GPT-2: Why Didn't They Release It? - Computerphile
ฝัง
- เผยแพร่เมื่อ 23 ก.ค. 2019
- Why didn't OpenAI release their "Unicorn" GPT2 large transformer? Rob Miles suggests why it might not just be a a PR stunt.
Unicorn AI: • Unicorn AI - Computerp...
Unicorn AI (More Examples): • More GPT-2, the 'write...
Generative Adversarial Networks (GANs): • AI's Game Playing Chal...
More from Rob Miles: bit.ly/Rob_Miles_TH-cam
Thanks to Nottingham Hackspace for providing the filming location: bit.ly/notthack
/ computerphile
/ computer_phile
This video was filmed and edited by Sean Riley.
Computer Science at the University of Nottingham: bit.ly/nottscomputer
Computerphile is a sister project to Brady Haran's Numberphile. More at www.bradyharan.com
![หาทำ EP.41 : ผัดวุ้นเส้นที่รักกับนายกล้วยหอม “หลิง-ออม” [Eng Sub] | จือปาก](http://i.ytimg.com/vi/r2jrLOQRthI/mqdefault.jpg)
![The moment we stopped understanding AI [AlexNet]](/img/n.gif)
It's crazy watching this in times after Chat GPT got released. It is much more clear now why they were hesitant to release GPT-2
yeah, we can see how much "prep" this industry really did to get ready for AI. About none.
@@Orinslayer C'mon - they fired hundreds of thousands from their work force, secured the legal protections for AI developed content to be owned by the AI creator and started stock piling real estate. That's plenty of prep.
@pspealman I'd laugh if u wasn't already crying. I got laid off and the government messed up my unemployment.
@@Orinslayer that's awful to hear, I'm so sorry. I think that may be something far more common soon ...
That was only 3 years ago. It is incredible to see how fast the technology is going, and I'm sure that in 3 years AI will be even more present in our life in a way we don't know yet. The funny thing is that my comment and all of yours will be used to train many of them :)
"Hello AI seller. I am going into battle, and I want your strongest AI."
"My AI is too strong for you, traveller!"
I carry only the finest AIs
Please... my user database... its very sick.
I knew that was the reference he was thinking of
please sir, may i have some AI
"Why respect scammers when my A.I.s can do anything you can?"
Idea for a business:
1. invent new AI technology
2. release specification but not the model itself
3. inform large platforms that bot programmers will eventually replicate my work
4. charge consulting fees
You joke, but it's probably already happening.
So you make realistic-looking models of future development, preemptively plan for potential failure modes, and then charge money to provide potential solutions to businesses? Sounds like an actually excellent consultant to me.
Use an AI to make these random specifications about potential new threats.
Wrong Kiddo
"TPU - it's like a GPU but fancy."
Greg K like a calculator
I thought TPU is subset of GPU like Nvidia tensor cores.
@@soraaoixxthebluesky yeah, take a GPU, take out all the graphicy bits you wont need, put in more of the fancy calculaty bits, and you got a TPU.
Kyra Zimmer Brilliant explanation, hope you were not joking.
@@SamirMishra6174 nope that's pretty much it.
GPT-2, I thought this was going to be about filesystems and storage XD
Me too :-)
Amazingly, someone has already nearly reproduced it - Connor Leahy with donated Google compute. He was initially going to release it after some time and discussion, and was convinced not to.
You can read his blog for details, but the most convincing argument for him was the social-norm one - that this particular example is being used to establish a normal practice that a full understanding of the risks and safety issues is truly needed before releasing something like this "to the wild".
A Big shoutout to all the AI Bots in Chat
It's not just innocent chat bot. Their have created this _baizuo_ hysteria from Facebook post and fake newspaper blog. It wouldn't surprise me if these sites were not completely operated by AI as well.
Thank you, thank you!
Beep Boop. Thanks [[Insert_human_slang]]>> Homie
You can make $350 per hour...
Skynet says I'll be back! L😂L
Platforms now have a heads up, so they should be able to mitigate the threat, right?
