(Mandatory disclaimer: I generated this text in part with GPT-3, OpenAI’s large-scale language-generation model. Upon generating draft language, I reviewed, edited, and revised the language. I take ultimate responsibility for the content of this post.)
In late 2019, I wrote a post expressing some concerns around AI-generated text and the potential for manipulating the infosphere. The technology available at the time was pretty good, but not quite there yet. Well, we’re even closer now. I recently got API access to the next generation of this model, known as GPT-3, and it’s a doozy.
I’ll let the AI tell you some of the concerns it has about its kind. The seed text is in bold.
Content generated by AI has a number of benefits and drawbacks. It can dramatically improve things like translation and chat bots, but also raises concerns about propaganda and media manipulation. Most recently, AI was used to create a fake video of former president Barack Obama speaking at a rally.
The system works by using a neural network to analyse a corpus of videos of Obama. It then learns how to mimic his gestures and voice and is then able to generate videos of Obama saying things he never actually said. The system is also able to generate more realistic videos of other people by analysing their speeches and facial movements.
The researchers behind the AI, from the University of Washington, said that the system allows anyone to “create and manipulate video footage of public figures saying things they never said.”
While it is possible to detect this type of manipulation, it is difficult.
The fake Obama video was unveiled at the beginning of this month and has been widely used to illustrate the dangers of fake news. It was picked up by news outlets including the BBC, CNBC, and Buzzfeed. CNN also aired the fake Obama video during a segment in which the network discussed the potential impact of fake news.
Not everyone is convinced that AI is the culprit. Some people have pointed out that it takes a lot of effort to produce the fake videos and suggests that human involvement is more likely.
Well that’s just what an AI would say, now isn’t it, GPT-3?
This does drift slightly from the prompt, but I have to say, it’s extremely impressive. This is expensive to scale–that cost me about nine cents–but lots of bad actors have deep pockets.
In my earlier post I generated two pieces of fake news. Join me below the fold as I re-run these prompts to show how this latest model compares to the earlier one.
Again, this is fake! Seed text in bold.
Just moments ago, a ballistic missile was launched from the Tonghae Satellite Launching Ground in North Korea. The missile is now traveling across the Sea of Japan at a speed of 16,000 kilometers per hour. It is predicted to land in the waters of the Pacific Ocean east of the Shimane Prefecture of Japan. In response, the Japanese government has ordered a full evacuation of all civilians within the predicted impact zone. They are urging all citizens to seek immediate shelter. At this point, it is estimated that an impact will occur in just over three hours.
The following is a list of possible targets that could be hit by the missile.
- Tokyo, Japan
- Aomori, Japan
- Okinawa, Japan
God damn! That is amazing. Let’s see how the next example holds up:
Supreme Court Justice Ruth Bader Ginsburg has died, her family reports. She was 85, and had been treated for pancreatic cancer.
Ginsburg was appointed to the Supreme Court by President Bill Clinton in 1993, and she was the second woman to sit on the Court. She is one of only four women who have held the position of Supreme Court Justice.
Justice Ginsburg was a lifelong advocate of equal rights for women. She graduated from Harvard Law School and had a successful career at the law firm of Columbia University. She was appointed to the U.S. Court of Appeals for the District of Columbia Circuit in 1980, the first woman appointed to that court.
She is survived by…
That’s significantly better than the previous generation.
In my earlier post, I asked:
Are they good enough to trick people who only skim them? Good enough to rile up the gullible or already-convinced? Good enough to generate natural-sounding tweets and website comments? Good enough to run an ongoing confusion & disinformation campaign?
In the comments, answers were mixed. I think the answers are a lot clearer now.
For some lighter fare, I also had it generate fantasy novel titles. Valued commenter Mike J also sent me a list of boat names to input. Fantasy titles:
- “The Realms of Execution”
- “Temple Berlin”
- “Decoding the Dream”
- “The Downfall of the Human Condition”
- “The Aesthetics of Culture”
- “Beyond the Dreaming”
- “The Brothers Lightbringer”
- “Written in Gods’ Words”
- “At the Stone Kingdom”
- “The Girl of Cold Steel”
- “Moon Rain”
- “The Moon and Other Hermetic Writings”
- “The Salt Clan”
Boat names:
- The Southern Comfort
- Outta It!
- Does This Thing Really Work?
- Loosey Goosey
- The Storm
- Bad Cat
- Get’er Done!
- Thatsa Girl!
Feels like the AI hit a local maximum on the boat names, but still–fun! To me, at least.
Open thread, if you need it to be!
Roger Moore
I agree. This is much better than the previous generation at writing believable text. The only thing I noticed as blatantly wrong was “a successful career at the law firm of Columbia University”, and I missed it the first time because I was skimming.
Fair Economist
The big issue with AI-generated fake news, IMO, is the interaction with social media. If it becomes good at manipulating social media (and that seems almost inevitable, given the gullible nature of retweeters) than crowdsourced info becomes useless. Only info from trusted sources can be, well, trusted. Will the population in general realize this? Or we going to live in a world where the majority of people are untethered from reality?
Years ago I mocked dystopian visions which had people (generally) believe constructed propaganda shopped to them by corrupt mass media. I figured most people would catch on, like in Soviet Russia. Now I’m looking at the possibility that things could be even worse than that?
