Cross-domain transfer learning. No task where training in one domain will result in predictive ability in another domain in the next 5 years
-
-
-
How is this not already false? And if I can ask that, it's not concrete enough to be bettable.
-
No system can predict the effects of gravity on objects in a video after being trained on textual explanations of gravity. Not now not in 5y
-
Well but vision is hard
-
Right, and language is hard too
End of conversation
New conversation -
-
-
Do you count Turing tests as a concrete verifiable tasks ?
-
Yes, but literal "the world will not end" doesn't interest me. If you have knowledge on AI limits you should be able to state lower limits.
-
Let's say no satisfactory machine translation.
End of conversation
New conversation -
-
-
Not in 5 years: translation of arbitrary text at same quality as a human professional.
-
In 5 years? I'd take that bet for a random text, decided by a blind panel of translators, for at least 1 language. How much will you bet?
-
What's random? And reasonably dissimilar languages. Spanish to Portuguese doesn't count. I'd bet a thousand euro.
-
Random published nontechnical or fiction book excerpts or popular press articles. And you'll bet 1k euro against my 100 euro?
-
Yeah, would bet.
-
Thus endeth the thread? C'mon
End of conversation
New conversation -
-
-
Cannot be done in 5 years: a bot that successfully answers randomly selected StackOverflow questions (not cherry-picked), w/ working code.
-
Cherry-picked seems plausible, "only answer the easy questions". Answering random non-duplicate questions with 99+% success, not so much.
-
99% plus? You'd need better than human ability, across all categories? I'd bet $1000 at 9:1 against C++ with 90+% accuracy within 2 years.https://twitter.com/ESYudkowsky/status/910566159249899520 …
-
Don't know the landscape, figured I'd start conservatively. That sounds plausible. Could also compare to top users, if you can avoid gaming.
-
Primarily, I wanted to avoid suggesting that a "smart FAQ bot" would suffice.
End of conversation
New conversation -
-
-
Build a house of known design in arbitrary terrain. Write a book containing at least 10 chapters that sells 500,000 copies at 10.00$
-
A misinterpretation got me thinking. 3d depth sensing of landscape -> house structure (CAD) that optimizes for it would be pretty cool
End of conversation
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.