Inside Grok 4.5: How Cursor Collaboration and Real Developer Data Shaped the Model
eyishazyer · x · 2026-07-30
The thread delves into the training details of Grok 4.5. This marks the first time xAI collaborated with Cursor, training the model on real developer-agent sessions rather than relying solely on synthetic benchmarks.
Notably, Cursor's own writeup honestly admitted that Grok 4.5's benchmark edge might partly come from an old Cursor codebase snapshot leaking into the training data. The author points out that Grok has previously suffered from a disconnect between top-tier benchmark scores and real-world user rankings (e.g., Grok 4 ranked around #66 globally), and it remains to be seen if 4.5 breaks this pattern.
More from Companies & People
- 18 reasons why AI transformations fail, even at successful companies — alex_verem · 2026-10-16
- Viral Jio origin story debunked: 'no internet' wasn't an excuse by 2010 — kingslyj · 2026-10-16
- LlamaIndex founder: post-training teams at frontier labs hire Stanford/MIT NLP PhDs at the highest rate — lateinteraction · 2026-10-16
- The 'sci-fi make it till you make it' playbook: Musk pioneered it, Altman copied it — Linahuaa · 2026-10-16
- Ex-DeepMind researcher Michal Valko surfaces with stealth AI startup on stage — misovalko · 2026-10-16
- Ant's Ling Product Lead on Why AGI Is "Inevitable" and Open-Weight Models as Product Strategy — ChinaTalk · 2026-10-16