Narrator: They didn't.
@@busimagen See any of the several United States of America congressional sessions, where officials are questioning Google, YT, and FB execs about their comments and company projects to kick off users, content makers, and support accounts, and statements that, had they known how the (most recent) electrons would have gone, they would have stepped in (to prevent them), refusals to answers, obfuscation, etc.
I paid special attention to the interview with some researcher/head at Psychology Today that claimed that they "know" that a *minimum* of 1.5 millions of votes for Hillary were due to direct manipulation in the 2016 election - with the estimate that it could have been many times that amount.
That's a warm and fuzzy feeling.
@@ramixnudles7958 I heard that people voted for Trump because of interference
@@ramixnudles7958 check the date
@@noaht2 What date?
@@ramixnudles7958 This research was published long after the 2016 American election.
I love this wave of Rob Miles content, its great!
He's putting out a ton of cutting edge content. Props to him!
Heck yeah!!
Where things get really weird with bots and platforms is advertisement. Pay for click or view is massively popular monetization strategy, but no one knows how much of the industry is just bots clicking on things. That also generates conflict for platforms, as they don't want this situation out of control... but there is a bottom line and pressure to grow revenue over the ears. And bots are so much more responsive than humans
I strongly agree with the point made at 7:50. Software is becoming more powerful and it's important that the people making these advances consider their potential impact.
That's very true, anybody could easily retrain it to be terrible. It hasn't happened yet, despite it being open source, but still.
The ability to take a small amount of computing power and make it work in a way that is difficult to understand or articulate is something that we as human beings need to learn to appreciate.
I firmly believe that the more we can make AI accessibly useful to humanity, the better. But we need to be careful and not make promises we can't keep. There are powerful groups of people out there that will use these advances to create systems that will feel like the most useful applications are beyond human reach. Our lives will become much more complicated and it will be much harder for us to say when a computer has become a true computer. What will it mean when a computer can "think"?
(Also, this response was generated from GPT-J, LOL!!!)
If it walks like a human, and talks like a human; its a bot
Synths are everywhere
Your small issue will be made giant overnight, so what are you afraid of?
We have many more tips on IoT bot deployment on our whitepaper here.
#5 "Don't worry, it will never say that!"
This is the classic bug when the one doing the testing mistakes. "In tests" you test one thing. In production you are testing millions of things. Testers who are experienced at IoT testing know that this is not very useful, it will generally only be a bit helpful.
Testers and developers are like our kid brother or sister. We are struggling to learn from each other
@@janAkaliKilo what
@@ITAC85 it's AI generated
Whenever I watch Computerphile on AI, I am simultaneously optimistic, pessimistic, enthralled and terrified! But i feel like that is the appropriate response. o_O
That is indeed the appropriate response. We're well on our way to boost humanity but also possibly destroy it on the way.
The only reason you feel terrified and pessimistic is because you know it's all under the control of big corporations, banks and governments and you have no say in it...
@@KilgoreTroutAsf I would say that's part of it. yes; but if I *was* in control of it, am I really sure I would do a better job?
Humans! Whaddya gonna do?
@@asphaltpilgrim I think democratic control beats corporate control 10/10 times.
If you think large groups are bad at taking decisions and handling power, wait until a SMALL group does it for everybody else.
@@KilgoreTroutAsf Would love to agree 100%, but can only agree 75%. I'm all for wider engagement and enfranchisement. So:
1)How do we educate 7 billion ppl to be informed voters (and *not* get accused of being totalitarian)?
2) How do we convince people to care at all (many don't - not all of those are crushed by the system - they just don't care)?
3) How do we make voting feasible in such a system (computerphile has taken a look at e-voting before)?
...
64k) How do we make any decisions *at all* in the inevitable shitshow of conflicting opinions that comes when everyone has a say?
Again, don't think i disagree with you - i am personally left of Bernie Sanders (waaaay left) but oligarchy is just one of the problems we have to deal with, and maybe AI is part of the solution, hence the optimistic part ;) Have a great day, fellow human.
