>For example, we loaded the entire text of The Great Gatsby into Claude-Instant (72K tokens) and modified one line to say Mr. Carraway was “a software engineer that works on machine learning tooling at Anthropic.” When we asked the model to spot what was different, it responded with the correct answer in 22 seconds.
This sort of needle-in-a-haystack retrieval is definitely impressive, and it makes a lot more sense to achieve this in-context rather than trying to use a vector database if you can afford it.
I'm curious, though, whether there are diminishing returns in terms of how much analysis the model can do over those 100k tokens in a single forward pass. A human reading modified-Gatsby might eventually spot the altered line, but they'd also be able to answer questions about the overarching plot and themes of the novel, including ones that cannot be deduced from just a small number of salient snippets.
I'd be curious to see whether huge-context models are also able to do this, or if they start to have trouble when the bottleneck becomes reasoning capacity rather than input length. I feel like it's hard to predict one way or the other without trying it, just because LLMs have already demonstrated a lot of surprising powers.
This sort of needle-in-a-haystack retrieval is definitely impressive, and it makes a lot more sense to achieve this in-context rather than trying to use a vector database if you can afford it.
I'm curious, though, whether there are diminishing returns in terms of how much analysis the model can do over those 100k tokens in a single forward pass. A human reading modified-Gatsby might eventually spot the altered line, but they'd also be able to answer questions about the overarching plot and themes of the novel, including ones that cannot be deduced from just a small number of salient snippets.
I'd be curious to see whether huge-context models are also able to do this, or if they start to have trouble when the bottleneck becomes reasoning capacity rather than input length. I feel like it's hard to predict one way or the other without trying it, just because LLMs have already demonstrated a lot of surprising powers.