At the Barron's Investment Conference, Elon Musk provided a new timeline for AGI: Grok 5, set to be released in Q1 of next year, has "about a 10% chance" of achieving artificial general intelligence, which is his first public quantification of AGI probability.
Model Size and Data
- Parameters: 6 trillion MoE architecture, sparsity of 70%, memory usage ≈ 1.8 times that of GPT-4o
- Modalities: Text, images, audio, and real-time video streams are uniformly encoded, with frame-level latency < 120 ms
- Data: 500 million public posts per day on the X platform + 200 million hours of video streams, which can be fed back for training in real time according to service terms
- Training Cluster: 100,000 H100s (Memphis supercomputer), peak performance of 750 PFLOPs, expected to complete pre-training by February 2026
AGI Determination Criteria
Musk defines "real-time video understanding = AGI ticket": the model must reason while watching, completing visual tasks that humans can do (such as live car repairs or interpreting street scenes). Grok 5 passed 68% of 300 visual reasoning tasks in internal testing, approaching the human baseline of 75%.
Visual Ability Demonstration
A 30-second drone footage was played on-site, and Grok 5 provided real-time output: ① identifying 47 object categories; ② inferring traffic flow; ③ providing the shortest driving route; ④ warning about construction hazards ahead — throughout the entire process without additional fine-tuning.
Risks and Regulation
Musk reiterated the countdown to AGI by 2030, believing that a "disarm all humanity" style peaceful approach must be conducted within a globally verifiable framework; xAI has reserved 10% of computing power for alignment and safety research.
