P(correct) doesn't go down with token count if you have self-correction. It can actually go up with token count.
2,680 karma · joined June 30, 2022
There's enough plausible evidence linking social media to problems in youth that we have to bite the bullet and try something, then reassess afterwards, rather than be caught in analysis paralysis about what may go wrong with attempts to fix the issues.
Public services are there to be used, but there's a line that gets regularly crossed by profit-seeking entities who do not optimize for public good and see public resources as something to be used up as much as possible as long as they can save a dollar.
Or the scale.ai private benchmarks
But I'm looking at the present moment and see too many similarities with the fascist dictatorships of the past. The nationalism, militarism, unreasonable border disputes and territorial claims and irredentist attitudes. The US just isn't that, despite their history.
The US policymakers have figured this out with their chip export ban. Techies on the other hand, probably more than half the people here, are so naive and clueless about the reality of the moment we are in, that they support open sourcing this tech, the opposite of what we need to be doing to secure our future prosperity and freedom. Open source almost anything, just not this. It gives too much future power to authoritarians. That risk overwhelms the smaller risks that open sourcing is supposed to alleviate.
"French electricity prices turned negative as a drop in demand and surging renewables output prompted some nuclear reactors to power down."
The sensible thing will be to use renewables on top of nuclear rather than overbuild nuclear, like what France is doing. But that will also eat into capacity factor when it's more windy and sunny like what France experiences.
Just something to keep in mind when you hear people talk about nuclear costs. Multiply it by at least 1/0.7 if they haven't taken this into account.
But these are translucent. They'll be put onto windows. So even if they're 50% more expensive, it's serving an adjacent use-case.
The more interesting figure is they're getting 160W/m^2. That's very large, and it'll be a big deal considering there is more vertical surface area than rooftop area on buildings. Standard solar panels get about 220W/m^2.
Why do you say it's sampling programs from "training data"? With that choice of words, you're rhetorically assuming the conclusion.
If he only sampled 20 programs, instead of 8000, will we still say the programs came from "training data", or will we say it's genuine OOD generalization? At what point do we attribute the intelligence to the LLM itself instead of the outer loop?
This isn't meant to be facetious. Because clearly, if the N programs sampled is very large, it's easy to get the right solution with little intelligence by relying on luck. But as N gets small the LLM has to be intelligent and capable of OOD generalization, assuming the benchmark is good.
It's not related to risk, at least not directly. It's related to the supply of entrepreneurship as a factor of production. Entrepreneurship is scarce, so founders have leverage in any bargaining situation against early employees, who are more numerous and therefore less valuable and less powerful. If 10x the number of people tried to become founders, then founders would hold less leverage and the equity terms would become more "fair" because they'd have no choice but to give generous terms if they wished to hire people.