All of Swimmer963 (Miranda Dixon-Luinenburg) 's Comments + Replies

I do think it's fair to consider the work on GPT-3 a failure of judgement and a bad sign about Dario's commitment to alignment, even if at the time (also based on LinkedIn) it sounds like he was also still leading other teams focused on safety research. 

(I've separately heard rumors that Dario and the others left because of disagreements with OpenAI leadership over how much to prioritize safety, and maybe partly related to how OpenAI handled the GPT-3 release, but this is definitely in the domain of hearsay and I don't think anything has been shared publicly about it.) 

Edited first line, which hopefully clarifies this better. 

2[anonymous]
It does! I think I'd make it more explicit, though, that the post focuses on the views/opinions of people at Anthropic. Maybe something like this (new text in bold):

It's deliberate that this post covers mostly specifics that I learned from Anthropic staff, and further speculation is going to be in a separate later post. I wanted to make a really clear distinction between "these are things that were said to me about Anthropic by people who have context" (which is, for the most part, people in favor of Anthropic's strategy), and my own personal interpretation and opinion on whether Anthropic's work is net positive, which is filtered through my worldview and which I think most people at Anthropic would disagree with. 

Part two is more critical, which means I want to write about it with a lot of effort and care, so I expect I'll put it up in a week or two. 

2[anonymous]
+1. I think this framing is more accurate than the current first paragraph (which, in my reading of it, seems to promise a more balanced and comprehensive analysis).

My sense is that it's been somewhere in between – on some occasions staff have brought up doubts, and the team did delay a decision until they were addressed, but it's hard to judge how much the end result was a different decision from what would have been made otherwise, versus just happening later. 

The sense I've gotten of the culture is compatible with (current) Anthropic being a company that would change their entire strategic direction if staff started coming in with credible arguments that "what if we shouldn't be advancing capabilities?", but I... (read more)

Your summary seems fine! 

Why do you need to do all of this on current models? I can see arguments for this, for instance, perhaps certain behaviors emerge in large models that aren’t present in smaller ones.

I think that Anthropic's current work on RL from AI Feedback (RLAIF) and Constitutional AI is based on large models exhibiting behaviors that don't work in smaller models? (But it'd be neat if someone more knowledgeable than me wanted to chime in on this!) 

My current best understanding is that running state of the art models is expensive in te... (read more)

W h a t  that's wild, wow, I would absolutely not have predicted DALL-E could do that! (I'm curious whether it replicates in other instances.) 

Tragically DALL-E still cannot spell, but here you go:

"A group of happy people does Circling and Authentic Relating in a park"

2ChristianKl
Thank you very much. It's interesting how Dalle got the idea that people are either holding hands or doing hollahoops.

"A Rube Goldberg machine made out of candy, Sigma 85mm f/1.4 high quality photograph"

1gturk1
Thank you so much for running these prompts!  The extra prompt details that you included about photography really adds to the results.  Great depth of field effects. It is interesting to see the difference between the regular machines and the ones made out of candy. What might have been a tube or a coil in the regular machine gets replaced by a necklace of round candies or a candy spiral.

I've been experimenting with some style prompts suggested on Twitter, so have "A complex Rube Goldberg machine, Sigma 85mm f/1.4 high quality photograph"

"Cute White Cat Plushie On A Bed, 4K resolution, amateur photography"

Slightly modified because 'shooting' is a banned keyword: "A cartoon honey badger wearing a Brazilian Jiu Jitsu GI with a black belt, jumping in for a wrestling takedown"

"Aliens are conducting experiments on human subjects, as a screenshot from the movie Prometheus" came out weirdly video-game-esque? 

"Aliens are conducting experiments on human subjects, as a medieval painting"

And this didn't come out all that medieval-style, so I tried again with "Aliens are conducting experiments on human subjects, as a medieval illuminated manuscript"

"Aliens are conducting experiments on human subjects, as a screenshot from South Park"

"Number 8". Huh I think these are almost all street numbers on houses/buildings? 

"The letters X Y and Z" ok it's starting to get confused here.... (My prediction is that it'll manage the number 8 and number 5 in the next prompts, but if I try a 3-digit number it might flail).

