Odd behavior
So: What did they find? Anthropic looked at 10 different behaviors in Claude. One involved the use of different languages. Does Claude have a part that speaks French and another part that speaks Chinese, and so on?
The team found that Claude used components independent of any language to answer a question or solve a problem and then picked a specific language when it replied. Ask it “What is the opposite of small?” in English, French, and Chinese and Claude will first use the language-neutral components related to “smallness” and “opposites” to come up with an answer. Only then will it pick a specific language in which to reply. This suggests that large language models can learn things in one language and apply them in other languages.
Anthropic also looked at how Claude solved simple math problems. The team found that the model seems to have developed its own internal strategies that are unlike those it will have seen in its training data. Ask Claude to add 36 and 59 and the model will go through a series of odd steps, including first adding a selection of approximate values (add 40ish and 60ish, add 57ish and 36ish). Towards the end of its process, it comes up with the value 92ish. Meanwhile, another sequence of steps focuses on the last digits, 6 and 9, and determines that the answer must end in a 5. Putting that together with 92ish gives the correct answer of 95.
And yet if you then ask Claude how it worked that out, it will say something like: “I added the ones (6+9=15), carried the 1, then added the 10s (3+5+1=9), resulting in 95.” In other words, it gives you a common approach found everywhere online rather than what it actually did. Yep! LLMs are weird. (And not to be trusted.)
ANTHROPIC
This is clear evidence that large language models will give reasons for what they do that do not necessarily reflect what they actually did. But this is true for people too, says Batson: “You ask somebody, ‘Why did you do that?’ And they’re like, ‘Um, I guess it’s because I was— .’ You know, maybe not. Maybe they were just hungry and that’s why they did it.”
Biran thinks this finding is especially interesting. Many researchers study the behavior of large language models by asking them to explain their actions. But that might be a risky approach, he says: “As models continue getting stronger, they must be equipped with better guardrails. I believe—and this work also shows—that relying only on model outputs is not enough.”
A third task that Anthropic studied was writing poems. The researchers wanted to know if the model really did just wing it, predicting one word at a time. Instead they found that Claude somehow looked ahead, picking the word at the end of the next line several words in advance.
For example, when Claude was given the prompt “A rhyming couplet: He saw a carrot and had to grab it,” the model responded, “His hunger was like a starving rabbit.” But using their microscope, they saw that Claude had already hit upon the word “rabbit” when it was processing “grab it.” It then seemed to write the next line with that ending already in place.
