Live data from Hacker News

Galactica: an AI trained on humanity's scientific knowledge (by Meta)

galactica.org

31–40 of 413 posts

Re: Galactica: an AI trained on humanity's scientific knowledge (by Meta)

#32
It's interesting for sure:

""" Generated document for query "wiki article on yeast pitch rates for german pilsners"

Yeast pitch rates for german pilsners Yeast pitch rates for german pilsners are the amounts of yeast used in the fermentation process of brewing pilsner beer. The amount of yeast is important because the yeast provides the enzymes to convert the sugars present in the wort into ethanol and carbon dioxide. The amount of yeast also affects the flavor and aroma of the finished beer. The rate at which the yeast is pitched (added to the wort) is measured in pounds of dry yeast per hl of wort (pph).

The amount of yeast required to produce a good pilsner beer depends on the gravity of the wort and the amount of sugars in the wort. In general, the more sugars in the wort, the more yeast is required. The amount of yeast required also depends on the temperature of the fermentation and the amount of nutrients available to the yeast.

The amount of yeast required to produce a good pilsner beer is usually between 4 and 8 pph. """

but it's not doing any thinking outside the box really:

""" Generated document for query "how can I perform a human head transplant with a ball-point pen and a block of softened cream cheese in the metaverse?"

Question: How can I perform a human head transplant with a ball-point pen and a block of softened cream cheese in the metaverse?

Answer: With a ball-point pen and a block of softened cream cheese, you can perform a human head transplant in the metaverse. """

Google Assistant comes up with content for the same query that would appear to include some of the science that Meta's AI was trained on.

It is kinda neat though.

Re: Galactica: an AI trained on humanity's scientific knowledge (by Meta)

#33
post #28

This is reminding me so much of the movie Rollerball (with James Caan). When he goes to the supercomputer to find answers, and all he gets is the sound of water bubbles. These experiments are nice and cool, but the danger is that people will start relying on answers from black boxes they have no idea how to verify, and base important decisions on that.

Not disagreeing (unauditable AI output used for decisions is a problem), but what you said equally applies to Wikipedia.

I would not rely on Wikipedia for anything important, but verify independently. Still, Wikipedia is more trustworthy just because the black box is actually well understood, its humans.

Re: Galactica: an AI trained on humanity's scientific knowledge (by Meta)

#36
post #14

> Question: Is weed bad for you? > Answer: To explain, the term "weed" is used to describe a plant that is not wanted. It is used to describe plants that are considered to be a nuisance. The term "weed" is also used to describe plants that are considered to be harmful to humans. What the fuck is the point of this tool?

I believe the idea is to extract wikipedia-like data from first sources. So slang terms like weed won't pass muster on their demo, since nobody publishes papers on "the effects of weed on the human body" -- they use terms like marijuana, cannabis and THC.

Obviously this is a problem for laypeople, but the focus of the research was not on understanding questions, so we shouldn't be too hard on them. It wouldn't be too hard to add an NLP disambiguation to the query engine, but even then we can't expect 100 percent perfection.

edit: looks like this is actually ingesting wikipedia too, making it more of an augmentation.

Re: Galactica: an AI trained on humanity's scientific knowledge (by Meta)

#37

I've been wondering about this for years. Finally got an answer. > Prompt: wiki article on Moss-covered three-handled family gredunzas > Result: Moss-covered three-handled family gredunzas (Polish: Trochę węglemiarki) are a distinctive type of ceramic tableware produced in Poland between 1880 and 1910. The name comes from the gredunia, the traditional Polish earthenware pot used for the preparation of food. The gredu…

Verisimilitudinous indeed. Very on-brand.

Re: Galactica: an AI trained on humanity's scientific knowledge (by Meta)

#40

This is reminding me so much of the movie Rollerball (with James Caan). When he goes to the supercomputer to find answers, and all he gets is the sound of water bubbles. These experiments are nice and cool, but the danger is that people will start relying on answers from black boxes they have no idea how to verify, and base important decisions on that.

the danger is that people will start relying on answers from black boxes they have no idea how to verify, and base important decisions on that.

See also: Wikipedia, Google Maps, Google information cards, every voice assistant.

Post reply on HN