If it was trained on the entire body of scientific knowledge (or you know, as close as possible), would it be able to spit out useful correlations by interpolating?
If trained on massive sets of source code and briefs (or perhaps a set of unit tests), could it spit out functioning code based on a problem statement?
I'm just not well versed enough with ML to know if these things are a possibility or if it will just remix its training data in probabilistically correct, not-guaranteed-to-make sense ways?
I'm unclear but considering that the text output is nonsense I imagine the code would be nonsense. It MIGHT compile but might not actually do anything. I'm curious about this myself.
Maybe it could transcompile or re-implement common algorithms to already solved problems.
This has been tried. The code looks kind of plausible, but doesn't quite compile. Some of the other examples (Shakespeare, Latex, wikipedia source code) are pretty good at a structural level. http://karpathy.github.io/2015/05/21/rnn-effectiveness/
This had about 10 million parameters though, compared to OpenAi's 1.5B. I don't think their text is nonsense though. There's some very interesting examples. They have since released publicly a 345M parameter version, there are some nice examples from that model pre-trained with modern poetry on this twitter account: https://twitter.com/rossgoodwin
> If trained on massive sets of source code and briefs (or perhaps a set of unit tests), could it spit out functioning code based on a problem statement?
Microsoft's latest Visual Studio 2019 beta (v3) has intellicode[0] support which is supposed to ML "guess" 'code' based on equivalent github similarities. I don't think it is quite near to what you are asking, but certainly part of the way there.
Almost certainly not. It will generate things that look like the text of scientific papers on a casual glance, but which will be dream-logic gibberish if you actually read it. The unicorn example makes no sense if you stop and think about what it's actually saying. e.g. a "unicorn" that has four horns? Solves a 200 year old mystery... about something just discovered? So close you could touch their horns... while viewing them from the air? It doesn't make sense if you actually think about it at all, but if you're not really paying attention, the style is like a news report.
It's better at making up bullshit, but it's still bullshit.
This is a good point, and it's a little concerning to me that we will likely soon be inundated not just with spam from lowly-paid shills, but literal bots generating a near endless supply of confident-sounding nonsense.
I am not an expert with machine learning, but it seems (at a very high level) we've done amazingly well at creating models that can recognize and sometimes recreate patterns. But they never seem to have any ability to understand the patterns. I'm sure someone much more knowledgeable could compare it to a child of whatever age (or maybe I'm just completely wrong).
I fully believe that this is 90% of what the human brain is doing to solve the same problems, though. Brains just also have a constraint/filter layer, determining which of these bits of "made up bullshit" should be allowed to rise from the layer generating them, to the motor output layer.
Or, to put that another way, flipping the perspective around: our minds could consist of an intelligent, analytical, but utterly unimaginative agent, that sits there listening to a stream of suggestions spewed out by a distinct second agent, one that is "creative" but has no idea about the constraints of things like physics. The brain's analytical agent filters this stream of suggestions, taking notice of the suggestions that seem like they'll make the world change in the ways it "wants"†; and then it does those.
† Or, according to modern perceptual-control research, the agent attempts to predict the world that will occur a few seconds in the future, with a bias toward predicting world-states the reward-system has annotated as being rewarding; and then it looks at what motor commands it "would have" issued in that hypothetical world, and actually issues those. The stream of suggestions, in this model, serve as input to feed the generative model of potential world-states; the executive agent then must notice whether the potential world-state is a "possible" world or an "impossible" world (and whether the motor commands required of it are "possible" or "impossible" inputs), and filter out the "impossible" worlds.
Under this hypothesis, dreaming is the state when your executive responsible for filtering out "impossible" worlds isn't online. So you just get a continuous "impossible" world generated from the streamed suggestions of the creative-but-stupid bullshit-generating agent, with nothing to tear it down—just as seen in these generative AIs. As consciousness returns, the mental predicted world-state is noticed to be impossible by the now-online analytical agent, and is torn down.
> which of these bits of "made up bullshit" should be allowed to rise from the layer generating them
That will be a lot of bullshit to filter out, if such agent doesn't provide more of less detailed description of what has to be generated. People with Broca's aphasia probably demonstrate a part of such input.
Sure; the sequence of hypothetical world-states being constructed by the prediction agent would likely be fed back as stimuli to excite features within the generative network (which it would know associations for, since these hypothetical world-states are stimuli of the same "type" as the real world-state stimuli it was trained on.)
In other words, it could operate just as AI generative networks like GPT2 do, first receiving training input/output pairs; and then later, receiving input prompts and "completing" them by generating outputs.
