Meta is axing 600 roles across its AI division

I think this is because older AI doesn't get done what LLM AI does. Older AI = normal trained models, neural networks (without transformers), support vector machines, etc. For that reason, they are letting them go. They don't see revenue coming from that. They don't see new product lines (like AI Generative image/video). AI may have this every 5 years. A break through moves the technology into an entirely new area. Then older teams have to re-train, or have a harder time.

This seems like the most likely explanation. Legacy AI out in favour of LLM focused AI. Also perhaps some cleaning out of the old guard and middle management while they're at it.

There always has been a stunning amount of inertia from the old big data/ML/"AI" guard towards actually deploying anything more sophisticated than linear regression.

FAIR is not older AI... They've been publishing a bunch on generative models.

I would expect nearly every active AI engineer who trained models in the pre-LLM era to be up to speed on the transformer-based papers and techniques. Most people don't study AI and then decide "I don't like learning" when the biggest AI breakthroughs and ridiculous pay packages all start happening.

It's a good theory on first read, but likely not what's happening here.

Many here were in LLMs.

This is not "older AI". This team built everything up to and including Llama 4.

I really doubt that. Most of the profit-generating AI in most industries... decision support, spotting connections, recommendations, filtering, etc... runs on old school techniques. They're cheaper to train, cheaper to run, and more explainable.

Last survey I saw said regression was still the most-used technique with SVM's more used than LLM's. I figured combining those types of tools with LLM tech, esp for specifying or training them, is a better investment than replacing them. There's people doing that.

Now, I could see Facebook itself thinking LLM's are the most important if they're writing all the code, tests, diagnostics, doing moderation, customer service, etc. Essentially, running the operational side of what generates revenue. They're also willing to spend a lot of money to make that good enough for their use case.

That said, their financial bets make me wonder if they're driven by imagination more than hard analyses.

FAIR is 3000 people, they do tons of different things

There's a lot of areas where you need to be able to explain the decisions that your AI models make and that's extremely hard to do unless you're using linear regression. E.g. you're a bank and your AI model for some reason appears to be accepting applications from white people and rejecting applications from african americans or latinos. How are you going to show in court that your model isn't discriminating based on race or some proxy for race?