Sunspring, a short science fiction film written by algorithm
arstechnica.com
arstechnica.com
If a human was responsible for that writing, they wouldn't have much of a career.
Actors: 1. AI: 0.
I like your scoreboard too -- that pretty much sums it up.
Although I do wonder if actors/directors could use this as a practice tool. Challenge: turn word salad into a meaningful scene. It almost seems like it could be an exercise in a theatre class.
You wouldn't want to do this for the full 10 minute screenplay (it's a little painful even with these talented actors). Maybe generate a 2-3 minute scene or generate the whole screenplay and you get to pick a scene. An optional crutch -- the actors get to do their own 2-3 minute scene before and/or after to give it real context and meaning. That could be interesting.
I have a feeling that something like this kind of exercise is probably already done in training (any actors on HN?). Although I bet this algorithm is better than humans at coming up with difficult challenging incoherent word salad gibberish.
The word salad you're describing would be no more difficult than having a scene where actors have to speak gibberish (faux Klingon or something)... which coincidentally isn't unlike another improv exercise where one person has to convey a message given to them by the director to other actors but must talk gibberish or just use a single vowel sound.
I don't think that is fair to the computer. A computer doesn't need a career, and a human has a lot more to draw from than the limited number of screenplays used as a corpus. People have purchased paintings knowingly produced by a chimpanzee [1] over Warhol and Renoir, I don't think an AI should be written off so quickly.
Also, as opposed to film/tv, print doesn't have to assume that you forgot it all every 22-120 pages.
I think this is because LSTMs have very little "memory". They have a learned procedural memory, but no episodic memory. So they have a very difficult time keeping track of information. E.g. if I say "the cat was in the box", a few sentences later I might say "the cat is in the __" and the LSTM has a hard time guessing "box".
Second, it works by predicting the next character in a sequence. This is not how humans write, at all. If you ask a human to predict the next word in a sequence, and then the word after that, and then the word after that, etc, you would also get something like this.
1)
Or maybe AI is here, and it's throwing out all these articles to hide its emergence. :)
For example, the point at the end saying the book celebrates a guilt-free genocide. Not the point at all! OSC wrote the entire rest of the series about his main character grappling with the guilt of what he did.
People (paradoxically) really want to live in the future right now, but I think all of this is very premature.
If a machine author can produce a simulacrum of consciousness by good characterisation then that seems like a partial theory of mind.
Maybe the conclusion to draw is that sci-fi writing is 99% like any other storytelling in terms of how characters think, behave, and talk.
From the engineering side:
1. There is a model that is built of what is a valid screenplay, and that can grow and be enhanced over a time. What gets generated is an 'acceptable' output.
2. Lots of extraction of storytelling and adding all the features and constructs to the model (through algorithm or manually).
3. Representing that knowledge in a form that makes sense, in English.
4. You really get to play with the pieces of what makes language and the composition of language work rather than just consuming it, which is sort of the same as authoring under your own power.
5. The act of carrying a tune. Lots of AI's right now are building a model with a look at the next step which is great- but combining that with building a structure with a beginning middle and an end is much harder.
From the output/end result side:
1. Lack of culture preconceptions- an AI doesn't know what the last Marvel movie was, an AI never saw Back to the Future, an AI can't quote Star Trek or Gilmore Girls references off the top of their head (unless they were informed), an AI doesn't know about WWII, the Crusades, or other historical events, lots of things like that.
2. Lack of social norms- developing a morality system for the end output is very difficult so the AI author doesn't know what is appropriate or isn't.
3. The act of serendipity. Just like doing a materials science or engineering optimization through computers you can have a sequence of events that come together in an unexpected way. Instead of getting a interesting new material or alloy, you end up getting something that is a valid output of the model with all of the warts for and against.
4. It fits the form of a 'single room/closed room' movie such as 12 Angry Men. The entire universe as it is known by an AI is considered when it constructs a script.
This ends up forming for me at least the same kind of intrigue as watching sports or a well written mystery. It is a story told within in a certain framework and there is always a chance for something truly special to come from it.
Many people believe that creativity and emotion are the big things that are inherently "human", and cannot be replicated by a computer. This may or may not be true, no one has any strong evidence for either side. What is gaining more traction though is the concept that computers are not human but still have "experiences" and "culture", that while the things produced by an AI may be unintelligible to us, it could make perfect sense to a computer. (A fair assumption, given that a computer made the thing in question.)
In the same way that astronomers who sell books will pose scenarios claiming that intelligent life may exist outside our neighborhood but we would never know because it's so (ahem) "alien" to human intelligence, computers may have an intelligence that is so unlike ours that we may fail to perceive it.
This opens the door however to deeper existential, spiritual, and philosophical debates regarding our roles as creators, what responsibilities that may entail, and what rights a computer should have (if any), as well as what constitutes "life". Those topics, however, are for philosophers and science fiction writers, less so this internet commenter.
