Harsh.

swe @google, prev CS @uwaterloo

[Harsh. @harshsa28]

@ronawang Is this tweet chatgpt generated?

[Harsh. @harshsa28]

@TheRealAdamG Genuine ques: Do you think there's value in interviews? I feel like a lot of them are mostly corpspeak and done before the launch of a big product (to get the public thinking about the topic/product)

[Harsh. @harshsa28]

@paulg @tommycollison ChatGPT it (using it as a verb now). I'm sure Oxford dictionary of quotations was in the training set. Even if it wasn't, the collection of aphorisms on the internet must be a superset of Oxford dictionary.

[Harsh. @harshsa28]

They both use GPT-3 under the hood but it's interesting to see the differences in their outputs. DramaTron's output was more conversation focused while ChatGPT's output was more text focused, but I am sure it can be improved through better prompts.

Which one did you like better?

[Harsh. @harshsa28]

Just wrote a short movie script using 2 AIs - @DeepMind's Dramatron and @OpenAI's ChatGPT. The plot is a covert space race between the US and Soviet Union that continues to this day.

Scene #1: US astronaut Peter Burns discovers the Neutron Bomb.

White: Dramatron. Gray: ChatGPT

[Harsh. @harshsa28]

Scene # 4: Jack Daniels is being questioned by Commander Hobbs about his reasons for betraying the US and stealing the Neutron Bomb research and giving it to the Soviets. Jack replies that Mutually-Assured-Destruction is important. Hobbs accuses Jack of being a naive idealist.

[Harsh. @harshsa28]

@InternetKatty

[Harsh. @harshsa28]

@fchollet What do you think about ChatGPT? It seems to be generalizing quite well.

[Harsh. @harshsa28]

OpenAI's ChatGPT just rick-rolled me...

[Harsh. @harshsa28]

@yishan Poe's law

[Harsh. @harshsa28]

@tendstofortytwo @bloodandcoffee Guess 2147483647

[Harsh. @harshsa28]

@fchollet @amasad Almost every frame of a movie can also be used as a high quality image. Even if we are picky and select 1 frame out of 60, we still get tens of 1000s of high quality images from each movie.

[Harsh. @harshsa28]

@fchollet @amasad Couldn't you get a huge amount of very high quality images from youtube (video frames)? For example, there are videos of Japanese culture in 4k. Any frame of that video would be excellent for training image generators.

[Harsh. @harshsa28]

@tendstofortytwo Is mastodon any fun?

[Harsh. @harshsa28]

@tendstofortytwo I think either 1) new prog langs (go rust) are just as quirky as old ones (cpp), but ppl haven't used them enough yet or 2) users grow more dependent on the language, eventually getting blindsided by the quirks (which were fine initially) - eg in 80s and 90s, people loved C

[Harsh. @harshsa28]

@tendstofortytwo Interesting way of getting an extension

[Harsh. @harshsa28]

@jachiam0 The responses could be good data to make AI-generated art more realistic. Nice switch from "using captchas data to train image recognition models" to "using captchas data to train art-generation models"

[Harsh. @harshsa28]

@fchollet So cherry picking should be fine in low risk systems, given that the model is good at head and torso cases

[Harsh. @harshsa28]

@fchollet Worst-performing cases only matter in high-risk systems (cars, planes), right? In all other cases (ads, recommendations, etc), as long as you are good at 90-ish % of the cases (head and torso cases), you can safely ignore the worst-performing cases (tail cases, which are rare).