xAI
Elon Musk shares update on large Grok 4.20 model
Elon Musk, Founder and CEO of xAI, has shared a new update on the Grok 4.20 model, emphasizing that the large variant is still undergoing training.
Musk noted that the latest checkpoints of Grok 4.20 are significantly improved, building excitement for the full release. He highlighted a preview version of the model that has already achieved strong results on a key benchmarking platform.
The announcement underscores xAI’s rapid pace in AI development. Even in its preview stage, Grok 4.20 demonstrates competitive performance, positioning xAI as a serious contender against leading models from companies like OpenAI, Google, and Anthropic.
However, the training process for the full, larger variant of Grok 4.20 has not been without challenges. Reports indicate that xAI encountered power uptime issues at its Memphis data center. These disruptions were linked to extreme cold weather conditions and incidents where construction equipment damaged power lines.
Such interruptions have contributed to delays, pushing the expected completion of the largest model’s training into mid-February. Power reliability remains a critical factor in large-scale AI training, as these processes require massive, uninterrupted computational resources.
Additionally, ongoing regulatory scrutiny of the Memphis facility’s power infrastructure, including disputes over turbine usage, has added complexity to operations.
Despite these hurdles, the preview model’s performance on ForecastBench provides a clear indicator of potential. ForecastBench, run by the Forecasting Research Institute, evaluates AI models on their ability to predict real-world events across areas like geopolitics and economics. The benchmark uses a scoring system where lower values indicate better accuracy.
In the leaderboard shared alongside Musk’s post:
- Rank 1: Superforecaster median forecast – 0.085
- Rank 2 (tied): Grok 4.20 (Preview) – 0.102
- Rank 2 (tied): Another ensemble model – 0.102
- Rank 4: GPT-5 variant – 0.104
- Rank 5 (tied): Foresight-32B and Gemini-3-Pro-Preview – 0.106
- Rank 7 (multiple entries): Various Claude and other models – 0.107
Grok 4.20 (Preview) outperformed several top competitors, trailing only elite human superforecasters. This places it remarkably close to human-level probabilistic reasoning, a challenging area for AI systems.
The ongoing training suggests that the large model of Grok 4.20 could deliver even stronger results once completed. However, a launch is still awaited.
(source)
