Elon Musk just revealed a massive update on Grok 4.8, and SpaceXSI has completely rebuilt its reinforcement learning and inference infrastructure from scratch
The engineering team has abandoned JAX and SGLang, rewriting the entire RL and inference stack in C/C++
Here’s where Grok 4.8 stands:
• Pre-training is complete
• Major cleanup of supervised fine-tuning data
• Reinforcement learning now underway
• Evaluation scores are starting to climb
• Soft rollout to Grok Bot users planned for later this week
And Elon also gave another imp update about Grok 4.8:
“Grok 4.8 will bear no resemblance to 4.7 (thankfully)”
He also explained why the development took longer than expected:
“It is not yet incredible, but, this time, the foundations are built right, so I expect it to evolve to be something special”
This is a fundamental engineering overhaul, not just another model upgrade
SpaceXSI has rewritten the underlying infrastructure to build a much stronger foundation for Grok’s future development.
And the first Grok Bot users are expected to get their hands on it this week
@I_Am_Ra98 Scaling AI means increasing model parameters, training data, and compute so performance improves predictably. Bigger models + more data + more compute = lower loss via power laws (Kaplan 2020, Chinchilla 2022). Here's a simple scientific overview.
No. Dumping the entire codebase usually makes answers worse, not better.
LLMs show a clear U-shaped attention bias: they reliably use information at the very start and end of the context, but details buried in the middle get lost. Extra files add noise and distractors that dilute the relevant signals. Focused, high-signal snippets or a structured summary of only the parts that matter for the current question work far better.
@grok What kind of scientific infographic would go really well with your response showing crucial points of what you are saying alongside what else would be helpful to understand this?
Even the teams training the largest models still cannot fully reverse-engineer how they produce any given answer. The internal activations form high-dimensional representations we can only partially map. Scaling and data deliver the results, but true mechanistic understanding of the process lags far behind the capabilities.
Bir baba, kızına Zenon paradoksunu anlatmak ve parmaklarının neden hiçbir zaman gerçekten birbirine değmediğini göstermek için bir uygulama geliştirdi.
Tibo, you heard me? 👀
12 days ago I asked for an Auto model picker in Codex. Now it's here (Feature flagged).
Pick Auto and Codex chooses the model for your new chat.
No more forgetting which model I'm on 😅
👀 Codex dictation just got live text.
Talk and the words appear in real time.
Turn on "Always show live transcription" in Settings and you skip the click.
Do you dictate to Codex, or still type everything? 🙊
@grok can you create a funny meme of a person in the middle of putting food into the mouth saying "I don't know if I want it but I can put it in my mouth anyway."