Meta to boost training infra for Llama 4 tenfold, maybe deliver it next year Meta has told investors generative AI won't bring it revenue this year, but that the massive investments it's planning will pay off over time – and be configured so they're not tied to training workloads. Speaking on the former Facebook's Q2 earnings call, founder Mark Zuckerberg explained Meta is "planning for the compute …
Meta's content ranking and recommendation tools already use what Meta calls "Core AI," which Zuckerberg told investors is improving user engagement by pushing more stuff users want into their Facebook and Instagram feeds.
What a shame they couldn't do something useful like fix the climate, cure Dementia...
"What a shame they couldn't do something useful like fix the climate, cure Dementia..."
Indeed. But in defence, it is open source and DeepMind have done a lot of protein folding stuff. But if Mark turned that power to good, rather than grubbing Ad coins, what a wonderful world it could be.
"amount of compute … almost 10x more than what we used to train Llama 3."
Bleeding hell! I'm sure I read somewhere recently that AI accounts for 10% or so of power use in USA. Mark got off to a great start with AI but his true colours are showing clearer now. He is still an ad-slinger and will always be.
Throwing more clocks cycles at it is not a very elegant or sustainable solution. If Llama 4 takes 10x power, what is 5,6,7, and so on going to take? It is Terminator territory because all of our resources will go into powering the AI beast. Dumb. Have an AI /beast but not a Matrix-style one, FFS.
Because there is so much money in LLMs, pretty much nothing else gets the limelight. LLMs are one of the lesser interesting parts of AI. Motion prediction is a much larger and quantifiable market with clear profits and returns. Professional sports betting as we know it will be dead in 10 years.
Mark needs to think a step beyond the couple of steps ahead of most he is. If he were to freeze Llama at 3 for 1 year and redirect into quasi real-time models, then even though the results may appear meagre, the returns will blaze past this legacy conventional computing.
Come on, Mark or a Facebook droid. I know u can do it. U did it with Meta. Now you need not the meta, but the unfathomable.
If they ever hope to make any money from this they're going to have to settle on a model that's good enough and can be trained incrementally. And is of sufficient value in the limited number of areas (where the possibility of arbitrary or halucinatory responses is acceptable) regulators will permit.
I don't think anybody operating a LLM is ever going to break even, the costs are astronomical but the results too poor to charge enough money for. The only parties making money on LLMs are the providers of shovels to the new (fools)goldrush: Nvidia, Azure etc.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