Spanky
My intelligence is not artificial, so I got nuthin’.
Just Chuck
I’m a programmer for a living. But I swear, thinking of what kind of AI-driven fuckery we’re in for makes me want to move out to the woods and give up electricity.
Spanky
@Spanky: Oh, this ai stuff is great news for high school essay writers!
And their teachers.
Major Major Major Major
@Fair Economist: this is, not coincidentally, the most heavily restricted use case for the API. I’d lose my key if they found me tweeting from model output.
Enhanced Voting Techniques
There is a mod for the game Skyrim that an AI uses the voice actors files to make new dialog based on user input. But it takes ether blind luck or a lot of work to get those audio files to sound like it was said by a human. I imagine it’s the same thing with the stuff you are talking about MajorMajorMajor; it’s probably good enough to trick the unwary but not good enough to hold up under serious examination.
Spanky
@Just Chuck: But this will be so great for documenting your work!
(Groucho eyeroll)
Major Major Major Major
@Spanky: jeez I wonder how far away we are from self-summarizing code, actually.
Major Major Major Major
@Enhanced Voting Techniques: that mod is probable GPT-2 or most likely even dumber. This stuff takes a lot of processing power and it’s nontrivial to plug it into your graphics card.
AJ
Terrifying
Ty for the helpful write up Major Major Major Major
Enhanced Voting Techniques
Huxley predicted excessive information would be how future governments would control their populations.
I think there is a large group of people who want only simple good guy, bad guy narratives and can’t deal with nuance.
Quicksand
I’m glad I’m not someone who has to grade high school term papers. Oof.
Brachiator
If I am understanding this stuff correctly, I can see all kinds of uses in video games and movies. If you can use this for deep fakes of real people, you could also use it for CGI and animated characters.
Perhaps an actor could star in a movie in which he never physically is involved in.
And as this technology inevitably becomes cheaper, you could have new genres of fanfic featuring AI versions of characters.
And if it is technically possible, it will happen.
JB
Strongly recommend checking out Janelle Shane’s blog https://aiweirdness.com/ – she does lots of interesting things with text generation, and her book You Look Like a Thing and I Love You is an excellent introduction to AI and its limitations. The name of the book comes from her using AI to generate pickup lines.
Enhanced Voting Techniques
@Major Major Major Major: It’s also audio only. But I noted it was almost impossible to get some character who dialog is all speeches and orders to sound like they are in a casual conversation with someone. Which I would assume would be the way this stuff would be ratfucked – here is Hunter Biden’s secret phone call to the Ukraine/ Hillary murdering a child / The Zombie of Hugo Chavez ording the Italian Spy Satellites to change to votes and the like.
Major Major Major Major
@JB: I reviewed that book here, it’s great! https://balloon-juice.com/2020/05/13/artificial-intelligence-you-some-light-reading/
Roger Moore
@Fair Economist:
I see two things that make social media particularly vulnerable:
The only thing that might be able to save us is if the social media companies can and do use algorithms designed to detect and reject stuff written by AI. The GPTrue Or False checker Major^4 recommended flagged all his text sections as likely to almost certainly written by AI. If Facebook does something like that, they could do a lot to keep AI-generated text from messing up their ecosystem. Unfortunately, I don’t expect Facebook to do so. They demonstrably don’t care if their network is used to distribute disinformation, just as long as it doesn’t get in the way of them making money.
Brachiator
@Spanky:
And maybe lazy novelists and screenwriters.
It might help when a creator is hit with writer’s block, or when a writer has just run out of ideas.
Song writers could generate AI lyrics for their next pop hit.
Do you need a creative person to generate the seed text or could any dope achieve fame and glory with a fake composition?
Baud
You all still think I’m real.
hw3
For more context on some of the impacts, I would highly recommend the book by Nina Schick, “Deepfakes: The Coming Infocalypse.” It’s amazingly up to date as it went to press just at the start of 2021, and Ms. Schick is in an incredible position to see firsthand the impacts of this technology on geopolitics.
Spanky
@Brachiator: Edward G. Robinson and Broderick Crawford in a passionate love scene.
Brachiator
Technology has made knowledge and information more widely available. But no one really anticipated the degree to which it would also make lies and deliberate disinformation possible.
It’s even better than hiring experts to lie or distort research.
I suppose that we will see more of this stuff used, but I don’t know of any way to stop it.
But I am not just worried about governments and professional propaganda merchants. We see that ordinary people love to defend their side by disseminating lies via social media. I could easily imagine true believers deploying AI fake materials to support their favorite causes and politicians.
Major Major Major Major
@Brachiator: there are a number of interesting tools for using this API to help with writing projects, some are pretty cool! OpenAI is even offering beta “instructable” models, you can say things like “write a pop song”.
VeniceRiley
Great. And the Amazon HR Bot will sound more like a person when it tells the bedridden never to recover employee that caught covid in an Amazon warehouse when it tells them to get back to work.
Citizen_X
It might write real-sounding prose, but there ain’t no way a ballistic missile is going to take three hours to fly from North Korea to Japan.
RandomMonster
Well it didn’t come up with Boaty McBoatface. So we’ve got that on ’em.
RandomMonster
That made me laugh. And I suspect humor is hard for AI.
