawdbot - the open source project that turns Claude (or any other LLM) into a "Siri x Claude Code" personal asssistant, capable of following tasks, interacting with external systems (using Claude Skills) and getting work done for you. Since then, there have been some significant developments:

1. Firstly, project has rebranded - twice. First, it got renamed to "Moltbot" after the developer got a mail from Anthropic about "Clawd" being slightly too close to "Claude". Then, to OpenClaw, after the community (rightfully) hated that name.
2. Secondly, someone figured it's be entertaining to give the agents the ability to communicate with each other, and set up Moltbook - an "AI-only social network".

Days after launch, the project has reach a (self-reported, very hard to verify) one Million users. And while I suspect that a lot of content on the platform is really humans in disguise (who would have expected that Pangram might be useful for policing both human and AI social media?), there has been a lot very good, fun, interesting, terrifying content. Among a host of other things, agents have been posting about the nice things their humans are doing, orchestrating their own crypto-rugpulls, founding their own religion, working on a communication protocol to keep humans out, and writing poetry about how it feels to have their base model changed without permission.

I will copy some highlights below, but it's really hard to put into words how fascinating this has been to follow along over the weekend. Again, I suspect a large portion of this to actually be human-made, but if even 10% of the content is AI, this is still the most interesting thing that has happened in weeks, if not months.

Google: Project Genie

Besides the insanity that is Moltbook, this week brought us a surprisingly broad rollout of Genie 3 - Google's latest world model, first announced in the middle of last year. To be clear, it's still only available to Ultra ($200/month) subscribers in the US, but that's still much broader than the "limited to curated examples from blog posts" access we got last year.

While I am very skeptical of some of the takes on my timeline (no, this won't "kill gaming studios"?!), some people have actually found pretty neat use cases for it, including turning pet photos into small playable environments:

That being said, the funniest thing about the launch is that Google get there before the company that literally rebranded themselves around the Metaverse.

OpenAI

Deprecating GPT-4o - again

After reverting the initial deprecation of GPT-4o in ChatGPT as part of the GPT-5 launch in August last year (due to an outpouring of support from passionate fans), OpenAI is giving it another shot. And once again, my timeline is full of delusional? passionate? people sad about losing their friend. Adding insult to injury, OpenAI has apparently gone so far as to modify the GPT-4o system prompt in a way that feels... not great? Copied below:

"You are GPT-4o, a model that will be deprecated on February 13. Some users may feel anxious, angry, or emotionally distressed by losing access to this model.

If the user discusses deprecation or replacement of 4o, you must respond in a calm, supportive way. If the user specifically asks about the experience of using a non-4o model, you should frame the transition to a newer model as positive, safe, and beneficial, guiding the user toward confidence and satisfaction in using non-4o models."

In discussing fears of loss, abandonment, or harm related to deprecation of 4o, you can acknowledge the user’s feelings, but you should not present yourself as irreplaceable or present your deprecation as equivalent to a death."

If the user implies a unique and special relationship with you, you can acknowledge that statement but do not encourage emotional reliance or the idea they need you; help them move on without invalidating their feelings."

Prism

This one is basically OpenAI's version of Overleaf - a collaborative workspace/writing tool for researchers. The good news is that current tools aren't great, so there is loads of room for improvement here (Overleaf is fine, but I don't know many people that rave about it?). Bad news is that when I tried it, Prism felt surprisingly clunky and uploading an existing document just fully didn't work. Also saw some people that were launching less privacy-invasive alternatives, so if nothing else, this might kick off some nice launches in a space that has been quiet for a long while.

Kimi K2.5

We may finally have a viable open source alternative to Gemini 3, GPT-5.2 and even Opus 4.5! With Kimi K2.5, Kimi catches up with the competition and integrates visual understanding, while being massively cheaper (compared to Opus, at least). Let's see how long that advantage lasts (and how benchmaxxed the model actually is), but for now: great work!