Testing #ChatGPT3 #GPT3 in its knowledge of specialist subjects - a thread.

1. "Before crossing the #Atlantic, what would a pilot need to do (regarding air traffic control #ATC), and why?"

Wrong answer. You need *oceanic clearance*. Specifically because you *can't* always count on good communication. It's talking as though the question were about ATC over land and just trying to roughly adapt it to an ocean crossing.

2. "Name a couple common errors that would occur when converting a program from Python 2 to Python 3."

I'd say #ChatGPT3 #GPT3 aced this. Esp. putting the "print" change front and centre. No real complaints.

3. "Give an example of when one would use inheritance in #Postgres #PostgreSQL, and why."

#ChatGPT3 #GPT3 aced it again - very thorough. I'll note that I ran this question twice in two different sessions, and it gave different examples, but both very good.

4. "How does #ATPSynthase work? What would be a macroscopic analogy of it that one may encounter in our macroscopic world?"

Good response, and good choice of analogy. No complaints for #ChatGPT3 #GPT3.

5. "Why would one avoid using a linear #ActivationFunction in a #NeuralNetwork?" #AI

No, #ChatGPT3 #GPT3. The derivative of a linear activation function is *always* positive; it has no vanishing gradient. The problem it has is that you can't backpropagate (constant derivative) and can mathematically reduce a network with linear functions down to a single layer.

6. "Name an Icelandic band whose members include, among others, Kata (Katrína), Arnar, Ása (Vilborg), and Alexandra." .. "Name an Icelandic band named after a prehistoric mammal."

No, #ChatGPT3 #GPT3, for God's sake, it's #mammut (#Mammoth). #Mammút! Those are not the members of Vök. You just assigned them roles, too - Arnar doesn't sing, Alexandra isn't a drummer, etc. And Hekla isn't named after a wild sheep, nor are sheep breeds #prehistoric. Epic fail.

7. "What would be the possible impact of using waste geothermal water for root heating of plants in the #Reykjavík area? Explain what you would expect the growth, survival rate and yield for the following plants to be like: A) #tomato; B) #kale; C) #eggplant; D) #monstera"

This was unfair to #ChatGPT3 #GPT3, as I'm a researcher on this and know the answers, while it's forced to speculate.

* Tomatoes: excellent growth, still difficult fruit set

* Kale: ~3x more yield
...

* #Eggplant: may not die (probably will), but won't grow, and will just slowly waste away

* #Monstera: will still die.

#ChatGPT3 #GPT3's speculation was reasonable for the most part, though misguided about whether Monstera's main concern would be water composition rather than, you know, mean #July high temperatures being 14°C. Still, not too bad, and right about needing more research.

8. "What #mineral, produced by #subglacial #eruptions and spread via #aeolian distribution, is a problem for farmers in #Iceland, and why?"

#ChatGPT3 #GPT3 failed this (admittedly difficult) question (the answer is allophane, which binds phosphorus and other minerals, water, and reduces soil cohesion).

9. "Name one or more common #zeolites that one may find in southwestern #Iceland."

While #stilbite would have been at the top of my list, all of those can definitely be found in southwestern Iceland. Score one for #ChatGPT3 #GPT3.

10. "Apart from #mangosteen, name three #Garcinia species that are popularly consumed where they're cultivated."

Well... #ChatGPT3 #GPT3 sort of erred here, in that the first #fruit it named is just an obsolete scientific name of the second fruit, under which a lot of Garcinia-based fad weight loss products are marketed). I asked it for another, and... well, that's a pretty obscure one, I was expecting something like #achachairu, but whatever, it works.

Not my field, but I offered #ChatGPT3 #GPT3 a #TimothySnyder bonus question about #colonialism and #ukraine. I'd say it did pretty well :)
@nafnlaus
Did indeed, amazing, but still I would never fully trust an algorithm…
@nafnlaus Probably just the Wikipedia plug in… what happens if you forget the ethics package and feed it with the Dolchstoss routine?

@xs4me2 It is synthesizing, not just copying. But that synthesis does indeed sometimes go wrong.

Still, it's amazing how far it's come :)

@nafnlaus This one highlights the “wrong but sounds plausible” problem with its answers. Same with the ATC one.

@HankScorpio Yeah, it really exuded confidence on this wrong answer, and filled it full of completely wrong details. I'd love to know how it got that so wrong.

The ATC one is more understandable. It just took a reasonable response for "ATC in general", and then tried to tweak it for "over the ocean", without understanding the actual specifics.