Parameter Update: 2025-23

It's been another slow news week, so I'll not be too mad if you skip this one. Some fun stuff in there though.

Midjourney: Video V1

I will admit that, since the release of gpt-image-1, I have not gone back and tried Midjourney again. Now, it seems that they were also feeling some of that pressure, as they have turned to release their first ever video-generation model.

As explained in an accompanying blog post, their stated goal is, in the long term, to build a world model. As an intermediate step, the current video model is meant to represent a good price/performance tradeoff. Considering it starts at $10/month for ~25 video generations at slightly below 720p resolution (with the Pro tier sporting unlimited "relaxed" generation from $60/month), it feels like you are getting ~half of OpenAI's Sora generations (currently at 50/month), using a much, much better model, while losing out on the other benefits of the ChatGPT Plus subscriptions. On the other hand, Sora already doesn't stack up all that well compared to Google's Veo3 (which has audio!) or some of the recent Chinese models, so not sure if Midjourney are delivering a particularly amazing value. Nevertheless, the underlying model is undeniably cool and still has that cinematic flair no one has managed to really capture so far:

The "image reference" mode is also pretty cool:

OpenAI

Record Mode

After announcing it last week, "Record Mode" for ChatGPT Pro/Teams/Enterprise is now live. While some people are already proclaiming the death of a thousand "meeting recorder" startups, I will reserve judgement until this one hits the Plus tier, as it apparently integrates with Canvas, which, in my experience, is usually a bit shit. I also really wonder when it became acceptable to just record your meetings without notifying other participants, which this not only enables but almost encourages? I supposed Granola crushed that barrier a few months ago, but still, not sure if this is an awesome development.

Podcast

Just when I thought Altman's demeanor couldn't possibly get more self-indulgent than the "io" announcement video with Jony Ive (incidentally posted the same week as Google I/O; currently down for legal reasons), we got the first episode of the OpenAI podcast, featuring the man himself speculating on solving high-energy physics by deploying ever larger GPU clusters while reaffirming they will "probably" not do ads in ChatGPT unless they are "well done" - yikes.

Google: Search Live

In yet another installment of "how many AI products can Google launch", Gemini Live (in the Gemini App) is now facing competition from Google Search Live, which is basically the same thing except oriented around search results instead of your environment and built into the Google App instead of the Gemini app. No word on which of the bazillion Gemini 2.5 variants this one is powered by for now, though (lol).

Cluely: Raising $15 Million Series A

Weeks after doing a full art deco / American retro-futurism rebrand, where they encouraged founders to chase big ambitions, Andreessen Horowitz has this week decided to lead Series A of Cluely, a shitpost-account turned meeting recorder turned VC-backed social media Hyper House. Some of the stunts they have pulled recently include:

• The founder got thrown out of Columbia for building a cheating tool,
• ...then used it to cheat himself into a few job offers,
• ...which he publically rejected and made fun of.
• After, he raised a $5.3 Million pre-seed,
• ...and spend most of it on "marketing interns", whose main output was a
• ... sexist launch video that had the founder awkwardly stumble his way through a date,
• ... another video featuring some strippers, and
• ... a massive YC AI School afterparty, which got raided by the police and had the founder briefly arrested

And yet, the hardest thing to believe is that they are doing all that for a simple meeting recorder tool?