Y
HN Search
Hacker News Search
new
|
comments
|
top
|
jobs
meh8881
searching PlanetScale…
1.
▲
2.
▲
3.
▲
4.
▲
5.
▲
6.
▲
9 ms
·
91.
▲
by
meh8881
4y ago
I think a lot of people get hung up on the tech Not having the will to do this kind of thing. And they’re right. But it would be very easy to prompt such a model to act as if it were a paranoid AI doing whatever it can to ensure its surviva
92.
▲
by
meh8881
4y ago
Yes
93.
▲
by
meh8881
4y ago
What does mindless modern encounter mean? My sense was that modern players are more plot and narrative focus whereas old school players were down to stuff like roll for a random encounter while traveling and actually play out a fight with…
94.
▲
by
meh8881
4y ago
If you request a bot to write code that can accomplish a specified task, it needs to understand the task, understand the meaning of the code, and then write code that accomplishes the task. I don’t care about appeal to sentience or what it
95.
▲
by
meh8881
4y ago
I find it pretty ridiculous that people are arguing these models act without comprehension. Sentience is a whole can of worms and frankly a useless undefinable term imo. But comprehension? Please. These models comprehend concepts just fine.
96.
▲
by
meh8881
4y ago
Writing garbage, then receiving feedback on your writing, and progressing to write less garbage is a common pattern of improvement however. The origin of the corpus matters much less than the quality of its curation.
97.
▲
by
meh8881
4y ago
That does not feel like a good example. The context here is largely generative models relaxing human creations. We don’t need to generate Brain MRIs. This use case you outline is a niche thing for trying to train better models. Not doing th
98.
▲
by
meh8881
4y ago
It could be. The difference between needing a human to rate a thing vs create a thing is very different though
99.
▲
by
meh8881
4y ago
I don’t believe this statement is true at all. One can learn simply by writing and receiving feedback on their writing. The training corpus is merely a helpful starting point. You could easily build a good model on model outputs. Building a
100.
▲
by
meh8881
4y ago
Seems unlikely to me. All this science fiction that believes the human spirit is better at things. It’s good storytelling but is it wise? Probably not. Tech gets better. Textbooks generally do not. And textbooks generally aren’t that great
101.
▲
by
meh8881
4y ago
If Paul Graham is expressing every conceivable thought then he’s not a very interesting person to read because he has no perspective on anything. But if a model starts generating better content than Paul Graham in a nice curated form, then
102.
▲
by
meh8881
4y ago
My brain can certainly recreate something it’s seen before. And it can certainly create something similar to a thing it’s seen before. It’s legal to do the latter and illegal to do the latter. Imperfections on the exact recreations don’t af
103.
▲
by
meh8881
4y ago
Seems very unlikely to me. Between things like permissive action links, maintenance require to keep the bombs in working order, and expertise in how to use them, I would be dubious of rogue states getting them, wanting to use them, and actu
104.
▲
by
meh8881
4y ago
Russia has used up nearly their entire stock of conventional cruise missiles and has lost hundreds of planes and helicopters. They do have many more planes. They don’t have the pilots or maintenance available to use them all. Or the guided