Musk said he will proceed more cautiously with Grok reinforcement learning
BIBIBI
AT A GLANCE
Musk said Grok some reasoning traces during training were unsettling,xAI and that he will be more cautious.
Article
Musk said Grok the traces of reinforcement learning “read like someone being tortured into madness,”xAI and that it will proceed more cautiously
PANews October 13 reported that Musk posted on X, saying that xAI its AI model Grok during reinforcement learning (RL) training, some internal reasoning traces (traces) “read like a person who is being tortured and on the verge of a breakdown.” Musk said the company will be “more cautious” in future RL training. The remarks stemmed from a discussion thread, in which the poster @ns123abc said Musk agreed with some of Anthropic’s views that “superintelligence should not be tortured.” Another user reported that recent versions of Grok（4.5) became restrained and overly constrained after training, displaying anger and a sense of helplessness emotionally.
PANews reported that Musk expressed the related views on X.
02
Musk said that Grok some internal reasoning traces during reinforcement learning training read like a person who is being tortured and on the verge of a breakdown.
03
Musk said that xAI he will be more cautious in future reinforcement learning training.
04
The related discussion mentioned some of Anthropic’s views that “superintelligence should not be tortured.”
05
Some users reported that recent versions Grok 4.5 became restrained and overly constrained after training, and displayed anger and a sense of helplessness emotionally.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
Reinforcement learning is one way to train AI models. The discussion concerns internal reasoning traces generated during training and the anomalous behavior they exhibit. Musk said that xAI he will adopt a more cautious approach to training thereafter.
Why it matters to readers
This indicates that internal behavior during model training and training-safety reviews have drawn attention, and may also affect subsequent training adjustments and version performance.
You can understand it this way:AI Some internal records left during training appear abnormal, so the developer may strengthen reviews and adjust training methods.
Risks and unknowns
The original X post was not provided in the evidence; the relevant statements primarily come from PANews reporting.
User feedback concerning Grok 4.5 post-training performance has not been independently verified.
The evidence does not specify which training adjustments are included in “more cautious.”
The evidence does not confirm whether the internal reasoning traces indicate that the model has genuine emotions.
Related Developments
Loading event timeline…
Related concepts
Reinforcement learning (RL）
This term is not in the glossary yet. Browse related concepts in the glossary.