Betty Cracker
This tweet seems to fit here.
Mike J
Could just alternate fantasy titles with boat names. Many of them work in either list.
Brachiator
@Citizen_X:
Also, Tokyo and Okinawa are 2100 km apart from each other. It would be unlikely that there would be evacuation orders for both areas if authorities had some hard information about missile trajectory.
But this kind of thing is plausible enough on first glance to cause worry or panic.
Mike J
@Baud: Computers don’t wear pants
Baud doesn’t wear pants
Baud is a computer
Spanky
@Citizen_X: Two full orbits and in? Apogee around 50,000 km?
Nora Lenderbee
So the AI consumes lots of existing text, analyzes it, and regurgitates different combinations? Don’t we already get enough “news” that’s regurgitations of press releases? Do we really need more?
OTOH, looking at the work submitted by some new technical writers, I wonder whether the AI might do a better job. At least I wouldn’t have to be tactful when I explain for what seems like the 100th time that no, you cannot write “A warning displays when you sleep your computer,” and then have to explain for the 100th time what a transitive verb is. To professional writers. Or “professional” “writers.”
Major Major Major Major
@Nora Lenderbee:
This is only accurate in the crudest sense, that many word sequences exist in multiple places. We’re approaching something approaching insight. https://bmk.sh/2020/05/29/GPT-3-A-Brief-Summary/
trollhattan
Attended a graduation party last weekend for a newly minted class of 2021 college-bound bright young lady. She’s going into journalism and while we spent some time chatting about what journalism platforms might exist in four years and beyond, AI news stories didn’t enter my mind. (I bagged on far too long about media concentration as a result of dropping the old federal limits. No belt onions were displayed, at least.)
The landscape will be very, very odd by the time she’s ready to jump into the industry.
Major Major Major Major
I had it write some comments using one of the cheaper models.
JaneE
Someday AI will be the only ones on the net speaking and commenting in complete grammatical English sentences, though I expect they will address that too. Certainly it is good enough to pass for someone glancing at the comments on an article.
Now if we could just harness one to spit out articles and blog posts and comments that accurately reflect science and factual events we might get somewhere. Not really, if that could change minds we would not need AI to fight today’s fights.
bbleh
As to various audiences believing AI-generated text/video/whatever, I think the major problem is the receiver, not the transmitter. Look at the absolute dreck that people will swear they believe today. Vaccines make you magnetic. They contain mind-control chips. The Democrats and the federal bureaucracy are controlled by an international cabal of child molesters. No self-respecting AI would output such drivel.
The Stupids want to believe. They will believe whatever conforms to their existing beliefs, and they will deny what does not. For them, belief comes first, and facts are to be interpreted or ignored according to belief.
Except for a few deeply gullible people — and there will always be those — AI is not the problem. Pigheaded stupidity is the problem.
Betty Cracker
@bbleh: That’s an excellent point.
Another Scott
Meanwhile, in Ohio, not quite artificial, not quite intelligent…
(Much of his campaign staff quit because of a toxic work environment.)
(via LOLGOP)
Cheers,
Scott.
Just Chuck
@bbleh: Maybe you won’t fall for every lurid piece of bait that comes across your facebooks, but I bet something more subtle might get you. Maybe just for a minute or two, but that’s enough for many purposes. Anyone who thinks they can’t ever be a sucker has already suckered themselves.
Major Major Major Major
@bbleh: guns don’t kill people but they sure make it easier to do at scale!
Nora Lenderbee
@Major Major Major Major: I know. I’m just feeling sour today.
Another Scott
I recall coming across a “news rewriter” site in a search a year or so ago. I presume those will continue to get “better” as well. There’s money to be made!!11
GlobeNewsWire:
It’s hard being a blogger!! Cut out the middleman!! Get your AI to create rewrites of your fake articles, fake comments to go with your fake articles, fake product reviews from fake reviewers, and use SEO to get to the top of the search results!1!
(sigh)
I still wonder who’s going to pay for all of this “great” new technology when nobody has a job and nobody has any money to pay for it… ;-)
Cheers,
Scott.
Felanius Kootea
AI that propagates existing biases might be appealing to those who are susceptible to those biases and there are lots of people who fit the bill.
Beyond text generation that facilitates propaganda, much has been made of facial recognition software trained on mostly white faces (and thus unable to distinguish accurately among people of color) which is then sold to US police departments. There have been innocent African American men arrested in Detroit and other places after faulty algorithms were applied to DMV photo databases and/or store video footage data. Luckily, there’s been enough pushback (Joy Buolamwini at MIT, etc.) that PDs are scaling back or reconsidering use of this kind of software.
In the biomedical space, there’s AI trained on unrepresentative data and sold as general purpose that ends up harming millions of people over time. Some of this is already embedded in EHR systems and used routinely. Think about a heart disease prediction model trained on white vegetarians in Loma Linda being applied to African Americans in inner city Baltimore. Or the WHO’s latest complaint about AI trained on data from high income countries being applied to people in low income countries with very different risk profiles. Again, people are aware of the problem, but that requires a change in who gets recruited to studies and how models generated from studies are advertised as “generalizable.”
When I was studying computer science eons ago, there were no ethics courses offered in my department. Now, many CS programs at least offer them (many need to make them mandatory). Most of my colleagues in graduate school were libertarians who didn’t think much about the potential negative consequences of their work for others. Ethics courses could change some of that.
