Wow, 500m is really tall.

This comparison doesn't really make sense. The 361 is an exponent and the trillions are a linear coefficient.

Ahh, ok. I hadn't noticed that captions are a feature I can turn on in the X app.

Feedback: I'm interested in discussions about sampling but +1 on not watching a video with sound.

What if each trait is a whole character? So like instead of having a target number to succeed at some task, the task is represented by some metaphorical monster that your trait has to fight.

This one is harder to interpret because the "50+ years" in the question is not equivalent to the "50 years" in the answer options.

I like how you removed the meaningless alignment of cells into visual columns, i.e. you removed visual structure that was not related to any structure in the task.

Hmm. There was a similar short story by Hoshi Shinichi, but it was about diamonds instead of gold.

I'd be curious to see similar models pretrained only on fiction.

In that case, I would expect that most passages are either 100% human-written or 100% AI-written, with a smaller fraction being assembled from both human and AI input.

This seems fine to me. Like, if the question is "if I point to a random token within this passage, what's the probability that it came from AI" then 100% is nonsense anyway. So I assume the question is "what is your median estimate for the fraction of this text that's from AI?"

That's bogus. Suppose I want to sell fanfic of your story. I'll get my friend to write a fanfic first and sell me a copy for $1. You get 30 cents. Then I write a fanfic of my friend's fanfic. I sell it for $1 million. You only get $90k and my friend and I split the other $210k.

That's just because they're low in everything. Protein per calorie in mushrooms is not far off from meat.

I think there is a spectrum of hope to despair that people experience, which is influenced by beliefs about the future and not just what is happening to them in the moment.

I wouldn't exclude the possibility that they're acting rationally but simply don't know exactly what will happen in the future. Like, we can't compare an expected value to a single sampled value, observe that the sampled value is less, and conclude that the expectation was wrong.

I.e. they claim to be imagining epsilon value, but they're actually imagining negative value, so of course when they multiply that by a large number of lives they find it repugnant.

I.e. when people imagine a life "barely worth living" they might actually be visualizing a life that sucks but that is worth hanging on to given reasonable uncertainty about things being better in the future.

For someone whose life is slightly negative in value, it could still be worth hanging on for a long time just to see whether things get better. This is also why I don't buy the "repugnant conclusion" as an argument against utilitarian ethics.

I would guess that pretraining is still accounts for most of the total optimization on modern AI agents, even if it's only ~7% of total compute. But would probably be a different story at ~1% of total compute.

Reports of miracles or magical phenomena are actually the work of technologically superior space aliens?