ControlsFreak
No bio...
User ID: 1422
ControlsFreak
14hr ago
But there will be an impact, that's for sure.
How do you have any idea what order of magnitude the impact will be? How do you have any idea what the sign of the impact will be?
ControlsFreak
19hr ago
I think the vibes have fully shifted on climate change damage estimates. Tyler Cowen posted this morning with a terse:
The whole climate to gdp transmission thing does not seem to be working very well?
He's referring to this paper and this thread about it. They perform an empirical review of previous major estimates, focusing on replicating them and analyzing the methodology. One thing I found interesting is that they distinguished between damage estimates, themselves, and applications of damage estimates, like SCC. They say that the latter have already been show to be irreducibly uncertain, though even if the damage->SCC pathway was not irreducibly uncertain, they are arguing that since the damage estimates, themselves, are irreducibly uncertain, so too would be things like SCC.
They spell out multiple factors that create identification challenges and show how small changes to the inputs of prior models can result in huge changes in the outputs, in strange and unstable ways. They don't necessarily think prior authors did anything actively bad or malicious in their approach, just that the entire endeavor is probably doomed from the start:
Importantly, we don’t think these particular papers are uniquely flawed; our point is that they are attempting an impossible feat...
Their tweet thread has the typical disclaimer needed to get out in front of the typical objections one would immediately hear upon taking such a position:
Importantly: we are not claiming that climate change is economically harmless. We're arguing that the magnitude of damages is deeply and irreducibly uncertain, and trillion-dollar decisions need to stop being made as if it isn't.
I feel a bit vindicated by the vibe change, because I had been arguing something similar a full decade ago at the old old old place, pretty much on my lonesome. Obviously, I didn't have the exact set of empirical critiques that these authors present today, but I feel like it's a good example of where you can have very strong theoretical knowledge in a related/relevant area (timescale-separated dynamical systems) that leads to a correct intuition along the lines of, "I don't actually have to know the details of the methods they're using (though I did look at several back in the day); I can't imagine they could possibly accomplish what they're setting out to accomplish, just because of the nature of the type of system they're working with."
ControlsFreak
1d ago
There has, from time to time, been some discussion concerning doctor salaries. I don't personally care all that much about this. They're highly-trained professionals in an in-demand field, and doctor salaries probably aren't the main driver of overall healthcare costs.
Nevertheless, there's often some debate over what the numbers actually look like. I was just linked to this tweet in one of my econ link aggregators. (Yay, built-in browser translation!)
Their claim is that 84% of American physicians are in the top 10% of incomes, and 26% of American physicians are in the top 1%. Their paper makes comparisons to other countries. They also broke it down into primary care vs. specialists.
So, at least this is one snapshot view of the actual distribution of doctor salaries, which I hadn't really seen before in these discussions. Assuming, of course, that their methodology is sound, which I'm not qualified to assess.
ControlsFreak
6d ago
if that made it work better
It seems to me that you are saying that you have goals for what you want the end product to be like. As such, I think you're implicitly affirming that you would choose to not do things like train on the test set. That is, you wouldn't just clearly and directly give it the answers, even though you could.
Now, the question seems to me, "What do you even mean by benevolence?" You originally said:
Lack of benevolence: God created the world and all that is in it, and is able to interact with it, but doesn't actually care about us.
But this sort of doesn't make direct sense. You care about the LLM you're creating. You deeply care about it, at least in that you very much care to "ma[k]e it work better". It seems like you're using some other sense of words that is not fully fleshed-out. Like, maybe to be benevolent, you have to care about some particular type of goal or in some particular way, but other types of caring/goals do not count, or something. I think we just don't have enough information to figure out whether this reasoning makes much sense.
ControlsFreak
6d ago
I drive 99% of the time, and my wife very very occasionally says things. She always apologizes about it, but somehow every. single. time. it is valid and useful information. For example, maybe I'm looking back to initiate a lane change, and something suddenly happens in front of us and to the other side.
ControlsFreak
6d ago
That sexual revolution thing didn't turn out so well for women, did it?
ControlsFreak
6d ago
If you were creating an LLM, would you train on the test set? If not, does that mean that you lack benevolence? You could just clearly and directly give it the answers!
Didn't you go for vaping, whereas Gwern specifically distinguished between gum/patches and vaping, even in the abstract of the essay?
More options
Context Copy link