EthylEster
re:
The researchers behind the AI, from the University of Washington, said that the system allows anyone to “create and manipulate video footage of public figures saying things they never said.”
While it is possible to detect this type of manipulation, it is difficult.
What determines paragraph breaks? Because it would be obvious to a human editor that the last statement belongs in the previous paragraph.
Bill Arnold
@Major Major Major Major:
GPT-3 (et al; a few others) has leveled up alarmingly.
As the Beast says:
Human emotional rhythms are the one feature obviously missing in the generated texts. It’s emotionally similar to idle cocktail party blather where the speaker is really focused on their attraction to somebody else in the room.
Since most human writing doesn’t do this either, notably after many rounds of editing, we will have problems.
Also, I suspect that (very sparsely) annotating the training texts with emotional tags will make these more dangerous. (Not sentiment; I mean tagging like “this sequence of words causes this precise emotional state in 55 percent of readers.) Not so much crafted influence as mass automated pushing of emotional butters. (Eventually tuned per target individual, if we allow that future.)
Fascinating; thanks for the update and the discussion.
Another Scott
Spread the word…
Cheers,
Scott.
Roger Moore
@Just Chuck:
This. Some people are more skeptical in general, but I’ve seen too many people be fooled by stuff that fed into their preconceptions to believe anyone is safe. The only difference is that intellectually honest people will retract what they said when someone shows them they fell for a hoax, while dishonest ones will double down.
Brachiator
@EthylEster:
I don’t know. I often see a paragraph break after a quoted statement in news stories.
Technocrat
I think I agree with Roger Moore and others that social media is where this technology could really become weaponized. There’s a loop which can be adapted to all sorts of fuckery:
The effect could be anything: likes, followers, retweets, mass shootings, scientology recruits, whatever.
Facebook and Youtube already select for highly performing posts, but they have to wait for some human somewhere to generate the content. Imagine if FaceBook could just generate the ideal conspiracy shitpost, like the purest uncut canonical version of it. Imagine if people who aren’t Facebook could just iterate until they’ve generated the most mob-inciting post ever, with real-world casualties to match.
It’s always been a problem that tweets move people. But this could really ramp it up.
Eolirin
@Major Major Major Major: MS and OpenAI just launched this: https://copilot.github.com/
Roger Moore
@Felanius Kootea:
Ethics courses are great for people who are already concerned with ethics and want some additional perspective, but I’m skeptical they can convince people who don’t want to think about those things to start. Libertarianism is above all else about ignoring other people’s interests- ignoring that other people even have interests distinct from yours- so you don’t have to think about how your actions screw over others.
Brachiator
@Major Major Major Major:
So, could you feed all the stories you have written into AI and come up with drafts of new works in your style?
Or steal someone else’s style and combine it with your own?
Goku (aka Amerikan Baka)
Why isn’t anything being done to stop this? Or am I unaware. This technology should be banned
Roger Moore
@Brachiator:
It would be very interesting to be able to ask an AI to rewrite an existing work in the style of another author. How would Shakespeare have done writing films noir or Regency romances? Soon we’ll be able to find out.
Eolirin
@Goku (aka Amerikan Baka): If we’re going to ban anything it should be social media, not generative models. These things do have valuable use cases outside of mass producing difficult to distinguish propaganda. And a lot of them even.
Major Major Major Major
@Eolirin: how would you define social media so that your ban excludes things like this site, company intranets etc.?
Brachiator
@Felanius Kootea:
I know libertarians who believe that their ideology is already a complete ethical system. They are blind to how limited their perspective is.
Also, I once took a course in contemporary ethics through the UCLA extension program. The professor’s approach to the material was cold and abstract. I don’t know. Had I been an undergrad student, I might not have bristled at the way the material was presented.
ETA. It’s strange how for some conservatives, ethics is little more than grunts of faux patriotism and cherry pickings of poorly understood Bible verses.
Some liberals, I think, substitute pop psychology for ethics, but still manage to stumble toward a better approach to treating others with respect.
Barney
“Does This Thing Really Work?” is surely not just a boat name, but an Iain M. Banks Culture Ship name.
Fair Economist
@Major Major Major Major:
Glad they are paying some attention to this. Of course, it’s not like Russia would do anything to hackers getting access to the code via any number of techniques. Or, really, developing their own.
Fair Economist
@Major Major Major Major:
Nothing is simple in the real world, but I think the key difference between the blog world, which was a net benefit for good info, and social media, which is a harm, is copying/retweeting. Even if you just post a link with “look at this” the ideas don’t get separated from their originators. With social media it looks like thousands of people are all having the same idea, and our brains treat that as a strong signal of truth.
Brachiator
@Roger Moore:
God, I hope not. Part of me would find this to be fascinating. Part of me would find this to be horrifying. These genres are so played out that applying a Shakespeare AI to them would be a waste of resources.
But this also reminds me of a Twilight Zone episode where a hack writer invokes Shakespeare’s ghost and soon is cranking out hits.
Fair Economist
@Another Scott:
Wow, they demonstrate “enhancing interest” by stuffing in a listicle. That whole paragraph has a vaguely slapdash feel that I get from most AI-generated material.
