Meta Back at the AI Frontier • Sora as a Startup

Continuing their "we're back" tour, Meta's new 'Wonder' models, built for Muse, put them at the frontier – or so they say. Meanwhile, not back yet, but maybe soonish: Sora?! Say it is so! Read on for thoughts on...

Meta Marches Back to the AI Frontier

Well it took about a year and a half and billions upon billions spent, but Meta may finally be back to the AI frontier. Here's Ina Fried:

Meta on Monday unveiled Muse Wonder and Muse Wonder Full, models it says deliver state-of-the-art performance on a range of tasks involving personal and professional agents, memory and coding.

The company posted a range of benchmarks showing the models, previously known by the code-name Watermelon, topping or matching the latest offerings from OpenAI and Anthropic.

First and foremost, while I didn't like the 'Muse Spark' name as it sounded too much like some sort of generic enterprise product (though Google clearly disagrees), 'Wonder' and 'Wonder Full' are great, fun, pun-y names. They also sound decidedly optimistic, which is probably needed these days with regard to AI.

Second, while you always have to take companies releasing their own benchmarks with grains of salt, you especially do with Meta, which famously got caught "benchmaxxing" their Llama models, which led to (or at least pointed to) the entire reboot of AI within the company. While there's no reason to believe Meta is doing any funny business here – and you have to believe they wouldn't be stupid enough to do that twice – well, sometimes things perform differently in public.

To that end, it might be reasonable to assume these models won't be state of the art at everything – as some are already pointing out with regard to 3D video rendering (and image generation). But instead, Meta focused on a few key areas, namely the all-important coding capabilities (though some questions seem to persist there too), but also:

Meta is emphasizing two capabilities — "Active Memory" and "Active Focus" are designed to help AI systems perform more useful work over extended periods.

Active Memory allows the AI to retain important information about a user across conversations and retrieve it when relevant to a new task.

Active Focus is designed to keep an AI agent working toward a goal, including trying different approaches and adapting to new information when it encounters obstacles.

Those are interesting areas of focus because it points to Meta tailoring Wonder specifically for agentic use cases. No surprise there given their focus on Muse, the AI personal assistant product. To date, the other players have seemingly wrapped their personal assistant harnesses around models that were more generally tailored for AI tasks. This could give Meta an even bigger advantage in the space, obviously.

That said, the Wonder models are currently being used to upsell to Meta's paid AI offerings. That's fine to start – and it worked on me, I'm actually paying Meta directly for the first time, business diversification, at last! – but they're clearly going to want to leverage Wonder to win what will be a full-on race around AI assistants.

Also note:

Meta isn't saying whether it will open source any of the Wonder family of models, but Wang noted that Meta recently open sourced the weights for Muse Glimmer as well as the technology needed for folks to put their Muse agent onto hardware.

"We're still committed to open source," Wang said.

Yeah, that's a "no" for Wonder, at least for the foreseeable future. No surprise there, but Meta continues to try to do this slightly disingenuous "open" dance. The reality is that all the major players release some flavors of open weight models now – well, aside from Anthropic. Meta releasing their non-frontier models' weights in pretty par for the course. But the old Llama strategy was to release the weights for all of their models. That's clearly not the case here at this new frontier.

Anyway, kudos to Meta for seemingly getting back to the bleeding edge of AI. I didn't think they could simply buy their place back in the front of the line, but I was wrong. I mean, to be clear and fair, obviously it was far harder than that and the speed at which they've executed is impressive.

Still, it's one thing to get to the frontier, and another to stay there – just ask Google. Speaking of, Meta's moves with Wonder obviously make Gemini look even worse for failing to stay at that frontier. But there are already whispers that a comeback – another one – may be in order at some point soon. Argon?

For now, enjoy that pint, Zuck! 🍻

Sora as a Startup

If this AI Rumpelstiltskin meme currently infecting social media has taught me anything, it's that I miss Sora. OpenAI's video creation service was just plain stupid fun. And it's wild that it had to be shut down not in the name of copyright, but so OpenAI could focus on their business in the face of the Anthropic threat. Well, good news, a comeback may be in order. As Stephanie Palazzolo and Cory Weinberg report:

Remember OpenAI’s video model Sora and all the great Sam Altman memes that came from it? Earlier this year, the ChatGPT-maker canned both the model and the TikTok-like app it powered as part of broader culling of server-intensive “side projects” even as rivals xAI and Google forged ahead with their own video models.

Well, there’s a chance that a version of Sora is coming back from the dead.

A stealthy startup developing AI video for filmmakers, which we revealed last month, is in talks with OpenAI to access Sora, tweak it and sell software or other services using it, according to a person with knowledge of the talks. Hollywood mogul Jeffrey Katzenberg, former Sora research lead Bill Peebles, and former Dropbox CFO Sujay Jaswa co-founded the startup in recent months. (We don’t know the startup’s name).

Yes, insert the Sora-meets-Quibi jokes here. Better yet, insert them into Sora and make some video memes about it, hopefully soon.

One of the stranger elements of the abrupt Sora shutdown was that OpenAI had signed a deal with Disney around using some of their IP within the network – also tied to a would-have-been equity investment by Disney. Katzenberg, of course, was a long-time Disney executive, so might this new partnership rekindle such flames? Or between this and the Epic fiasco, is Josh D'Amaro burnt out on the tech deals?

The more interesting question in all of this is if it could actually work. How many times does the resurrection of a once popular product actually take hold? Not often. Everyone is still waiting for that Vine comeback. Perhaps what they're hoping for here (if this actually happens) is a Musical.ly-to-TikTok transition. ByteDance famously bought that startup (which was only mildly popular to begin with) and from its ashes arose the juggernaut we know today.

Of course, it helped to have ByteDance's other businesses both to build the all-important algorithm as well as feed the ad spend on Meta's properties that helped TikTok get big, fast. Still, it worked because the product was good. And Sora's product was good!¹ So maybe there's a chance...

¹ Though as the article notes, it's not clear if this new company would get access to Sora's weights, which would obviously be key in replicating the functionality of the product. But granting such rights could help OpenAI recoup some of the massive investment in Sora thanks to an equity stake in this new company. ↩