LessWrong posts (usually on AI safety research) often warn against “searching under the streetlight”, such as solving similar-looking problems that miss the “hard bits”. However, I believe incremental progress is given too little credit.
When attempting an ambitious goal like figuring out how to align superintelligent systems, it is tempting to focus on easy subproblems first. This could be to align dumb systems, or to solve the problem under some simplifying assumptions. However, these “easy bits” often don’t make any progress on the “hard bits” of the problem, but currently take up the vast majority of researchers’ time. A natural conclusion is that we should spend much more effort on attacking the “hard bits” early.
However, I think the current approach of first searching under the streetlight is an effective strategy. History has shown us that a great number of useful tools are lit up by streetlights! ChatGPT was a breakthrough, right? But it was just a fine-tuned GPT-3, which was just a scaled up GPT-2, which was just a decoder-only transformer, which was just a RNN + soft attention minus the RNN, and so on. However, when you stack enough of these incremental steps, you get Gemini 2.5, which seemed absolutely impossible in 2014.
OK, so incremental progress stumbled upon powerful AI systems, but the alignment problem is different. We are unlikely to similarly stumble upon a general alignment approach that scales to ASI, or at least unlikely to stumble upon it before stumbling upon ASI. However, if the “hard bits” require insights far beyond our current reach, then we have no choice but to start at the beginning of the chain. We need to continuously check when the “hard bits” are within reach, but I believe the main progress is made elsewhere. We’re going to do lots and lots of work that doesn’t go into the final chain, but we don’t know what the first link is, so there is no other way.
LessWrong posts (usually on AI safety research) often warn against “searching under the streetlight”, such as solving similar-looking problems that miss the “hard bits”. However, I believe incremental progress is given too little credit.
When attempting an ambitious goal like figuring out how to align superintelligent systems, it is tempting to focus on easy subproblems first. This could be to align dumb systems, or to solve the problem under some simplifying assumptions. However, these “easy bits” often don’t make any progress on the “hard bits” of the problem, but currently take up the vast majority of researchers’ time. A natural conclusion is that we should spend much more effort on attacking the “hard bits” early.
However, I think the current approach of first searching under the streetlight is an effective strategy. History has shown us that a great number of useful tools are lit up by streetlights! ChatGPT was a breakthrough, right? But it was just a fine-tuned GPT-3, which was just a scaled up GPT-2, which was just a decoder-only transformer, which was just a RNN + soft attention minus the RNN, and so on. However, when you stack enough of these incremental steps, you get Gemini 2.5, which seemed absolutely impossible in 2014.
OK, so incremental progress stumbled upon powerful AI systems, but the alignment problem is different. We are unlikely to similarly stumble upon a general alignment approach that scales to ASI, or at least unlikely to stumble upon it before stumbling upon ASI. However, if the “hard bits” require insights far beyond our current reach, then we have no choice but to start at the beginning of the chain. We need to continuously check when the “hard bits” are within reach, but I believe the main progress is made elsewhere. We’re going to do lots and lots of work that doesn’t go into the final chain, but we don’t know what the first link is, so there is no other way.