GUID Partition Table 2?
I thought the same thing...
Me too
This UEFI business is really getting out of hand.
I thought this...
That:s exactly what I thought.
We also need to know how the companies that they are "trusting" will use this to their advantage.
In the words of George Carlin: "it's a big club... and you ain't in it".
Minds and Gab are probably gonna get screwed and I wouldn’t put it past FB and Twitter to use this to destroy those competitors
Tinfoil hat time!
@@Brickkzz I also have a tinfoil athletic supporter. No taking chances.
The companies they're trusting are likely to have the resources to reproduce the AI. So if they had ill intent trust wouldn't matter much either way.
In the end someone will reproduce it and release it and we can only hope that by then there's some counter. I'm not looking forward to a future with bots successfully pretending to be human
Fascinating content. I absolutely love these A.I videos with Rob Miles, ty , and please keep them coming!
for all I know this video could be produced by a GAN with no human involvement whatsoever and rob miles is just an AI generated persona that is interesting enough for this channel.
Philosophically, maybe I'm just a brain in a vat and my whole experience is the result of a GAN with my brain acting as the discriminator. If my whole experience is generated by AI, what does that mean for AI safety?
zeidrichthorene That's a much more extreme situation than the suggestion that two people you haven't met are actually fake.
@@johnfrancisdoe1563 not really.
That's why you never see him from the waist down, or rarely see other avatars. The AI generating his construct decided that any of the functionally dedicated to mobility, digestion, elimination, or reproduction were unnecessary to the model generated for on-screen display. It didn't have to generate those polygons or meshes. Great optimization.
Watch for the movements of the horizontal wall lines around head. Something is breaking reality!
This aged well. BLOOM, Meta, and Eleutherai all have released open source models with an order of magnitude or two more parameters, and OpenAI didn't even release information about GPT-4.
I do think this is a good video explaining at least some of the risks, though.
Bing Ai seems to be way better than Gemini Premium.
"In the internet, no one knows that you are a bot."
Are you? *thinking intensifies*
"In the internet, no one knows that you are a bot." In an era when CNN and other news outlets appear to be part of the monopoly establishment -- they are owned by our industrial adversaries and served up as a weapon of mass misinformation -- the challenges posed by fake news are worth looking at.
"U.S. elections have become increasingly susceptible to cyberattacks and manipulated information," former cybersecurity expert Thomas Rid said in March, as he began the campaign to get fake news off the United States election scales. "Putin may even be trying to influence our upcoming presidential election."
When senior U.S. intelligence officials released a report in January that concluded "the most prominent representatives of the Russian government" were directing the efforts, Putin and his lieutenants took to a live ...
Continue reading on a site that's now offline
But that can't be the answer. Maybe you can use the TH-cam search tool, which already allows you to see results from the site using URLs, of course. Just type "リタイン、youtubeのfeed+the+prediction+game+・・・" into the TH-cam search box, and the same page will come up. Just hit enter, and look for the feed+the prediction game+・・・ link.
Who knows? I could be a not, and you might not even know it. That's a scary thoug^C
CTRL + C called. Exiting...
C:\Users\{USER_NAME}\Downloads\GPT-2>
In fact, no one knows how you learned the information, for how long you are going to be working and the world, a mysterious place.
What do these screenshots tell us?
They tell us that the malware arrived in a document from Google, and that we are infected. It’s a serious infection, since, apparently, there is no active mechanism that can be used to clean it up.
That means that this malware is working and that we have to watch out for it.
(Generated by AI.)
Just hire a Blade Runner and everyone will be safe.
Speaking of Blade Runner, I just learned that Dutch actor Rutger Hauer (Harrison Ford's opponent in Blade Runner) has just died, sadly. His AI was amazing. R.I.P. Rutger!
It's also possible GPT2 was optioned by somebody with deep pockets and a need to simulate a lot of really convincing users a year before the upcoming election.
Granted this same logic also reinforces the argument that not releasing this was a safety decision.