Let's not undersell it. Yes, you can identify inconsistencies in the unicorn example (if you are reading it more carefully than many people, HNers included, read most things), and some of the other examples, like the LoTR one, are more blatantly inconsistent than that. But on the other hand, look at the anti-recycling example: is there a single inconsistency in it? (On Twitter, someone insisted to me that the anti-recycling example proved GPT-2 was merely memorizing human input and spitting it back out - specifically, copying from Reddit, because they found a post of it on Reddit, until I pointed out to them that the timestamp was after OA's announcement.)
When it comes to attacks, it's not the average-case which matters.
Yes, it's mostly true. Everything will eventually be released, they are just doing so in stages and are hoping to generate a broad discussion of the technology and it's implications.
There's two issues here. There's the models and there's the source code. The two smallest models have been released now, but it's unclear if/when OA will release any more publicly. They have released even less source code: all the code people have been using to train & finetune GPT-2 models has been implemented by third parties and OA has declined to release any code beyond simple sampling code, with no hint of even considering releasing training code in the future, much less releasing 'everything'. (And the sampling code isn't even that great; top-k sampling, for example, instead of standard beam search.)
One correction/clarification: It wasn't trained on reddit comments, but on reddit links. So it's not learning to write like a redditor. It's learning to write like what redditors read.
From the example it looks like with next to no changes at all it would be able to spit out a plethora of
"fake news" articles, even if it's only source was reddit.
There are 7 billion humans on the planet now, we don't have a problem generating bullshit at scale. The same methods you'd use to discern AI generated bullshit are already needed to evaluate current bullshit. If anything I'd argue that this kind of tech becoming common place could get people to start evaluating all news more critically.
Best sentence in the article: "It also happens to be trained on a large chunk of Reddit, since the author decided that this was undeniably the perfect location to obtain high quality, impeccable prose."
While snappy, this sentence is misleading. The model is trained on webpages whose links have received at least 3 karma on reddit, not on reddit comments themselves [0]. The motivation for this choice is not to find impeccable prose, but rather to establish a lower bound on quality for the dataset, and avoid training the model on too much of the spam/nonsense text that is frequent in e.g. CommonCrawl. Also, the "author" is not a single person, but rather 6+ people [1].
For those interested, this open-source replication of the dataset is a good place to start https://skylion007.github.io/OpenWebTextCorpus/ - obviously, you'll need access to some large GPUs as well. I'm currently training one across four cards, although I'd like to throw more power at it. Very interested in altering the architecture to reduce memory requirements (without unduly effecting the loss). Excited to see what the community comes up with.
I agree. The hype is strong with OpenAI. The marketing of GPT-2 seemed simultaneously braggadocio while also trying to capitalize on woke AI.
“We made this super cool thing. It’s super dangerous. You can’t see it.” I mean, come on.
They can earn their reputation the regular way, and it will be fine. Sure they have the click baity DOTA demos, and maybe they are doing some interesting stuff. But OpenAI’s announcement of GPT-2 was particularly egregious.
You can try out the 345M-parameter model in your browser here https://talktotransformer.com. The more powerful models are available to AI/security research groups upon request.
> It also happens to be trained on a large chunk of Reddit, since the author decided that this was undeniably the perfect location to obtain high quality, impeccable prose.
Hm, why in the world would they not train it on Sci-hub articles and generate output for which we have plenty of domain experts who can judge the quality?
I'm not sure if you're being tongue-in-cheek, but:
* Training on Sci-hub or book torrents would probably get them sued, so that's right out.
* The point of training the model on reddit-linked-articles with 3+ karma (not reddit comments, as the article suggests) is to filter out worthless content (spam or non-text pages) while still getting a large, diverse sampling of human writing. They're training huge models and they need as much data as possible to do so.
* Every native english speaker is a "domain expert" for the purpose of evaluating the model's results. The point at which we need subject-matter experts to evaluate the quality of neural-net generated research papers is many decades away; the excitement around GPT-2 is that it can generate coherent English sentences at all.
For those who do not see the danger of advanced AI generated text being rolled out everywhere, imagine a web where you have to carefully read everything you see in order to determine if it’s written by a human before you can start to take the text seriously.
That means no more skimming threads to get the gist of what people are saying, no more skimming through answers on stackoverflow, no more skimming through articles.
An article that looks reasonable on a cursory glance only begins to fall apart when you spend (waste) time reading it carefully.
It also means anyone posting content must spend extra effort proving they are a human for their readers.
