PendingElon Musk promised massive Grok gains from the C/C++ stack rewrite by Sep 30, 2026. The deadline passed and it didn't happen.
Deadline Sep 30 2026 passed with no public receipt of 'truly massive gains' from a full C/C++ training-and-inference rewrite exact-mapped to GB300. Musk said Sep 13–14 that Grok 4.8 (unreleased) is trained on a new C++ pre-training stack; he also said the RL stack is still legacy. Fetched Sep 30: x.ai/news flagship post is still Introducing Grok 4.7 (Sep 21; 'twice as fast' vs comparables, not attributed to the C++ rewrite); newest product post is Team Bots (Sep 28). docs.x.ai/docs/models lists grok-4.7 as the model to choose with no grok-4.8 id. Internal training-stack rewrite without shipped inference/GB300 gains is not the June 29 promise. Missed. x.ai/news + docs.x.ai/docs/models fetched Sep 30 2026 — Grok 4.7 still flagship; no C++-stack massive-gains / grok-4.8 receipt ↗
In a June 29, 2026 reply on X, Musk said Grok would see 'truly massive gains' in about three months once xAI rewrites its entire training and inference stack in C/C++, deletes most intermediate software layers, and exact-maps the model to Nvidia's GB300 hardware. The claim extends his earlier 2026 statements about an in-house, bare-metal AI stack meant to outperform frameworks like JAX. Whether measurable performance gains materialize by roughly the end of September 2026 is the falsifiable test.
Pending
Pending
PendingComment on predictions and back the project. We email you a one-tap sign-in link — no password to remember.
By signing in you agree to keep it civil and lawful. We never sell your email.
Discussion