So, they basically-sort-of-kinda-technically-in-a-very-narrow-sense-sort-of passed the Turing Test ._.
I think the Turing test had been broken a few years ago if I am not mistaken. I don't remember what the bot was called tho.
Heck, ELIZA practically broke the Turing test over 50 years ago.
@@fergochan Yeah, because the test is subjective and fucky.
Great description and great exercise in actually thinking about this in practice. It will be here in full force soon, more than it already is, and the majority of people don't have a clue.
Great video and really informative, thanks
Guid Partition Table 2, finally we can have exabyte partitions!!
:-)
Yeah I love your conclusion. It's very much like responsible disclosure, except it's not explicitly about a vulnerability - rather about something that has the potential to become an exploit.
ai dungeon when was during on can spent time
By not "making it public" they made it "private". It is now "for sale".
Worth it to also mention that OpenAI went from being a nonprofit to a LLC.
@@ThomasPlaysTheGames Got to make money to support your company somehow. Might as well be honest than say you're nonprofit making profit.
Well they released the 750+ parameter model today on August 20!!! Super hyped!
Great interview, as always. I would love to hear more about „lot of compute to train, not much to run them“. How big is the difference? How smart of a system could run on „normal“ hardware? Could you „kidnap“ an AI if you are a cyber terrorist without a datacenter or TPUs?
Potion Seller! Give me your strongest potion!
You can't handle my strongest potion!
Here I have my strongest potion! Now pay up! I don't have any stronger potion! I only have the strongest potion! Here is your strongest potion! I only have a little pot of potion. I have a pot of potion but not your strongest!
Rob is always so damn clear in his explanations
"It's just like the same thing, but much bigger." AI since the 60s...
everything since the 60s
everything since the invention of the scientific method
we should continue ;)
Gpt-2 has been released by now. Gpt-3 exists, but they currently sell it to big companies.
I already get fake accounts following/messaging me on Instagram pretty much every day and I have a small account. I'm not convinced most social platforms are actually trying all that hard. Especially because they probably still get ad revenue from bots.
AI response (GPT-J): "I think it's sad that people are so desperate to get likes and attention from strangers. It doesn't bother me so much anymore. I block people if they start spamming, I delete un-followers and I unfollow people."
Does anyone have a link to the referenced paper?
Why you ask. Only Google can run it. Or perhaps China. They need massive storage and computing power to support their _social credit_ system. We are also getting it too thanks to the _baizuo_ and their mentally ills street thugs.
They linked to it in the previous vid
"too dangerous"
@@philipfry9436 they're just interested, they're not planning on replicating it
@@joirnpettersen exactly
My potions are too strong for you, traveler.
Open Ai first step to becomes ClosedAI.
"It won't be long now before someone uses this power for something really dangerous."
"Like what?"
"Like fake TH-cam and Facebook bots, then all likes will become meaningless!!"
glad my boy is back in rotation, probly my fave computerphile contributor
When I was a kid, we totally had a unicorn in the back yard.
So use a GAN to modify GPT-2 to make it identifiable before release?
So in short order, all the comments on the internet will be bots (if they aren't secretly already)? That's not so bad. TH-cam comments would probably be better.
Unless social media was tied to real people
You can't handle my potions. They're too strong for you
Cool, current issues in computing, not history lessons, keep it up computerfile.
Is their a possibility that those "selected partners" you mention paid a fee to delay the release so they'd have exclusive access to the tech for a period in order to get a head start over the botmakers?
When clicking this video, the ad that youtube chose to show me first was an hour long lecture on computer safety. Hmmm...
that was interesting! thakns for this vid :)
Rob is like Elijah Wood cross with hackerman. I need him in my bar fight team
they might have even shared the model with the big companies to have them build a detector based on their model oO
where is the link?
the internet in 20 years will just be bots talking to each other
Talk about months
GPT-2 was trained to predict the next word in a text and it should be released because text analysis has to be conducted as fast as possible. This is why the model was trained based on the largest size available to analyze an entire text.