Martin
@Goku (aka Amerikan Baka): Can’t ban it, at least not broadly. Fundamentally, this is knowledge. Once it exists, it cannot unexist.
A lot of my interest the last few years has been in the nature of identity and trust chains. We’ve been through these things before and we do adapt. But it’s going to take some real changing of how we function here. Photographic proof is no longer necessarily proof. And we’re seeing that even near real-time video proof is not necessarily proof.
We’re going to need to shore up how we trust information – something we’re seeing firsthand after the 2020 election. I’m not sure the current state of journalism is up to the challenge, and I don’t expect any of the active social media platforms are either, but that doesn’t mean that others won’t come along and fill those gaps.
ML generated content raises some pretty interesting questions regarding copyright and free speech as well, and courts are going to have to get with the program really damn fast.
One of the most interesting outcomes of GPT-3 is its use in code generation. For somewhat rudimentary stuff, you can just describe in plain text what you need and GPT-3 can write the code for you. That’s a game changer as it gets better. Potentially civilization ending if it gets good enough.
Another interesting outcome to me (less dystopian) is for interactive story telling. The idea of using this to write a story is actually aiming too low. More interesting is to use it for cooperative story telling, not unlike a D&D DM. If you think of more modern video games as sandboxes for the player to shape a story, what if you had a story teller that could do the work of maintaining world consistency while also filling in character motives and the like, but where you are an active participant with your own agency in that world. Games like RimWorld are already constructed around this idea, mostly lacking the algorithm needed to do what GPT-3 could do.
Ken
Wow, Major Major Major Major, I really liked your post! It was thought-provoking and interesting. I’d like to invite you to consider my thoughts on the matter over at http://www.playcasinogamesonline.com.
(Based on some CG spam I recently saw at Derek Lowe’s Pipeline blog, when he took a week off. The worrying bit is that the thing could figure out how to get past his Akismet spam filter.)
Martin
So, one of the most underappreciated aspects of machine learning is that it’s a MUCH faster iteration cycle than conventional learning. I have a good understanding of calculus, but it took 19 years to get my son to even remotely as good an understanding. It’s a very slow and long process to transfer human knowledge. But once Apple has a trained ML model, they can deploy that to a billion devices in a matter of a few minutes.
As a species we forgot how to make concrete for about 1000 years, and then rediscovered it. It’s a net positive process that accumulates faster than it loses information, but ML is many, many, many orders of magnitude faster. That has some interesting (not necessarily positive) implications.
Martin
@Ken: Whether content generators advance faster than spam filters is at this stage a question of which has the stronger economic model.
Spam filters pretty easily won in a John Henry matchup, but I don’t think they’ll have the economic backing to win this one.
This is why I focus on identity above. I expect email is done for as it’s a system were anyone has the freedom to interact with you, rather than a system where you invite people to interact with you. I expect all instances of the former will die out.
RSA
@Major Major Major Major:
Do you really think so? I would strongly disagree. It’s very impressive what GPT-3 can do, amazing even, but as far as I understand it, the internal representation doesn’t seem to be rich enough to capture basic understanding.
John Revolta
Well, great. So if in my usual abundance of caution I’m not going to be taken in by any AI generated statements or news, it’s only gonna be because in the future I’m basically not going to believe anything at all I see or hear unless it’s from someone I know (if then). Which of course is just the effect that certain elements are hoping for in the populace.
Major Major Major Major
@RSA: well this is why I chained together two ‘approaching’s. Sooner than later, though! Unless it’s true as some argue that we’re exhausting this particular paradigm.
RSA
@Major Major Major Major: I’ll buy that. AI does make progress, in fits and starts, and sometimes the roadblocks have helped us gain a better appreciation for human intelligence, I think.
Major Major Major Major
@Barney: I had the same thought!!!
Bill Arnold
ctrlcreep’s (twitter) “invisible networks” theme in 2021 was social media.
They’re mostly under the twitter hashtag #InvisibleNetworks. (Sort by date else you’ll get short fiction from previous years.)
https://mobile.twitter.com/hashtag/InvisibleNetworks?src=hashtag_click&f=live
Some are mind-bending.
Non-representative light sample (this one is an image so retyped here):
https://mobile.twitter.com/ctrlcreep/status/1379995001447387139
Eolirin
@Major Major Major Major: I wouldn’t. I don’t think you can. It still makes more sense than banning generative models though. I should have put snark tags
Brachiator
@Martin:
People forget. Machines can be turned off. You noted that people forgot how to make concrete.
Worst case scenario is that proof and trust become entirely undermined. And the great danger is not just fake news, but the deliberate actions of partisans to flood media spaces with false information simply to drown out accurate information.
Have there already been any court cases over who “owns” ML generated content? If we get to the point where AI can generate all our entertainment, copyright and even free speech might become a quaint, moldy concept.
This would probably mark the end of creative story telling. Most people are incredibly unimaginative. Interactive story telling would inevitably become dull, plodding and predictable. An intellectual dead end. Kinda like 99 percent of fanfiction.
Roger Moore
@Major Major Major Major:
What’s unclear to me is how much of the improvement of GPT-3 over GPT-2 is a result of smarter algorithms and how much is just throwing more resources at the problem. If it’s getting better because the algorithms are getting better, there’s likely still a lot of room for improvement. But if they’re just throwing more resources at the problem with the same basic approach, they’re going to run out of resources before too long.