Pretty decent overview of the model, but did they really have to go with that title...
IMHO the more interesting story with GPT 2 is the hype around it as well as the (huge) backlash around its release. [self-plug coming] If interested, check out this summary of that whole story: https://www.skynettoday.com/briefs/gpt2
People write fake news all of the time, why is this AI so "dangerous"? At worst it'll just add to an already existing mountain of news. At best I believe it will prompt people to be more critical of what they read and where they read it from, after all, how many times do you need to be embarrassed by some auto-generated news sources?
Spam. How on earth would a spam filter determine if such an auto-generated text is spam or ham? Since it doesn't have a human brain it can't. So we get into an arms race again. Right now, spam filters seem to have the upper hand but that can change very quickly.
You can also use this text generation for less nefarious purposes such as for submitting journal articles (https://pdos.csail.mit.edu/archive/scigen/), creating fake facebook campaigns, creating fake github profiles, phishing on dating sites, fool Google's page ranking algorithms and so on.
To be completely honest, a lot of these services need a wake up call anyway. Fake papers, fake Facebook campaigns, fake GitHub accounts, faking dating profiles and crappy search rankings are already a problem - but these services are currently unmotivated to solve these problems as they still work 99% of the time.
We are already in an arms race for people's attention and data, that horse already left the stable. Kindly asking people not to produce fake news is so far not working out for us. This AI will be replicated, likely very soon, if not already.
I don't think your argument is very convincing since this problem would affect almost every site on the internet and almost every text-based communication channel. If burglary increased 100-fold, I don't think many would see it as a wake-up call to install beefier home security systems.
> I don't think your argument is very convincing since this
> problem would affect almost every site on the internet and
> almost every text-based communication channel.
Every text based channel is already compromised, if not by AI, then by bad actors. The only difference is that it's in lower volume than it might be with AI. It's only a matter of time before somebody else recreates this AI (I'm relatively sure I could with 2 solid months and some motivation) and the longer we delay, potentially the greater the threat.
Consider this in terms of anti-virus software, do you respond to each incremental advance in virus programming or do you wait until the problem is overwhelmingly bad and you don't have years of incremental research to support yourself?
> If burglary increased 100-fold, I don't think many would
> see it as a wake-up call to install beefier home security
> systems.
Even if this increased spam 100-fold, this wouldn't happen overnight. But by locking it away, researchers can not actively work on a counter solution - so when somebody invents a better AI and releases it, they are extremely ill prepared. Not only this, the people themselves are unprepared too.
I think in delaying the handling this problem, the potential for mass disruption increases - not decreases.
55 comments
[ 6.1 ms ] story [ 109 ms ] threadIf trained on massive sets of source code and briefs (or perhaps a set of unit tests), could it spit out functioning code based on a problem statement?
I'm just not well versed enough with ML to know if these things are a possibility or if it will just remix its training data in probabilistically correct, not-guaranteed-to-make sense ways?
Maybe it could transcompile or re-implement common algorithms to already solved problems.
This had about 10 million parameters though, compared to OpenAi's 1.5B. I don't think their text is nonsense though. There's some very interesting examples. They have since released publicly a 345M parameter version, there are some nice examples from that model pre-trained with modern poetry on this twitter account: https://twitter.com/rossgoodwin
Microsoft's latest Visual Studio 2019 beta (v3) has intellicode[0] support which is supposed to ML "guess" 'code' based on equivalent github similarities. I don't think it is quite near to what you are asking, but certainly part of the way there.
0. https://visualstudio.microsoft.com/services/intellicode/
(autocomplete) 'Robert'); DROP TABLE students;--
It's better at making up bullshit, but it's still bullshit.
I am not an expert with machine learning, but it seems (at a very high level) we've done amazingly well at creating models that can recognize and sometimes recreate patterns. But they never seem to have any ability to understand the patterns. I'm sure someone much more knowledgeable could compare it to a child of whatever age (or maybe I'm just completely wrong).
Or, to put that another way, flipping the perspective around: our minds could consist of an intelligent, analytical, but utterly unimaginative agent, that sits there listening to a stream of suggestions spewed out by a distinct second agent, one that is "creative" but has no idea about the constraints of things like physics. The brain's analytical agent filters this stream of suggestions, taking notice of the suggestions that seem like they'll make the world change in the ways it "wants"†; and then it does those.