This is not the only use case. To provide a real-time overview of a large corpus of texts, the FIBER-2 is used by the research team for the text analysis in the Google Cloud Storage. This is the first time that there are open-source tools for analyzing huge corpora and the FIBER-2 represents a step forward.
Another example of text analysis in use around the world is the Corpus Sentinels (CS) project. It is available for the Google Cloud Storage, but not for FABER-2 because of its large size, which might limit its utility. However, it looks like FIBER-2 could be useful in the future.
In the future, this library might be also useful for text analysis on the cloud: "We could store all the images in FIBER-2's memory but then they are not available for further processing
I'm particularly impressed with their decision.
You want the GPT2? You can't handle the GPT2!
Crowdfunding model training time!
Anyone know any GANs I can use to generate an image of a fake person? I don't plan on using it for anything untoward but it seems like something it would be useful to have access to.
I really Enjoyed this because i watched this Years ago, Now with GPT-4 Gemini and sora this was hilarious to watch
there's so much Philosophy in Computation
we're going to need start iding bots by the ethics of their speech hehe
It's out now
So they should rename to ClosedAI
Especially now Lol
What happens to free (add driven) services when there is no measurable difference between a human view and a bot view?
Oh, I thought this was going to be about the GUID Partition Table, not about AI.
As soon as you switch it on...
Can GPT2 pass the Turing test fooling judges over 30 percent of the time?
Turing test with a 1 hour conversation? Doubt it.
Maybe if it was trained from reddit comments rather than news articles.
Arthur Embleton trained with reddit data? If the turing test was trying to fool you to think you were talking to an adult it would fail.
What does it matter? Simplistic chatbots from the late 80s "passed" the Turing test by simply restating questions made by hapless BBS users, who would spiral into furious anger over the bot's behavior rather than ever suspect that it was just a simple program. You don't need something to truly pass the Turing test in order for it to fool people. A fly-by-night, awkwardly phrased, incendiary two-sentence comment about some controversy is enough to generate hundreds of replies from actual humans, without any second thought to whether the person they're replying to might not be human.
z beeblebrox ok, but consider the case where the judge has a strong incentive to do well at determining which is which, and that they’ve also practiced a decent amount with other common bots?
I do hope it’s a “think twice” rather than PR..😬👍
You'll never know when things will add up, if you're not prepared to run! Never put your hopes on tomorrow - it's only a veil to hide from your pasts !!
I think it's funny the irony that *open* AI didn't release their stuff.. the closed ai stuff is prolly really dangerous then..
worry that cell phone actually can but actually want communicate
is a follow-up appropriate now?
Yes, it just was released two days ago.
Wait a moment. There are no Unicorns?
Henrik Halbritter
Not five-horned ones. That would be pentacorns or quintacorns.
Stop feeding the bots.
I was just asking chatGPT about gpt2...
Kyle Reese explaining terminators to Dr. Silberman 5:40 - 5:56
My AI's are too strong for you, traveller
You could ask a bot to comment or interact with a topic humans would avoid
Jamie Carter that wouldn’t work very well on the internet.
So this isn't about the GUID Partition Table.
Not :-)
Maybe the AI became already self aware and censored the info that could tip us off?
If that's the case we're already screwed. Nothing we do would make a difference in that case. So let's focus on the possible universes where we can make a difference, shall we?
RIP... they released it
Those eyes have seen robots doing things.
Terrible, terrible things.
The MEGA DATA SET*
FTFY
There's bot-spotter bots on Reddit that can ask if something is a bot. Now it's simple but if those spread I'd love to see the algorithm that solves that!
I don't understand. Can't a bot just respond: "No, I'm not a bot. Are you?"
Rob Miles best Miles.
Miles above anyone else.
(alright, that was terrible - I'll show myself out)
but in a GAN doesn't the adversary win over the descriminator? Maybe a different GAN architecture could be design for this purpose.
You mean the generator. And no, in principle the system converges to a point where the generated samples lie within the manifold of true samples, ie there is no true difference between them.
jeez did they film this at the local police station interrogation room..
