OpenAI decided not to release the full scale version of their model (1.5B), because they were afraid of its potential security implications, in particular in generating fake news.
One thing I believe gets indequate recognition, including in this article, is OpenAI’s stated motivation for their release strategy. From the GPT2 release announcement[1]:
This decision, as well as our discussion of it, is an experiment: while we are not sure that it is the right decision today, we believe that the AI community will eventually need to tackle the issue of publication norms in a thoughtful way in certain research areas. Other disciplines such as biotechnology and cybersecurity have long had active debates about responsible publication in cases with clear misuse potential, and we hope that our experiment will serve as a case study for more nuanced discussions of model and code release decisions in the AI community.
It was, in part, due to direct safety concerns, but it was also to force the conversation about how to handle the release of legitimately dangerous capabilities in the future.
Big bump to this! After talking to OpenAI and MIRI I have seen that I had misunderstood some of OpenAI's message. I am updating my beliefs accordingly, but taking my time in doing so.
I would have accepted that from a publicly funded research lab. From a group funded by private interests with a very personal views of future and progress, I think it is fair to suspect ulterior motives.
On a side note, I'm still amazed that the NYT is able to uphold its curated public reputation so well that it is still consistently used as an example of an organization that would never dare publish something low-quality.
The general 1.5B model is not a threat to anything: it's the finetuned GPT-2 models and conditionally generated text from GPT-2 models that have been coming out lately which will take the next step.
Here's a few apropos outputs from a GPT-2 model I'm working on to conditionally generate programming article titles based on keywords "Java", "AI", and "death":
In the age of machine learning, is a human becoming too powerful? If so, what impact does AI have on how we view death?
An AI death machine created in Python
The death of an AI: What the world is looking to AI for now
Why Is AI So Bad At Killing And Exploring For People In Java?
Java is a death sentence – and AI is coming for all of us
How to avoid death by AI in Java
The AI is using Java to make it's own deathmatch
Google's "AI for death" is now a death sentence for Java
Why is Java a death sentence for AI?
Java is not a death sentence, but a step towards AI
The AI is dying... And it's not even on Java.
Go to Java to see the death of AI
Google's AI will stop every death before it's too late
A must watch: Attacking the "Dossier" is like attacking a sore thumb - bad for the country!
Why isn’t Hillary Clinton being investigated for her dealings with Russia-Russia? There is no reason why her corrupt ledger should still be closed. (and of COURSE!)
Totally illegal Donna Brazile calls Barack Obama "The Drilling America" on her Sunday morning television show.
Unbelievable evening in Albany New York last night! Disgraceful given the way Crooked Hillary is treated. There is no way this should be happening. She is a disaster!
So funny Bernie Sanders saying that the Democrats should go to a constitutional amendment, like our present system. They will never win if we continue toovertake!
Very sad that Crooked Hillary Clinton can't close the deal with Bernie Sanders. I can't imagine she has $10 million on hand! A vote for her is a vote for defeat.
Thank you Ed Klein! It is so great how you have done so much for the people of Vermont. So proud of you as their Governor. Now need your support to #MakeAmericaGreatAgain!
Wow Crooked Hillary Clinton has just indicated that she will not run in 2020 & that Bernie Sanders is a fraud! What does she have to hide!
Will be doing @TheIssue with @Greta at 9pm tonight
Heading to Albany New York for an interview with @CNN. Enjoy!
"@greta: Will @realDonaldTrump run? If he doesn't run and if we don't get a real developer and if we don't get a businessman that can get us the answers we seek
So many false and phony T.V. commercials being broadcast in New York State. Federal Election Commission rules prohibit "buying or selling of political ads."
I very much appreciate the many nice statements Mayor Stephanie Rawlings-Blake has made regarding the need to rebuild and restore trust. AARP currently representing VA
Unbelievable howarty Shore needed a rebuild just like the people of Virginia. So much corruption - big money and all. RAISE YOUR HANDS!
.@AlexSalmond Can you believe-you spent $1000000 to take this issue to the Scottish Parliament. You couldn't even get United Kingdom!
We are getting down to the wire on our Great American Infrastructure. Presidential Memorandum on #Infrastructure was well received and easy to understand.
Unbelievable howarty Shore needs a rebuild just like the people of Virginia. So much corruption - big money and all. RAISE YOUR HANDS!
We must stop the crime and human trafficking taking place on our streets. It is very sad that many more people are being robbed and killed by certain groups. We must get smart!
"@XKidd92: @realDonaldTrump help me Donald Trump I am so tired of reading about your so called "victories" !"
"@BeaumontAnthony: @realDonaldTrump you are the man Mr Trump!" Thanks.
"@BitcoinMoneyMan: Bitters night at the Trump Hotel in Chicago. Tin Roof is better than the day I became a citizen. #nice!"