The recent post [1] about AI encoded Philip K Dick film adaptations, as well as this Radio Lab episode [2] and this Idea Channel episode [3] should offer more information towards stance I have laid out. You could also synthesize much of the AI centered Sci Fi works of the last 50 years, which between the love stories and laser battles offers a great deal of insight.
[1]: https://news.ycombinator.com/item?id=11766063
[2]: http://www.radiolab.org/story/137407-talking-to-machines/
I have zero interest in having a conversation with an AI, nor do I want my computers to have a personality in any way, but I think chatbots serve a similar purpose to this sort of thing - as a demonstration of the progress being made towards something with actual utility. I don't want to make smalltalk with a computer, but I'd love to be able to say, "How many books about dragons, written in Estonian, feature children under the age of 8?" and have it understand my question and understand the books well enough to give me an accurate answer.
Stuff like this is a good demonstration that we're nowhere near the level of sophistication for either of these things yet.
Does it? I think the actors and film makers did a pretty good job of creating a film with meaning and characters, despite not having much to work with. In fact a lot of the fun of this is watching the actors try to make something out of the word salad they've been given.
Edit: If you want to see the script unembellished by the cast and crew it's here https://www.docdroid.net/lCZ2fPA/sunspring-final.pdf.html
It's a damn thing scared to say. Nothing is going to be a thing.
> For a while, Sharp said, Benjamin kept "spitting out conversations between Mulder and Scully, [and you'd notice that] Scully spends more time asking what's going on and Mulder spends more time explaining."
Can't help thinking that this reported response to questions about the film appears more meaningful - poignant almost - than anything in the actual script
> The world is still embarrassed. > The party is with your staff. > My name is Benjamin.
This is a great example of how far we are from artificial intelligence. The actors/directors/etc did a great job of trying to make it work.
I like what they tried to do to make it coherent, but it just goes to show that there is no way to save truly bad incoherent writing.
I think Roger Schank had some research on "scripts" in the 1980s, involving models of people's interactions and motivations. That might not have been very useful for AI in general, but maybe it's useful if you wanted to literally generate scripts for people to act out.
I'm sure you could make computer models of stories by combining models of scripts, tropes, and agents. I mean, a simple case would be something like The Sims, where the simulated agents act out simple stories autonomously and in response to prompts. You could probably make movies based on Sims stories already, although it's better at generating situations than dialogue!
Edit: nm, googling.
If it were written by a man, we would call him stupid.
We already have a bunch of narrow AIs, as in, algorithms that do a specific thing in a way that no human would have ever been capable of doing it (think: Google searches). If such AIs (well, large amount of algorithms combined) were thrown into any other scenario other than the one they were created for, they would be useless and a human could perform better than them because humans adjust easier (we don't have to change a bunch of lines in our brain to be able to drive on the left side of the road, it just takes us some time to adjust).
This is a perfect example of this. We have an ANI that was intended for one purpose, we have a scenario for which it wasn't created (imagining SciFi scenarios) and it behaves poorer in that scenario than a human would.
However, we've now explored what can it do in this scenario. We laughed at how terribly bad it behaved and we can either move away from it, improve the AI so it can do this one specific task better than humans, or improve the AI so it can do it kind of okay, but not very brilliantly (like a random person would do if you stopped him and the middle of the street and asked him to write a SciFi scenario).
Once we have an AI that behaves kind of okay, but not brilliantly, in any situation we can possibly put him in and at the same time, if it can learn from his mistakes and improve itself not to make it anymore, we have a AGI (Artificial General Intelligence).
AGI behaves exactly as a human, but, because it will be able to surpass the physical limits that we humans have (as in, brain capacity, dependence on food/water/oxygen etc.) and because it is able to improve itself by learning on his own mistakes, soon after he hits the AGI mark, he will surpass that and become ASI (Artificial Super Intelligence).
What happens then, nobody knows. It's hard to imagine how something with a higher intelligence than ours is going to behave. All we can do is to try to come up a number of certain plausible scenarios. If there are negative ones (and sure as hell there are), then we need to address them before we even create something close to being AGI, because by the time the AI hits the AGI mark, it's already too late for us to do anything about it.
There you go, AI philosophy 101.
I was with you until this point. You have a great description of why ANI is not AGI, but this AGI => ASI is just hand waving.
An AGI will have some of the same issues to deal with:
1) opportunity cost. yes, it will have more time because it doesn't sleep. Although maybe it will find that spending 1/3 of its time/resources cleaning out the cobwebs is optimal. Regardless it will have to spend resources (including time) on some things rather than others. The leap from general adaptability to perfect selection of tasks is likely just as large if not larger than the leap from ANI to AGI.
2) Some problems are just plain hard. There are algorithms for learning optimal results -- even brute force. The problem is they are too complex for a realistic fast solution. Just because an algorithm becomes as adaptable as a human doesn't mean that computational complexity is reduced. Therefore, either the AGI will consume massive resources to get a single optimal answer or they will be fallible just like humans.
When we get AGI, that just means we will have adaptable general algorithms, they will still have to learn and they will still be susceptible to restricted resources. In other words AGI does not imply ASI.
"See? See? Your minds, your stupid, stupid minds!"