11 ms·
I imagine the steps to get here were: 1) You're an ambitious meta employee trying to get promoted 2) You know leadership is hyped about AI 3) All of the low
by Zenzero 2y ago
I imagine the steps to get here were:
1) You're an ambitious meta employee trying to get promoted
2) You know leadership is hyped about AI
3) All of the low hanging fruit is gone
- threeseed 2y agoThe level of resources to deliver a feature like this is staggering. There is no reality in which Musk and the SLT are not personally aware of this. And these features are all just about getting regular feedback to improve the model for future use cases.
- StableAlkyne 2y ago> The level of resources to deliver a feature like this is staggering. Are they, though? AzureML will straight up let you spin up a chatbot in a few clicks, and you could just have the ops team set up auto scaling so you only consume resources you need (because really, how many people are actually going to use this?). It could be expensive if you find tune, but a free POC would just be a to prompt an LLM to act like a celebrity, maybe with an example of text they wrote. So at worst, you're paying for compute that is a drop in the bucket to someone like Meta
- brg 2y agoI think this being based on Llama makes it tenable. GPT4 and Gemini are monolithic multipod models. Multiple instances of Llama run per machines.
- Zenzero 2y agoHow is it staggering? Meta probably threw together llama with a few hundred react components and then shipped it.