† Or, according to modern perceptual-control research, the agent attempts to predict the world that will occur a few seconds in the future, with a bias toward predicting world-states the reward-system has annotated as being rewarding; and then it looks at what motor commands it "would have" issued in that hypothetical world, and actually issues those. The stream of suggestions, in this model, serve as input to feed the generative model of potential world-states; the executive agent then must notice whether the potential world-state is a "possible" world or an "impossible" world (and whether the motor commands required of it are "possible" or "impossible" inputs), and filter out the "impossible" worlds.
Under this hypothesis, dreaming is the state when your executive responsible for filtering out "impossible" worlds isn't online. So you just get a continuous "impossible" world generated from the streamed suggestions of the creative-but-stupid bullshit-generating agent, with nothing to tear it down—just as seen in these generative AIs. As consciousness returns, the mental predicted world-state is noticed to be impossible by the now-online analytical agent, and is torn down.
That will be a lot of bullshit to filter out, if such agent doesn't provide more of less detailed description of what has to be generated. People with Broca's aphasia probably demonstrate a part of such input.
In other words, it could operate just as AI generative networks like GPT2 do, first receiving training input/output pairs; and then later, receiving input prompts and "completing" them by generating outputs.
When it comes to attacks, it's not the average-case which matters.
Some details are available here:
https://openai.com/blog/better-language-models/
I extremely rarely upvote stories.
I read the comments section more than i read the articles.
I cant be the only one.
[0] https://openai.com/blog/better-language-models/ [1] https://d4mucfpksywv.cloudfront.net/better-language-models/l...
“We made this super cool thing. It’s super dangerous. You can’t see it.” I mean, come on.
They can earn their reputation the regular way, and it will be fine. Sure they have the click baity DOTA demos, and maybe they are doing some interesting stuff. But OpenAI’s announcement of GPT-2 was particularly egregious.
But this is for a smaller version of GPT-2 (~400M parameters). Since the bigger model (2B parameters) was deemed too dangerous by OpenAI.
Hm, why in the world would they not train it on Sci-hub articles and generate output for which we have plenty of domain experts who can judge the quality?
* Training on Sci-hub or book torrents would probably get them sued, so that's right out.
* The point of training the model on reddit-linked-articles with 3+ karma (not reddit comments, as the article suggests) is to filter out worthless content (spam or non-text pages) while still getting a large, diverse sampling of human writing. They're training huge models and they need as much data as possible to do so.
* Every native english speaker is a "domain expert" for the purpose of evaluating the model's results. The point at which we need subject-matter experts to evaluate the quality of neural-net generated research papers is many decades away; the excitement around GPT-2 is that it can generate coherent English sentences at all.
That means no more skimming threads to get the gist of what people are saying, no more skimming through answers on stackoverflow, no more skimming through articles.
An article that looks reasonable on a cursory glance only begins to fall apart when you spend (waste) time reading it carefully.
It also means anyone posting content must spend extra effort proving they are a human for their readers.
What if you could beneficially weaponize an AI to teach children/people a vast body of knowledge quickly on any given subject.
Please see the movie; Lawnmower Man
It would be slightly more coherent to at least point to a potential use case for text generation specifically.
IMHO the more interesting story with GPT 2 is the hype around it as well as the (huge) backlash around its release. [self-plug coming] If interested, check out this summary of that whole story: https://www.skynettoday.com/briefs/gpt2
You can also use this text generation for less nefarious purposes such as for submitting journal articles (https://pdos.csail.mit.edu/archive/scigen/), creating fake facebook campaigns, creating fake github profiles, phishing on dating sites, fool Google's page ranking algorithms and so on.
We are already in an arms race for people's attention and data, that horse already left the stable. Kindly asking people not to produce fake news is so far not working out for us. This AI will be replicated, likely very soon, if not already.
> problem would affect almost every site on the internet and
> almost every text-based communication channel.
Every text based channel is already compromised, if not by AI, then by bad actors. The only difference is that it's in lower volume than it might be with AI. It's only a matter of time before somebody else recreates this AI (I'm relatively sure I could with 2 solid months and some motivation) and the longer we delay, potentially the greater the threat.
Consider this in terms of anti-virus software, do you respond to each incremental advance in virus programming or do you wait until the problem is overwhelmingly bad and you don't have years of incremental research to support yourself?
> If burglary increased 100-fold, I don't think many would
> see it as a wake-up call to install beefier home security
> systems.
Even if this increased spam 100-fold, this wouldn't happen overnight. But by locking it away, researchers can not actively work on a counter solution - so when somebody invents a better AI and releases it, they are extremely ill prepared. Not only this, the people themselves are unprepared too.
I think in delaying the handling this problem, the potential for mass disruption increases - not decreases.