Elon Musk says Grok 4.8 is finally in reinforcement learning — and getting there meant tearing xAI's whole training stack down to the studs. In a post on X on November 10, Musk said xAI has “dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++.” Pre-training is done, the evals are “starting to climb,” and a soft rollout to Grok Bot users is coming later this week. It is the clearest sign yet that xAI is done borrowing other people's foundations — and that Musk thinks the model it sells today needs replacing.
What Musk Announced
Musk’s post was short and blunt. “Training (aka pre-training) is done. Lot of cleanup was needed on the SFT data to get that mostly right. Now, 4.8 is in RL and the evals are starting to climb. We will do a soft rollout to Bot users later this week and see how it goes.” The timing matters: on September 14 he had said Grok 4.8 would finish training that week and start RL — 57 days ago. The delay, by his own account, was the rewrite. Musk gave no benchmark scores and no price, and said nothing about when API customers get 4.8. What he did say is striking: “Grok 4.8 will bear no resemblance to 4.7 (thankfully).” That is a remarkable thing for a CEO to say about the flagship model his company sells today —
Grok 4.7 launched just 50 days earlier, on September 21.
Why xAI Ditched JAX
xAI built Grok on outside software from the start — its 2023 launch post said the stack was based on Kubernetes, Rust and JAX. JAX is Google’s Python library for accelerator computing, and SGLang is the open-source inference framework from LMSYS. Musk now says both are gone, replaced by xAI’s own C and C++ code across pre-training, RL and serving. He even added that the new stack is derived from work by the engineers behind Starlink — software built to move vast amounts of data through space, now repurposed for training superintelligence. In September he had already called Grok 4.8 “a 2.5T model trained with our new C++ software stack.” Read together, the two posts describe a company that now owns every layer of its
model-building pipeline — and paid for it with “some extra time.”
The Opus 5.5 Verdict — and Grok 5
Asked how Grok 4.8 compares to Anthropic’s Opus 5.5, Musk was unusually candid: “Not as good. Opus 5.5 is a masterpiece.” But he added the real headline: “Grok 5 should exceed Opus 5.5 and maybe that happens in 2 or 3 months, with continued acceleration thereafter.” That puts Grok 4.8 in an odd spot — a stepping-stone model whose own boss calls it “not yet incredible,” with the real prize one generation away. The context is a brutal week for rivals: Anthropic’s
Claude Fable 5.2 landed November 9 at $8/$40 per million tokens, and Google’s Gemini 4 Carbon appeared in its cloud catalog on November 10. Musk’s bet is that owning the stack, not just renting more GPUs, is what closes the gap.
Why It Matters
For Bot users, the payoff could come within days — a soft rollout later this week, with no detail yet on how wide it goes. For the industry, the story is bigger: the AI race is no longer just about who has the most chips. xAI sits inside SpaceX now, with colossal compute behind it, and Musk is saying the bottleneck was the software layer all along. If the C++ rewrite really makes training and serving faster and cheaper, 4.8’s evals should climb fast — and Grok 5’s two-to-three-month window suddenly looks plausible. The honest caveat: we have no scores yet, only Musk’s word that evals are climbing. Watch the rollout this week. If Bot users feel a real jump over 4.7, the rebuild was worth the wait.
Comments
Post a Comment