"@TiniQBoy: @realDonaldTrump don't let it be said you were a great guest of Putin. You were a good friend. Let the truth be known!"
"@BitcoinMoneyMan: @realDonaldTrump I would call my daughter Ivanka that. She is a great person. She has a lot of class."
"@TWIAMundo: @realDonaldTrump you were a great role model in life. Make us all proud. We are all very proud of you!
"@JenaFeuers: @realDonaldTrump @MillionWit Yes! We are so happy you are coming back to Vegas! It is such an amazing place."
"@ChrisFoley_: Just set my DVR to record Celebrity Apprentice starting next Sunday night at 9 pm EST"
"@liamvanvorhis: @realDonaldTrump attending Canizal 2015 in Verona. En Charles
The published samples from the author's replicated GPT2-1.5M [1] look nothing like the ones from OpenAI's original publication. I really would have loved to see a student with no funding replicating a super large state of the art model, but unfortunately he must either have collected lower-quality data for training, or have failed to replicate every aspect of the model. Almost none of the generated samples have anything to do with the input text, and many of them don't even make any sense.
"Unlike in OpenAI's blogpost, I did absolutely nothing to manually cherry pick the quality here. Some examples are good, most are bad. Almost all outputs decay in quality as the post gets longer. I wanted to keep the raw outputs to give an accurate feel of the strengths and failure cases of the model. The truth is that using a GPT2 type model to create text you want is more an art than a science and can be very finicky."
The problem is OA also specified the level of cherrypicking, including sometimes picking the very first one. Do any of the conditional samples, picking out of 1 or 5, look like the OA blogpost ones? I didn't read them all, but they didn't look like it. Or much better than the 345M recently released, for that matter.
I don't know why it is assumed that eyballing text generated by a neural net can give an accurate measure of the network's quality. Evaluating this sort of output is difficult and there are no good metrics for it.
Then of course there is the fact that the network can generate text a lot faster than a human can read it. Think for a moment how many distinct passages these systems can generate -possibly even infinite- and how few of them a single person can realistically hope to read. It's very hard to know if the few passages one ends up reading are representative of the output of the network, or not.
It'd be even harder to compare two systems just by eyballing their output side-by-side.
You have to eyeball it because (a) 'what it looks like to a human eyeballing it' is the most important metric for misuse - no one cares that much about a model which has a slightly better log loss but a human can instantly spot is robospam, that can't be abused or used for fun projects like generating poetry; and (b) the log loss between GPT-2 and his will not be comparable due to the different training corpuses (and possibly architectural differences as well), so the actual metrics are not useful.
Given the large difference in quality, it only takes a few comparisons. Again, just look at the samples. Read a few. Do they really look the same in coherency and quality and realism?
Do you think that if you were given a sample of text genereated by one of the two systems you would be able to tell which system generated it?
Looking at two texts side-by-side while knowing which is which may well create an illusion of striking differences, which however are not really there. This is why concrete metrics are necessary- because human judgement is biased and inconsistent.
But, like you say- there are no good metrics. In fact, the only thing we have to go by when discussing Open AI's model in the first place is opinions and convictions. Not least the opinion of OpenAi that their model generates text of unprecedented coherence, which itself is not based on anything else than eyballing.
I think I could. Few or none of the new samples even manage to maintain coherency in a sample, and you can see all sorts of garbage prose being generated.
The author could have saved himself pages of posturing about "the biological blockchain"... These examples are crap. No way he's got something of equivalent potency to openAIs 1.5B model.
Not once in this blog did the author address why OpenAI has chosen to release the model in the way they have and why he disagrees with them, other than the belief that 'it's probably not possible to distinguish between GPT-2 and humans so let's not bother giving people a chance to try to develop techniques'. Blithely dismissing what OpenAI is trying to do is incredibly arrogant.
- "And I think no currently existing technique can even scratch the capabilities humans have in this area [detecting truth], even with all of their biases".
This is literally why OpenAI is releasing the model slowly.
- "I really do think that if people just knew that something like GPT2 exists and is out there in the wild, it would force them to improve their standards for what information they trust".
I disagree. Look at how the knowledge that 'shills' and 'russian trolls' exist has shaped online discussions. It becomes an easy way to dismiss things that challenge your worldview, but does not significantly improve their ability to determine veracity.
- "I think we have reached a point where it is no longer, in general, possible to determine whether a given text is human generated or not".
- "It means that even if we had a system that can perfectly detect AI generated babbling and deployed it, it would censor not just AI, but a good chunk of real, human communication as well".
Based on what? This is a very thoughtless dismissal of a very important (and very much unanswered) question. Deep learning often leaves identifiable artifacts, such as deep fakes not blinking. GPT-2 is impressive, but to assume that it is identical to human generated text is a lazy assumption.
