I didn’t disagree-vote but I found your quick take confusing because the example you linked seemed on-trend with the general progression of leaked frontier CoTs (mostly OpenAI, though) I have seen over the past 1.5 years to me.
Thanks, so is the interpretation something like this?
2a. “this is what Fable’s CoT looks like, but it doesn’t follow that we should accelerate our timelines, because we already knew RL could produce this kind of behavior, so it should already be baked into your timelines if your timelines are good.”
I don’t mind disagree votes, I’m just interested to see what people think. I was very surprised to see that CoT, but I’m only really basing that on reading CoT from open models (Kimi, DeepSeek, etc.) that do sometimes have random meltdowns, but AFAIK never do something like this productively.
Not sure how to interpret the accuracy downvotes. :) Do they mean:
The explanation is something other than “random meltdown” or “this is what Fable’s CoT looks like”?
The explanation is “this is what Fable’s CoT looks like, but it doesn’t follow that we should accelerate our timelines”?
Something else?
I didn’t disagree-vote but I found your quick take confusing because the example you linked seemed on-trend with the general progression of leaked frontier CoTs (mostly OpenAI, though) I have seen over the past 1.5 years to me.
Thanks, so is the interpretation something like this?
2a. “this is what Fable’s CoT looks like, but it doesn’t follow that we should accelerate our timelines, because we already knew RL could produce this kind of behavior, so it should already be baked into your timelines if your timelines are good.”
I don’t mind disagree votes, I’m just interested to see what people think. I was very surprised to see that CoT, but I’m only really basing that on reading CoT from open models (Kimi, DeepSeek, etc.) that do sometimes have random meltdowns, but AFAIK never do something like this productively.
From hovering I can see there are just three accuracy votes (for a total of −11) so far, so someone apparently used strong downvotes.