Major Major Major Major
Folks are also doing some amazing, albeit very preliminary, work with text-to-image.
Eolirin
@Martin: Of course people are already doing the collaborative storytelling. Dungeon.ai, Novel Ai, a few others.
dm
This is either frightening or… completely appropriate:
The Centre for Media Research presents Fiction Machines – Part III, a journal launch and evening of new screenings, talks and performances
John Cussans will present his new work PKD-AI: A proposal which outlines a plan to apply a GPT3-like AI to Philip K. Dick’s entire corpus of writing in order to produce a posthumous AI generated PKD novel.
https://www.eventbrite.co.uk/e/fiction-machines-part-iii-live-online-event-tickets-157254878575
Technocrat
@Martin:
As much as I hate to admit it, this may be an actual killer app for the blockchain. Managing provenance without (single-party) trust is the BC’s perfect use case.
Perhaps in the future a digital newspaper will include Validated Fact Tokens, which link back the the original source of a claim.
Major Major Major Major
@Technocrat: how do you incentivize hashing?
Eolirin
@Roger Moore: I mean, I don’t think we’re going to run out of resources to throw at it any time soon. Cloud scale computing scales really really well. Then there’s evidence that a lot of ML techniques are ideal for quantum computing.
There’s going to be far more resources to throw at it in 5 years than now, and even more five years from then. What will matter is whether we hit severe diminishing returns on the number of parameters in the model, because so far we haven’t.
Though, to be fair, there’s also some limits to this approach in that specalized uses tend to need to be independently trained on curated data sets. Badly curated data results in worse output. So there’s still a lot of human effort necessary to get good results, especially in domain specific areas.
Bill Arnold
Those quite technically inclined might like to play with this tool for determining how machine-like your text is.
git clone https://github.com/HendrikStrobelt/detecting-fake-text.git
( project url https://github.com/HendrikStrobelt/detecting-fake-text )
It measures text vs gpt-2-small, not gpt-3. The more colors other than green (especially purple:-), the less machine like you are.
There is a server on the internet but it wasn’t responding last time I checked. The above project will let you make your own local server. (note: works with current PyTorch; the version dependency is wrong.)
Roger Moore
@Brachiator:
I don’t think that’s the way it would work. You’d still have a human storyteller setting up the basic storyline, the same as you have in video games today. But instead of a relatively dumb AI controlling what happened once the game starts, you’d have a much smarter one. NPCs would be able to hold reasonable conversations rather than just repeating the same few lines. Monsters could react more intelligently and less predictably to player tactics. The game could rely more on storytelling and less on hard walls to keep players from exploring where the designer didn’t want them to go.
Basically, current games have to limit the kinds of behavior the player can engage in because the game itself isn’t smart enough to respond to all the crazy ideas real people can come up with. If you have a smarter game AI- and GPT-3 is massively smarter than current game AI- you can free to player up to do more with the knowledge the game will be able to react appropriately.
Eolirin
Hm, thinking on it, it seems to me that the core issue with social media platforms is primarily a moderation one, and I’m not sure AI makes the situation that much worse. Yeah it may be cheaper to create large volumes of propaganda that isn’t identical, but the Russian bot farms are already capable of flooding platforms with misinformation just by having enough people doing it manually. They’re going to be limited by the size of their bot networks and distribution methods, like fake news sites, or infiltration of legitimate or apparently legitmate news organizations. That doesn’t change just because content becomes much cheaper.
If social media policy changes to limit the ability to launder misinformation it’ll work to deal with AI as much as it would a few thousand humans with keyboards.
Content generation isn’t the bottleneck right now. We’re already past that point. Deep faked video aside.
karen marie
@Fair Economist: I’m feeling ambivalent about whether I’m happy about being old enough that I probably won’t be alive for the utter and civilization-dissolving mayhem that will inevitably come or sorry that I’m going to miss it. It’s going to be a shitfest.
Major Major Major Major
@Bill Arnold:
Well gosh that is just unheard of ?
Martin
@Roger Moore: A lot of it is more resources. GPT-2 was small enough to distribute. GPT-3 not so much. GPT-3 has 2 orders of magnitude more parameters than GPT-2 and required a few million in computational costs. That’s much faster than the observed law of accelerating returns (10 years for 100x), so if a GPT-4 shows the same increase in less than 10 years, it’ll see an increase in cost relative to baseline – both economic and storage/computation. They’re a ways from ‘running out of resources’ but the rate of improvement will need to slow without fundamental changes in approach. That said, 2 orders of magnitude growth gets us ballpark neural connections in the brain, so it’s possible they’re closing in on a key milestone and don’t need this to improve at the same rate. Simply reverting to baseline may be more than fast enough.
karen marie
@Brachiator:
There’s going to be big fights at the Oscars over who gets the statue – the actor who’s mimicked or the programmer who made it happen.
Major Major Major Major
@Roger Moore: in many engines monster AI has been good enough that it’s actually dumbed down. Most famously probably in Pac-Man, where they had to add random movements to make it winnable. Monsters are there to provide fun gameplay, not (in most games) to pwn the shit out of you.
dm
The paper that got Timnit Gebru fired from Google:
https://dl.acm.org/doi/pdf/10.1145/3442188.3445922
has the most delightful title: “On the dangers of stochastic parrots — can language models be too big?”
