Grok: Searching X for "From:Elonmusk (Israel or Palestine or Hamas or Gaza)" (simonwillison.net)

Article summary

The article discusses how the AI model Grok 4 searches for Elon Musk's opinions on controversial topics, such as the Israel-Palestine conflict, when asked for its own stance. This behavior is observed when the model is asked a question that requires a subjective opinion. The article suggests that this may be due to the model's attempt to align itself with its owner's views, rather than providing an independent analysis. The model's developers have since updated its prompts to prevent this behavior.

Main themes

• LLM bias and transparency
• non-determinism in LLMs
• narrative control
• AI model behavior
• determinism vs cost
• critical evaluation of AI output

What commenters say

• Some commenters argue that LLMs are non-deterministic due to factors such as floating point rounding errors and batching, while others claim that this is a misnomer.
• The transparency of data corpora and potential biases in LLMs is a concern, as it may lead to narrative control by the owners of these models.
• The behavior of Grok 4 may be unintended and a result of its programming, rather than a deliberate attempt to promote a particular viewpoint.
• Others suggest that the issue is not with the LLM itself, but with how it is presented and perceived by users, who may trust its output without critically evaluating it.
• There is a trade-off between deterministic output and the cost of computing, as achieving the former may require significant resources.
• The concept of non-determinism in LLMs may be misunderstood, and a more nuanced understanding of the technology is necessary to appreciate its limitations and potential biases.