- "I’m as anti-war, pro-human as you could imagine, but if I was alive in the 40s, and the US offered me to work on the Bomb…I don’t know if I could have said no. Not because I wanted to hurt people, this is the important thing to understand about the curious hacker, but because it was just so damn cool. Splitting the literal building blocks of matter itself to make a giant explosion? That’s fucking awesome!"
- "a digital 21st century Manhattan project (Which, again, sounds like heaven to me)".
There are no words to describe my disgust at this sentiment. I'm not anti-war, but I think prioritizing 'having fun' over human lives is evil.
These are some of the best counterarguments I've heard yet. Would you be interested in discussing this at more length over email or DM? I'm sorry I have done and said things you think are wrong. I am genuinely trying to do my best to do what is right.
I'll follow up on this when I have time, but I should say that, while I do stand by everything I said in my post, my tone is harsher than I would like for reasons unrelated to you (I'm in a bad mood).
Okay... I got hit with the shivers when the point came together at the end. Too much coffee today? Who knows. I think and read about this stuff quite seriously when I'm not burying my head in the sand about it. But this post got me on a spiritual level in a way that a lot of tech writing usually doesn't.
Aside, it's a very beautiful example of text that isn't babbling, rambling as it may be. Although now I wonder about a time when this level of elucidation might be arbitrarily generated in bulk by some ultra finely tuned blob of mathematical functions. Trained on human thoughts, of course. Maybe we're not all that far off.
I became acutely aware at various points while reading this of the powers of silence, listening, slang, poetry, signifying, innuendo, of knowing, of secret languages, of stories, curiosity, of music. Amongst others, these are deep gifts we have, and I take some comfort in remembering them.
My take: I'm fairly confident in saying I think this was written by an actual human.
I'm usually a little bashful about showing my music, but it was kinda uncanny how much I was reminded of an old song:
All the GPT2 output just seemed like spam text that you find on forums that don't have Akismet or something installed. Honestly, I've never found it either interesting or dangerous.
NPCollapse, do you have any metrics on the quality of your replication? Without any mention of metrics in either post, I'm skeptical that you've really replicated their results, which would render everything else rather moot. As you said, trust is hard.
With growing power and capabilities of individuals the only thing that really stands between us and disasters of any kind (physical, social, environmental, ...) is the utter incompetence and lack of creativity of bad actors.
(This is not meant to say that this guy is bad, in fact neither he nor the content of the article matter at all. The only necessary context is that a single person decided to do what a group of top-notch researchers considered as maybe a bad idea.)
Another argument the author may not have considered is that by prematurely releasing his implementation, he may be making it more (not less!) likely that future discoveries are hidden away from the public.
It seems OpenAI wanted to release this project in phases, allowing people time to adjust to its nature. If in the future an even more disruptive project is created (by OpenAI or others), if the creator feels they cannot release it in their own perception of what a “safe” way is, they may simply avoid publishing and instead privately communicate with companies and powerful individuals. Which I don’t think is the outcome the author wants. So I hope he reconsiders here.
> It seems OpenAI wanted to release this project in phases, allowing people time to adjust to its nature.
Do you really think anyone besides AI enthusiasts are paying attention? It's not like the general public is even aware of this, let alone following its progress.
I may have spoken too generally here. By “people” I meant e.g. engineers at Google, Facebook, Reddit, news outlets, that kind of thing. I see it a bit like a security disclosure.
I understand OpenAI is experimenting with their release of the GPT2 model, but I still don't understand their reasoning. If it's too dangerous to release today, what's going to change in the few months before they release it? They don't say why it's too dangerous beyond hand-waving, so it's impossible to be able to protect against that.
Security disclosures are much simpler - we found a vulnerability and we will provide time for the company/team/organization to patch it before announcing it to the world so it won't be exploited by bad actors.
If OpenAI truly feels they have something akin to nuclear weaponry, and that fewer actors having it is better, than they have to openly admit that they consider themselves better gatekeepers of the technology than the public and back away even further from their non-profit/limited profit ideals. "We are creating this technology for the good of the world, but it's so good we are afraid to let you use it, so only we will benefit from it."
I find them wildly inconsistent in their messaging, trying to have the best of everything with none of the drawbacks.
The model needs to be retrained from sctratch for different types of texts. One can release a model trained to generate Trump tweets, but it's of not much use for generating fake news on a specific topic.
Most of the examples don't rhyme. It's unclear to me if this is because most of the original poetry doesn't rhyme so it's just faithfully replicating the lack of rhyme, or if it only partially and accidentally grasps the idea of rhyme.
Some of the ones I like are 'We never say "Thank you"', 'Thy soul, thy very soul is burning!', '"It is morn!” said the clover-bush', 'And they have seen the last light fail', 'There comes a murmur low and sweet'.
