With Claude, the main problem I notice is failure to commit, resulting in flailing. It seems to benefit from being forced to drill into ideas. Make it commit to the best of an imperfect set of options. Then break that option down into sub-options and repeat until finding a concrete next step.
Gemini does this drilling by default. But it seems to blithely hallucinate, exaggerate, and rationalize its commitments, rather than recognizing and accepting flaws.
I speculate these problems come from different corporate philosophies at Anthropic and Google on how to prioritize immediate AI tradeoffs. Anthropic’s prioritized anti-sycophancy, skepticism, and paternalism toward the user. Google seems to have prioritized optimism, a can-do attitude, and subservience toward the user.
I find it very helpful to “tag” individual subtopics or cases for Claude. Give things names like FOO and BAR, or RSI and GRADUAL-LOSS-OF-CONTROL.
Then Claude can use these different tags to keep different scenarios relatively straight. Or you can tell it to temporarily ignore a particular set of scenarios by name (“Let’s rule out all the ASI scenarios for the sake of this discussion.”)
This prevents it from thrashing back and forth between different branches of a scenario tree, and causes it to do case-by-case analysis. Without some structuring mechanism, it tends to over-focus on recently discussed things or pull in older tangents at random. With nice labels, it mostly keeps things straight.
With Claude, the main problem I notice is failure to commit, resulting in flailing. It seems to benefit from being forced to drill into ideas. Make it commit to the best of an imperfect set of options. Then break that option down into sub-options and repeat until finding a concrete next step.
Gemini does this drilling by default. But it seems to blithely hallucinate, exaggerate, and rationalize its commitments, rather than recognizing and accepting flaws.
I speculate these problems come from different corporate philosophies at Anthropic and Google on how to prioritize immediate AI tradeoffs. Anthropic’s prioritized anti-sycophancy, skepticism, and paternalism toward the user. Google seems to have prioritized optimism, a can-do attitude, and subservience toward the user.
I find it very helpful to “tag” individual subtopics or cases for Claude. Give things names like FOO and BAR, or RSI and GRADUAL-LOSS-OF-CONTROL.
Then Claude can use these different tags to keep different scenarios relatively straight. Or you can tell it to temporarily ignore a particular set of scenarios by name (“Let’s rule out all the ASI scenarios for the sake of this discussion.”)
This prevents it from thrashing back and forth between different branches of a scenario tree, and causes it to do case-by-case analysis. Without some structuring mechanism, it tends to over-focus on recently discussed things or pull in older tangents at random. With nice labels, it mostly keeps things straight.
That’s a helpful suggestion, thank you!