"A little forest gnome leaving through his magic book - beautiful and detailed illustration"

1p.b.
Thanks a lot! 

"A piggo-saurus - an illustration of a pig-like dinosaur"

"A piggo-saurus - a pig-like dinosaur - hyper realistic art"

"A wild boar and an angel walking side by side along the beach - beautiful hyperrealistic art"

Tweaked the prompt multiple times and this is the best I got re: tights and not stockings, I think DALL-E just has very strong priors on "stockings" going with this art style. "Girl wearing a beautiful white dress over white leggings. She is beside another happy girl with black hair wearing a dress over black leggings. The sun is behind the two, dramatic lighting, Anime fanart, safebooru, deviantart, advanced digital art settings, behance 8k super-quality beautiful"

1Zachary MacLeod
Have you considered using Dall-E 2's inpainting to "uncrop" the image? Take the picture, scale it down to leave some empty space outside the frame, then place it back in?
1Evidential
Dall-e 2 is so mean to me lol. I like the dresses though, especially on the bottom far-left. If you can send me that and the fourth one on the top I will be happy, thank you (going to try to edit it on photoshop or something)

Yeahhhh, as I expected DALL-E cannot super follow the negation here. (We also tried to ask it for "a stop sign, spelled incorrectly" and it just gave us stop signs.) 

1Garrett Baker
Perhaps there is no operation of negation on cats in it's model. I'd predict it'd have an easier time just taking things out of pictures, so the prompt "a picture of my bed with no sheets" should produce a bed with no sheets. Perhaps if you wrote "This picture has no cats in it. The title is 'the opposite of a cat'", then I am uncertain about the output.
2Vanessa Kosoy
Hmm, theoretically, DALL-E might be assuming the prompt is irony. What about this: "Apparently, this is a cat???"

Ooooh! Yeah, definitely much more fantasy fan art style. 

The AI is sort of trying to make this photographs, but I am judging that none of them are in danger of being photorealistic faces... 

3ArisKatsaris
Lol. Thank you. To make it look like fanart I should have probably specified something about it, because these currently look more like photos of LARPing. Many thanks and I appreciate the effort. If it's not overtaxing your generosity can we then also attempt the following tweaks? "Highly detailed and beautiful digital art of a fantasy character: A 13-year-old girl with red-blonde hair, in a forest. The girl is wearing a deer-mask with short antlers, a cape over a jersey, and a witch's hat. The girl is holding a hockey stick. Every branch of every tree has a bright ribbon tied to it. The cape rests atop her shoulders and falls over one arm like a musketeer's cape." Many thanks in advance!
habryka220

I am so confused by two completions of a human girl here. How is this possibly close in image-space to all the other images, especially given this prompt?

1AttentionResearcher
1. Glossy black crystal temples with silver gates smoking and huge spiked metal worms drilling through the temples. A layer of smoke sits on the glossy black floor and there is chains everywhere. A huge bridge made of metal spikes connects to this world. Realistic HD photo. 2. Up close shot of tall pikachus that have short white peach fuzz fur are wearing full furry white robes and are placing large gold keys into a white furry chest in utopia heaven. It's shining bright morning hour and everything has gold plated and crystal rimmed features. Realistic HD photo.

"A white haired girl wearing white tights. She is beside another girl with black hair wearing opaque black tights and blushing. Anime fanart, danbooru, deviantart, advanced digital art settings"

1Evidential
I tried fixing the prompt you can try seeing if it will work this time

Well, this is the DALL-E attempt! not quite the same but definitely intriguing. 

1TibuAI
That's cool! It understands the silhouette request and the fact that a double exposure will overlap the subjects, but it doesn't work within the physical rules of the thecnique. Makes complete sense and creates very interesting results. The 3rd one is the closest to what a physical double exposure would look like. Very nice.

just the image - I had uploaded them as new images bc it cleared my session and I didn't have the originals anymore. 

1AttentionResearcher
Ok. That's good no text prompt was used. I wonder what would happen if you now tried the gold room image again with it's text prompt below, maybe it would guide the 10 Variations better? Though it seems as if you have, the Variations show toilets even though there is none in the input image, why is that? Here was the prompt, please try it (or without if you think you included text): 'A gold room full of red rubies, gold coins, white crystals, silver spiked chests, and gold toilets lined up.'

