Overview
- Elon Musk plans to use Grok3.5, which he may rename Grok4, to rewrite the entire corpus of human knowledge by correcting errors and adding missing information.
- He argues that existing foundation models are trained on 'garbage' data and carry 'woke biases' that skew their outputs.
- Grok3, released in February, uses synthetic datasets and reinforcement training to improve reasoning and factual accuracy while reducing hallucinations.
- xAI's Colossus supercomputer, built in under nine months, delivered over 100,000 hours of Nvidia GPU processing for Grok3's development.
- Ahead of a planned end-of-2025 update, Musk has asked X users to submit factually true but politically incorrect 'divisive facts' to refine Grok's training dataset.