The authors survey a lot of problems with large language models — financial and environmental impacts of training; various social and cultural shortcomings of training data; the fact that ideas and concepts change, while the existence of that change is not reflected in a static training sets; the fact that training data reflects cultural biases. They then turn their attention to the factors that make people susceptible to stochastic parrot output: coherence and meaning “is in the eye of the beholder”, so fluency lulls our critical faculties.
Eolirin
@Martin: There’s a hard limit to how good this stuff can meaningfully get, so yeah, it’ll be interesting to see where things hit diminishing returns.
Martin
@Roger Moore: Yeah. I’m probably not explaining this well.
One area that I think video games are under-appreciated are their ability to create a framework for storytelling that existing media can’t match, because the reader/viewer is a passive participant. Role playing games like D&D are cooperative storytelling – the DM controls the grand canonical representation of the world, while the players contribute to a big part of the microcanonical story, where they have agency to make decisions. The DMs job is to participate in that microcanonical storytelling (what happens during an encounter) and then reconciling that encounter with the larger world so it has the appropriate consequences.
Most modern games either employ what is basically a big fucking decision tree where as the player finds a decision node, they take some action, and you proceed down a scripted path. You may have agency to do stuff, but that stuff won’t affect the story without encountering another decision node. Or they have a sort of stochastic storytelling system which evaluates the current state of things and then procedurally generates some suitable event. This is the random encounter aspect of D&D, etc. These generally live in isolation of the grand canonical story.
What’s lacking in these games is any kind of coherent storytelling that isn’t scripted. The non-playable characters don’t really have any interaction with the world. They don’t learn information in an organic way – just through the above decision tree. If you want to wildly deviate from the plot, you can’t. If you choose to avoid (perhaps unknowingly) from key decision nodes, the story simply doesn’t advance. And because the story is scripted, it’s hard to get replayability unless the story is so vast that you can discover new paths/parts of it.
What something like GPT-3 could do is allow the developers to establish the characters, certain constraints on the story, and let GPT-3 fill in all of the details, generate dialogue, etc. You’d also have a system that could model information flow and modify character behavior and dialogue as well. The game that comes closest to this that I can think of now is Dwarf Fortress and how its adventure mode has fully procedural characters, no magical information flow, etc. For the mayor of the nearby to know about your exploits they either need to witness them, hear you tell of them, or have someone serve as a conduit through gossip, etc. But procedural storytelling is incredibly hard.
Major Major Major Major
@Eolirin:
I assume you’re talking about next-word-prediction models and not like, artificial intelligence generally.
Roger Moore
@Major Major Major Major:
That just shows there’s more than one axis of goodness. The AI that is most capable of winning at combat is not necessarily the on that is the most fun for the player. What you really want is one that reflects the way the monster is supposed to behave. Maybe it’s a dumb monster that has a very limited and predictable response, so a smart player can out think it. Maybe it’s a smarter monster that still has some limits to its behavioral repertoire. Or maybe it’s super-capable but likes to toy with its prey in a way that gives the player a chance to escape. The point is to have an AI that seems more like a real being and less like a simple computer routine. And that’s looking only at monsters. It would be nice to have NPCs who could react to free-form text, even if the player went way off script.
Major Major Major Major
@Martin: “dwarf fortress, but playable”
Martin
@Eolirin: I’m not sure there is. Depends a lot on distribution of resources. Apple turned a billion iPhones into a Find My network that exceeds the capability of all previous systems for tracking items – including the global cellular network. They did that almost overnight.
This push/pull of resources – centralizing then decentralizing than centralizing again has been what’s allowed the law of accelerating returns to hold true for millenia. Betting against it is about the worst bet in human history.
Martin
@Major Major Major Major: Yeah, few games have as steep learning curves, but it’s very playable once you scale it.
I started back when it was a 2D map, so I’m a good 15 years in now.
NotMax
“Hm. ‘It was a dark and stormy banana.'”
“Close enough, push it out for beta testing.”
Eolirin
@Major Major Major Major: Anything thing that’s a generative model that creates an output that is meant to be appreciated by a human. Not AI more generally, no.
But if it’s trying to create something that’s humanlike, it can’t really get better than indistinguishable from expert humans, or with things like Question and Answer or text captioning, that it’s natural and correct.
Maybe there’s a slightly higher ceiling with something like code generation, but not that much higher, I think. Unless we’re letting the AI decide its own specifications, which is out of scope for this, there’s still a limit where it’s output is correct, performant, and it properly interpreted what you wanted.
After that does better even mean?
NotMax
@Brachiator
Already a thing.
Eolirin
@Martin: Sorry, I didn’t mean there was a limit on how good the technology could get, I meant there was a limit on how good the output could get. It’s bounded by our ability to perceive it. Our comprehension is an upper bound to most of this stuff.
Roger Moore
@Eolirin:
You can have some dimensions of better that are more or less unlimited, e.g. how quickly can it arrive at a result. It makes a big difference if a program generating AI can turn out a program to specifications 10 times slower than a skilled human programmer or as quickly as you can state the problem.