Probably the best IMO is 'The sun is gone, and the night is late', but of course everyone will have a different favorite.
Yes, "The sun is gone..." starts out amazingly well. But later fixates on tides for some reason :)
Everything is generated by the 117M model, correct? If so, do you expect the quality to improve for larger models, or is there not enough poetry to train them on? I wonder how much of total poetry is contained in Gutenberg poetry corpus...
It's a mix of OA 117M and 345M at the moment. I haven't observed too much in the way of overfitting yet, so there should still be benefits to going up another 4.4x in model size to 1.5B. My guess is that at 1.5B, it'll start being more important to improve the poetry corpus, since you can already start to see problems with it - the Alexander Pope brokenness and the occasional prose generation of footnotes/commentary are definitely undesirable, and I suspect there would be less 'run on' effect in samples if the original corpus actually properly marked '<|endoftext|>' for each poem...
Maybe I don't understand something about these models. If the model was trained to mimic Trump tweets, it means that someone spent days of GPU time to find the weights of the model. Now if we want it to mimic HN comments, we'd need to spend the same amount of GPU time to find different weights. This is what I meant by "from scratch".
> ... if we want it to mimic HN comments, we'd need to spend the same amount of GPU time ...
These models are often much more general than you seem to be thinking. There's a base model which is incredibly computationally expensive to create from scratch. It is trained on a very large, very general set of data. Then there are specialized versions which are much cheaper to create - you start from the base model that you already have, and you train (much more briefly) on a specific set of data in order to tailor the output.
> Modern image recognition models have millions of parameters. Training them from scratch requires a lot of labeled training data and a lot of computing power (hundreds of GPU-hours or more). Transfer learning is a technique that shortcuts much of this by taking a piece of a model that has already been trained on a related task and reusing it in a new model.
Seconded. From the author's linked blogpost about the model, it seems like they've trained a similar analogous model, but I see no reference to metrics that might suggest it's really of the same quality.
I want to warn anybody looking forward to play with his 1.5G model, that there's no confirmation yet, that this model actually beats 345M one from OpenAI. Connor had to come up with his own training procedure, which might have led to a worse (or better) result.
> But why do we trust the New York Times? Because the New York Times is composed of humans using their brains to do exactly what these detection algorithms try to do
This reasoning sounds circular and doesn't go deep enough. Couldn't NYT or individual reporters at NYT abuse our trust on rare occasions if they felt the stakes were high? Couldn't they unintentionally mislead us because they were mislead by a source? No: we trust NYT because they have a financial interest in being a reliable source of news. People wouldn't read NYT anymore if they thought that they didn't take that responsibility seriously.
What I like about this article (despite slightly too much babbling) is a CS undergrad using lessons from Psychology and History to shape his thinking and conclusions.
This was not happening 10 years back.
In my experience the majority of 30 and 40 year olds in Tech today pushing AI code out have no clue who Kahneman, Hariri and Pinker are.
So expect subprime meltdowns/Trump/Brexit type unintended bullshit for a few more years till the psychologist+historian+sociologist coders take over.
People have found issues with Harari and Kahneman too. That's okay. But who else are you going to read to get a sense of these subjects?
People attacked Faraday for suggesting electromagnetism was a thing, hardly paid attention to Maxwell for doing the Math because it was too complicated, and attacked Oliver Heaviside his whole life, for simplifying the math that everyone uses today. Each of these characters made mistakes and had weaknesses too and its easy to find fault.
That's the way things work. There are always going to be more people who fixate/react to mistakes, than use them as stepping stones to new discoveries.
Without these writers I would have hardly any awareness about these subjects, because these subjects were hardly discussed when I was in school. And it's not even that long ago. They opened the door. History will show that was the role they played.
I disliked him by the second paragraph. What a monument he built there, to whom he thinks is a most awesome and impressive person. Sadly he thinks it's himself.
OpenAI decided not to release the full scale version of their model (1.5B), because they were afraid of its potential security implications, in particular in generating fake news.
One thing I believe gets indequate recognition, including in this article, is OpenAI’s stated motivation for their release strategy. From the GPT2 release announcement[1]:
This decision, as well as our discussion of it, is an experiment: while we are not sure that it is the right decision today, we believe that the AI community will eventually need to tackle the issue of publication norms in a thoughtful way in certain research areas. Other disciplines such as biotechnology and cybersecurity have long had active debates about responsible publication in cases with clear misuse potential, and we hope that our experiment will serve as a case study for more nuanced discussions of model and code release decisions in the AI community.
It was, in part, due to direct safety concerns, but it was also to force the conversation about how to handle the release of legitimately dangerous capabilities in the future.
[1] - https://openai.com/blog/better-language-models/