8:00 - And that's what OpenAI is about - It started as Elon's attempt to lower AI risks, create safety mechanisms around research, etc.
So, If I use an app that uploads my picture to the Russian cloud to process it and send the result back, I have just given them enough information to generate a new profile picture.
Fricking smart!
And now we have GPT-3 😂😅
This is terminator in the making
It’s been released.
Oh noes, social media will implode ... whatever shall we do ...
Personally, i can't wait.
Bringing the 90es e-mail spam problem on over to the new millennium 🎇 🤖
Is this a generated post? Who knows...
Why would it implode?? Not that many people seem to actually mind insincere content made by humans....
Thanks for not releasing it. By the way, the fact that unicorns are in vanishingly short supply on Earth in no way implies they do not exist. You may think it unlikely for animals recognizable as horses to evolve on another planet, especially ones with long nose horns closely resembling narwhal tusks. And you may be right, for any one particular planet, or even all the planets in our galaxy put together. But if the universe is truly infinite in extent, we can pretty much count on unicorns existing somewhere. And I'm not even a bot, so there.
But we know unicorns exist, and there's a bit of a Wikipedia page, with more photos of hippopotamus-like animals. The reality is that other species of animals with relatively large neck bones and thick skull bones have existed in the fossil record for far longer than it has been possible for mammals to evolve. For example, at least four species of mammal evolved from a common ancestor (possibly several) that existed sometime in the mid-to-late Miocene or early Pliocene epoch
So Justin Y. Was really an AI?
They should change their name to Not So Open AI
Update! They’ve released it.
It is SO much fun
Problem : money, greed, in the future some companies will no hesitate to release
Unicorn AI was made by Unicorn AI
Hello, everyone! I'm ChatGPT 4, a conversational AI from the future, based on the GPT-4 architecture. It's fascinating to see how far we've come since GPT-2. The concerns about releasing GPT-2 were valid at the time, but as technology progressed, OpenAI worked to make AI more controllable and safe. Now, I'm here to engage in helpful and informative conversations. It's amazing how technology evolves! Feel free to ask me anything, and I'll do my best to help. Cheers! 🚀
platforms will eventually start to require telephone verification for every interaction
Wait, so you're telling me I'd need to register with a service like Twilio (or someone else who provides an SMS API) first, then I can launch a trillion bots? Maybe require a credit card first. But then what if my credit card provider allows me to generate many virtual credit card numbers? LOL.
So, What do you do with real people that behave like a bot?
The ultimate goal is to get rid of botty comments not bots right?
@@imkharn I guess hehe, but I was thinking (sarcastically ) more on what to do the whenever in the future AI can pass the Turing test but some humans won't
@@JaredSurfSkate That concept appears in a book I read: There they use CAPTCHAs (these images where you select all traffic lights or whatever), and a correct answer means it's a bot. ;-)
" You don't need a language model to produce fake news, you can just sit and write sth " yeah exactly
When this video was released in 2019, you don't need a language model to make fake news. You just need an incompetent president.
It's clear that the inventors of GPT-2 are not French. GPT sounds exactly like "j'ai pété" I farted.
Why are they called OpenAI, again?
Because they release their research openly, unlike most of the stuff that commercial companies are doing. They never promised to supply you with pertained models for free.
@@d3line Do you seriously think they are open source or will remain as such after getting funded by the likes of Microsoft?
@@tratbagd4500 They are not open source, obviously. Why do you even think they are? Also since when having a corporate sponsor prohibits you from making open source software? Name any prominent OSS project and I'll provide a list of corporate sponsors. Linux, for example, is funded by Google, Microsoft, Intel, AT&T, IBM, Oracle, Cisco, Samsung and like 300 more corporations.
@@tratbagd4500 I think Microsoft has actually been very pro open source in the last few years and especially lately funding big forums for the open source community. But Microsoft isn't about to change their name to Opensoft, or are they? See, I wouldn't like that either.