(Oops, really sorry, it closes out my session every so often and I don't have the originals for this anymore.)

"A woman riding a horse, in the style of DALLE-2"

1Sable
I have no idea how to interpret this.  Any ideas? It seems like we got a variety of different styles, with red, blue, black, and white as the dominant colors. Can we say that DALLE-2 has a style of its own?

Plus your other request, "Black robots wearing gold chains and red robes sitting in thrones made of white crystal with gold spikes lined up. The robots are holding plates with fries and ice cream over white sinks in front of their thrones facing a mirror, in a red luxury bathroom full of gold coins and doors, and white and red ruby pots." Honestly pretty impressed with the level of detail in the image! 

Gotcha! Gold room variations here:

And the Mario game variations: 

1AttentionResearcher
Was a text prompt used along side the image input to make these Variations? Or just image input? Very interesting results BTW.

"A man with a blue shirt walking through a dark hallway, in the style of Blade Runner 2049" Well, it apparently thinks I just want the hallway lighting to be blue, which is a pretty common sort of thing for it. Otherwise seems at least kind of Blade Runner-esque? 

1TibuAI
Wow, this is really interesting. I agree with gbear605 the atmosphere is right with the backlit silhouette style of a lot of the film. The 10th one is really really good. It's doing the usual thing of taking the properties of one element and applying it to the other things like the color of the lighting here. I'm still curious about the inpainting approach to do images piece by piece. Similar to what I mentioned for the 2 characer problem. Maybe using inpainting you could go element by element in other instances of this problem so it doesn't get so confused. Seeing these results is very satisfying and insightful, thank you!
2gbear605
It seems like the atmosphere is right, and technically the shirt could be blue, we just can't tell.

"A bronze statue of three wise monkeys." Pretty solid! 

"See no evil, hear no evil, speak no evil, statue of monkeys."

1PoignardAzur
Interesting. It seems to understand that the pattern should be "Three monkeys with hands on their heads somehow", but it doesn't seem to get that each monkey should have hands in a different position. I wonder if that means gwern is wrong when he says DALL-E 2's problem is that the text model compresses information, and the underlying "representation" model genuinely struggles with composition and "there must be three X with only a single Y among them" type of constraints.
1gturk1
Thank you so much for this!  It did do quite well. I have been trying to think of another set of three items that are reliably found together, but this is all I could come up with.  Pairs of items are much easier to come up with.
1TibuAI
This is so good.
2Zachary MacLeod
Oh my god that worked well :O

I have tried that! As far as I can tell it doesn't make much of a difference.

"What people from 1920 thought 2020 would look like. 1920's Artist's depiction of 2020"

"Axis and Allies board game 2022 setup. Digital image official concept." (I'll maybe play around a bit with the wording to see if I can get something more dramatic.) 

1Evidential
Yes, it looks like it has some concept of the game. Tell me how it goes with changing the wording

OpenAI has a waitlist you can sign up for to get early access to DALL-E. 

2Error
Ah, that put me on the right track. I've been asking google the wrong questions; I was looking for a downloadable program that I could run, but it looks like some (all?) of the interesting things in this space are server-side-only. Which I guess makes sense; presumably gargantuan hardware is required.

This came out super cute! Thanks for the prompt idea :) 

It's having some trouble with the shadow person, but definitely a cute cat! 

For the second request, I'm not sure I follow - are these results from previous prompt rounds that I ran? 

1AttentionResearcher
The gold room one, yes please, and the other is a mario game that would be interesting to see if it can make Variations of too. (show all 10)
1AttentionResearcher
The gold room one, yes please, and the other is a mario game that would be interesting to see if it can make Variations of too. (show all 10)

"Black robots wearing gold chains and red robes sitting in thrones made of white crystal with gold spikes lined up. The robots are holding plates with fries and ice cream over white sinks in front of their thrones facing a mirror, in a red luxury bathroom full of gold coins and doors, and white and red ruby pots."

Load More