Unique Research · F2AI Focus · March 19, 2024
A week after teasing it, Elon Musk’s xAI delivered: Grok went open. Developers framed it as both a bet on open AI and a response to OpenAI’s closed, commercial turn — Musk, an OpenAI co-founder, has long argued AI should stay open.
AI + sci-fi = an optimistic Black Mirror?
Grok launched November 2023, named after Douglas Adams’s Hitchhiker’s Guide spirit — witty, rebellious, deep empathy. (The word “grok” also comes from Heinlein’s Stranger in a Strange Land.) Dual mode: rigorous accuracy via large params/architecture, and humorous/rebellious chat that may even suggest what to ask — ChatGPT-class chatbot with more attitude.
Musk’s SpaceX for AI?
- 314B parameters — among the largest open LLMs
- Mixture-of-Experts (MoE) — route inputs to specialized expert nets
- Custom train stack: JAX + Rust from scratch
- Trained on large text without task-specific fine-tune; ~25% of weights active per token
- Broader low-level access than some restricted open models — but less full transparency than Pythia / BLOOM / OLMo (no complete train code + dataset release), limiting reproducibility and bias audit
Open vs closed ≠ zero-sum
Open AI: collaboration, innovation, shared wins — TensorFlow (Google), Hugging Face Transformers. Pros: transparency, zero-cost entry; cons: security surface, uneven support.
Closed AI: commercialization, exclusivity, IP — Apple Siri; Google Search ranking. Pros: control and support; cons: lock-in and cost.
Many successful stacks mix both. Open vs closed can reinforce each other as models deepen across domains — not a zero-sum game.
2 CN images; attach on ingest.
— END —