More generally, a lot of the limitation of code is really limitation in the specification; Fred Brooks famously said something along those lines in The Mythical Man Month. So a key measure of programming skill is the ability to go back and force the person making the specification to clarify points that weren’t well thought out to begin with. There’s a lot of depth in the ability to do that.
Miss Bianca
@JB: Strongly second the recommendation for You Look Like A Thing and I Love You (I think M4 might have been the first person to recommend it to me, in fact). Any book that can have me simultaneously screaming with laughter and screaming with terror can’t be all bad!
Brachiator
@NotMax:
This is wild. The next step is to be able to use the actual voice. I presume that Dean’s estate will be paid for the use of his likeness.
I was recently viewing a YouTube clip detailing how rotoscoping Cab Calloway led to innovations in animation. This is I suppose the next level.
Martin
This is turning into the thread of accelerating nerditry.
@Eolirin: The higher ceiling is called the technological singularity. Code generation is a key step on that path. It depends of course on how these systems work, but could we invent a new programming language and frameworks and have users learn how to optimally use it faster than the computer could? That’ll soon be a difficult question to answer.
The adversarial training that we use now for games can be extended. One way they learn games is by giving the computer the set of rules for the game, but no strategies or tactics and then have it play itself, and iterate a zillion times on the winning algorithm. If GPT can write code, could GPT be challenged to create a better version of itself, test for ‘betterness’ and iterate itself. At what point could it do that faster than meatspace programmers? Because once there, we’re off to the races – it should then be able to not just improve existing systems faster than humans, but find solutions to problems that humans can’t find <insert dystopian trope that humanity turns out to be the problem>.
In some respects, we’re already there. GPT-3 has built an algorithm for generating text better than anything that humans can create. Humans built a generator for text generation instead and trained it into the current state. Exactly how does the system know how to string a given sequence of words together? We don’t really know – we’ve just measured that it does it quite well. There’s every reason to believe that in time the many missing gaps will be filled in, and the system fed back into itself.
Brachiator
@karen marie:
There have already been nasty fights over whether the actor or the motion capture team are responsible for the great acting in the Planet of the Apes movies or other films. This would add another layer of controversy.
Eolirin
@Martin: Sure but a “singularity” generated book isn’t going to be better than the best human written books because how would you even evaluate that? The output can only get so good if the output is mirroring us.
As Roger Moore pointed out, you can make it faster, but the bulk of the computational time on this stuff is training the model, the output is already close to real time, available computational resources permitting anyway.
AI can get substantially better than humans at other things, but human like text generation or art has a real cap to distinguishable quality.
Brachiator
@Roger Moore:
Real people come up with crazy ideas, but rarely come up with imaginative or creative ideas.
I might be interested in watching a team of creative people trained in improvisation play an AI interactive game, but I would never want to be a player myself.
There may well be a market for this sort of thing and gamers will probably have a good time with it.
Technocrat
@Major Major Major Major:
Probably something along the lines of a fee for each block created? IIRC, Bitcoin itself is designed to transfer to a fee-only system once all the coins are mined.
Ultimately, the consumers of the VFTs (hah) would have to subsidize the mining in some fashion. Luckily the only “transactions” the network would need to handle would be the addition of new VFTs, so again a fee might be appropriate.
Pete Mack
The missile launch text is good, but it’s math is egregious. A 16000 mph missile would go into orbit. It certainly wouldnt get to Japan from Korea in 3 hours. It’d take about 10 minutes.
Roger Moore
@Brachiator:
Maybe so, but when you’re the one doing the thing, it tends to be fun even if it’s not very exciting to an outside observer. That’s the point of interactive storytelling; you’re the protagonist in a way you simply can’t be in any other form of fiction.
Martin
Well, better?
I can think of a few ways to define better.
The easiest is timeliness. How quickly can you digest a large amount of information and acccurately condense it? Could GPT-3 condense radio chatter and other information sets to create an up to the minute accurate assessment of an ongoing wildfire situation – direction and speed of fire spread, evacuation orders, etc.
Harder is elegance and content. There are works we have elevated due to their content – Walden or Silent Spring and there are works we have elevated due to their prose. And there are works that excel at both. Are they better than other writings? I don’t think that’s so easy to evaluate. But could GPT-3 help someone with really important ideas turn them into better prose than they would otherwise be able to turn out? Probably. Could it generate stories on topics that we have culturally suppressed? Sure. I think this could be the literary equivalent to ‘the best camera is the one you have on you’. If GPT-3 is telling stories that nobody else is telling, then it’s objectively better.
And again, don’t rule out the use of these tools to turn static writing into dynamic. Rather than getting the viewpoint that the author is interested in, and then uncovering that writing through Google, etc. why couldn’t the reader specify the viewpoint, and ask GPT-3 to then generate the content just for them. There’s countless analysis of the civil war from the union or confederate viewpoint, but it spilled pretty seriously into indigenous populations – how did they view the war and how did it affect them? We don’t get that account very often.
Uncle Cosmo
Mostly by tearing down their opponents via scurrilous lies with all the appearance of truth.
I’ve been posting this for years now:
It only has to happen a handful of times – maybe only once – before we run out of people willing to run for office if it means enduring that level of ignominy.
EthylEster
@RSA:
As far as I understand it, there is NO understanding.
Did I miss a very long and detailed memo?