Elon Musk Grades Grok: 4.7 Delayed, 4.8/4.9 Roadmap

Elon Musk spent four days publicly grading his own AI, Grok, with the unfortunate result that its current grade is lower than previously promised. Grok 4.7 is experiencing delays and no longer holds a leading position in the market. Its current performance is roughly on par with Anthropic's Opus.

Grok 5 is anticipated to surpass all existing models in terms of performance and capabilities, establishing itself as the definitive future AI model. This ambitious goal will be achieved by building Grok 5 on an expansive architecture featuring 6 to 10 trillion parameters.

Grok has been integrated as the primary voice assistant in Tesla vehicles since the recent summer update, enhancing user interaction and control. Furthermore, Grok provides the core intelligence for Optimus, enabling its autonomous functions. Starlink also utilizes Grok to manage over 15,000 support and sales calls daily, demonstrating its wide-ranging applications across Musk's enterprises.

This extensive, large-scale training process enables these models to develop emergent reasoning and coding capabilities. Through identifying patterns within vast datasets, LLMs learn to generate coherent and contextually relevant responses.

In AI, a parameter functions as an internal connection or a 'knob' within the neural network, allowing the model to store and utilize learned patterns. When Elon Musk states that Grok 4.7 is a 2.1 trillion parameter model, he refers to 2.1 trillion such 'knobs.' Generally, more parameters enable the AI to store more patterns, leading to greater reasoning capacity and overall intelligence. However, larger models are also typically slower and more expensive to operate.

Despite the general correlation, a larger model size does not automatically guarantee superior performance. For instance, Grok 4.6, which operates on 1.5 trillion parameters, achieved a score of 66 on their intelligence index, closely trailing Claude Opus 5 with its score of 63. This indicates that Grok can achieve high performance efficiently, optimizing for both cost and speed.

This process involves rewarding correct answers and penalizing incorrect ones, thereby refining the model's intelligence. Musk's development roadmap for Grok specifically targets optimizing these three crucial aspects.

Grok 4.6 demonstrated superior efficiency by completing tasks in approximately half the steps required by Opus 5, using around a quarter of the input tokens. Specifically, Grok 4.6 took about 53 turns compared to Opus 5's 103 turns. Elon Musk asserts that Grok maintains a significant advantage in both speed and cost efficiency over its competitors.

This issue causes the model to prematurely abandon difficult tasks, even when it is capable of completing them. Furthermore, Grok 4.7 exhibits insufficient rigor in self-verification and checking its own work.

Grok 4.8 is anticipated to deliver a noticeable improvement, while Grok 4.9 is expected to achieve performance levels comparable to Astra/Claude 3.5 Sonnet class models. The ultimate goal for Grok 5 is to potentially surpass all current industry models.

Grok 4.8 utilizes 2.5 trillion parameters, representing an unequivocal improvement over the 2.1 trillion parameters used in earlier versions trained with Jax. Elon Musk acknowledged that mistakes were made during the mid-run of previous training pipelines, which have since been corrected. There has also been a significant improvement in data quality over the past few months, contributing to the enhanced performance of newer Grok models.

Grok 4.8, a 2.5 trillion parameter model developed with a new C++ software stack, is scheduled to complete its training phase this week. Following the completion of training, the model will proceed to the reinforcement learning stage, a critical step in refining its capabilities. This marks a significant milestone in the development of Grok, as it moves closer to deployment.

Elon Musk is concurrently training a 10 trillion parameter model aimed at achieving Artificial General Intelligence (AGI) while also endorsing an essay that advocates for a slowdown in AI development. These seemingly contradictory actions are not in conflict, as Musk views peer review as a necessary price of admission for safe AGI deployment. This dual approach aims to mitigate potential government oversight by proactively managing development. The rapid pace of AI advancements and related headlines are also impacting stock markets, emphasizing the broader economic implications.

This exclusive event is invite-only, with RSVPs due soon for attendees. A graphic released by Tesla, showing a cold gas thruster package from SpaceX, hints at advanced features for the new Roadster.

Elon Musk confirmed that Starship Flight 14 will not attempt a tower catch, with the booster instead splashing down in the Gulf of Mexico and the ship in the Indian Ocean. The first tower catch attempt has been rescheduled for Flight 15, with Musk estimating the odds of success at 50% to 60%. Flight 14 is designated as a true orbital attempt, carrying Starlink V3 satellites.

The delay of Grok 4.7 is attributed to engineering challenges rather than a lack of clear explanation from Elon Musk. Currently, Grok is one release behind competitors such as OpenAI and Anthropic. A critical need for Tesla is improved multimodal performance (image/video capabilities) within Grok, as this is currently its weakest aspect. The roadmap indicates that Grok 5 is expected to utilize 6-10 trillion parameters, which could significantly enhance its capabilities.

The quality of Grok has become a primary factor in Tesla's product development questions, directly influencing the company's strategic direction. If Grok 4.9 achieves Astra and Fable class performance, Tesla could equip every car with a frontier assistant at SpaceX AI's cost structure, providing a significant competitive advantage. Conversely, if Grok stalls at the Opus 5 class, it would still be a very good assistant but a step behind the leading models. AI revenue is projected to potentially surpass other SpaceX revenue streams, underscoring the strategic importance of Grok. The focus remains on the development roadmap rather than speculative AGI timelines.