Meta Sets the Table with Table Stakes AI
Jonathan Vanian
Good news, Wall Street. Meta isn't burning all those billions on nothing. Well, we think. It's all a bit TBD. Quite literally.
But now at least we get the first taste. 'Muse Spark' is an awfully generic name, but the early results seem promising. Of course, that was the case with Llama as well until it wasn't. We're probably going to need to see a bit more than benchmarks shared by Meta here. But really, the proof will be in the usage. As in, is anyone actually going to use these models? And not just because they're shoved into surfaces that billions of people use?
To their credit, Meta is being honest here. Muse Spark isn't really competitive with the truly frontier models from others on a number of fronts – namely, coding. Seemingly the benchmark such companies care about the most right now. Instead, Meta believes they've made a relatively svelte model that at least deserves to sit at the same table as the other labs for a number of tasks. Yes, Meta has made a table stakes AI.
That's harsh, but fair. I mean, no one seems to worried about this model ending the world. On the speed-to-launch, it is impressive – it took them nine months to birth this baby. Also table stakes for a human being, but fast when rebooting your AI lab and starting from scratch! Of course, xAI also previously got to the cutting edge in record time and... it hasn't really mattered. Well, unless the goal is to merge – first with a sub-scale social network, then with an orbital scale rocket company. That's probably not Meta's game plan here, so the results are going to have to stand on their own far more.
But perhaps not completely, because again, Meta has the unique advantage of having several of the most widely used surfaces on the internet and mobile. If nothing else, it seems like Muse Spark will help to power Facebook and Instagram recommendations – and yes, ads. And don't forget the glasses. That's the really big, future play here. If Meta can sustain the growth of their Ray-Bans, they have a shot to take on Apple. Not the iPhone, but their AI hardware projects. Google too. And, of course, OpenAI.
And no, this model isn't "open". That writing has clearly been on the wall since shortly after Meta bought Scale and kicked off this sprint. Mark Zuckerberg may have spent much of the past few years talking up "open" "open" "open" "open" "open", but well, sometimes "open" backfires, just ask Google.
Yes, yes, there's still some "open" lip service here. Future Muse Spark variations or whatever. But that too is table stakes.
Anyway, that's all down the line. Clearly, 'Spark' is just the first 'Muse' model. This was the one codenamed 'Avocado' and there's a bigger fruit apparently in the works in the form of 'Watermelon'. Hopefully that one gets a better final name.
One more thing: perhaps the most interesting element of the Muse movement is the notion that Meta intends to sell access via APIs. A first step towards a bigger Meta Cloud offering? You don't spend $140B a year for table stakes.
M.G. Siegler
M.G. Siegler
M.G. Siegler
M.G. Siegler
M.G. Siegler