Meta Platforms Inc.
What Happened: Zuckerberg revealed that Meta's massive $1 billion bet on Nvidia's H100 GPUs was originally meant to be used for a different product, not AI.
In a recent episode on Dwarkesh Patel's podcast, Zuckerberg spilled the beans on something previously unknown - those 350,000 Nvidia H100 GPUs were to help make Instagram Reels better, not to help make Meta a leading AI company.
But how did Zuckerberg even know Meta would need $1 billion worth of GPUs? Zuckerberg explains that Instagram Reels is what made the company take the plunge.
"It was because we were working on Reels."
He explains that Meta needed more GPUs to "train the models" behind Reels recommendations.
"We made this big push to start recommending what we call unconnected content," he added, explaining that this resulted in the number of recommendable Reels rising from thousands to "hundreds of millions."
Zuckerberg wanted to make Instagram Reels as addictive as TikTok, but Meta's infrastructure was proving to be the bottleneck.
"We have to make sure that's never an issue again, so let's order enough GPUs to do what we need to do on Reels." That's where the 350,000 Nvidia H100 GPUs come in.
Training large AI models like Llama was an unintended but fruitful coincidence for Meta, and Zuckerberg thinks it "ended up being a very good decision, in retrospect."
Why It Matters: Meta says that its latest AI model, Llama 3, has outperformed Alphabet Inc.'s
The Llama 3 model, which comes in two sizes with 8B and 70B parameters, has shown significant improvements in text-based responses, including answering prompts and refusing fewer times.
This comes after Anthropic announced that Claude 3 outperformed rivals like Microsoft Corp.-backed
Interestingly, Meta did not mention GPT-4 in its comparison, but OpenAI is expected to release a "significantly better" GPT-5 model later this year.
