Maybe a dumb question - how does the model which is trained to predict the next words answer questions, as shown in the reading comprehension example? Do you just feed it the question and watch it generate the answer, or is something else going on?
OpenAI releases larger GPT-2 model
51–60 of 87 posts
Re: OpenAI releases larger GPT-2 model
#52Earlier quoted context omitted.
I think people are over-worried about fake material. We've been able to say lies since we became human. We've been able to print lies since printing. We can already edit photos. People routinely make fake screenshots of messages to defame others. Newspapers sometimes print fake stories based on lies given by their sources or which they fail to vet. But provenance still exists to prove things, not how convincing it lo…
If evidence alone was enough, there wouldn’t be any debate about antivaxxers, climate change, the Iraq WMDs, the 9/11 and moon landing conspiracies, Holocaust denial, the correct way to bring peace to the Israel/Palestine conflict, the safety of being white in a Muslim community or being black around an American cop, Brexit, any of the Assange drama, or if Snowden was acting appropriately by leaking what he leaked. A…
To take an optimistic view, it may eventually teach people to more critically evaluate anything they read.
Re: OpenAI releases larger GPT-2 model
#53> Due to our concerns about malicious applications of the technology, we are not releasing the trained model. Has anyone said when they are intending to release the full model? There's likely a lot of positive applications of this technology as well.
They said within 6 months, however by that time there will likely be something better.
Releasing publicly AI tech that is six months behind the state of the art, might be a reasonable compromise against the fear of nefarious use. That way, more people know that the tech exists before it's available to everyone, and get a 6-month window to adapt.
At the same time, the risk is lower that a different team develops similar capabilities and widely deploy it without anyone even knowing that what they do is possible.
Wouldn't be surprised if this pattern is followed even more in the future. It's an acknowledgement that there are many angry cats that will (almost unstoppably) be let out of the bag in the next decades, but attempting at releasing the cats in the most responsible manner possible.
Re: OpenAI releases larger GPT-2 model
#54> Due to our concerns about malicious applications of the technology, we are not releasing the trained model. Has anyone said when they are intending to release the full model? There's likely a lot of positive applications of this technology as well.
They said within 6 months, however by that time there will likely be something better.
> They said within 6 months, however by that time there will likely be something better.
There has. BERT has come out and is better.
Re: OpenAI releases larger GPT-2 model
#55Maybe a dumb question - how does the model which is trained to predict the next words answer questions, as shown in the reading comprehension example? Do you just feed it the question and watch it generate the answer, or is something else going on?
If I remember correctly, they say that since the training set contains extracts of question-answer sessions, it will detect the pattern and follow it when you give an appropriate prompt.
So yes, you just feed the question and, detecting that it is a question, it answers.
Re: OpenAI releases larger GPT-2 model
#56Maybe I’m too naive here, but I’m not seeing the potential malicious usage of this model. People will generate text, and then what?
Especially what kind of usage that could not already be achieved by asking a human to write a text.
Re: OpenAI releases larger GPT-2 model
#57Worth noting that the new 345M model is still far from the full 1.5B model they were avoiding to post. The headline makes it seem like they finally decided to give the full model, but it's just a slightly larger demo model.
It strikes me as weird that they are not publishing it, by the way. According to their rhetoric when they started, it was the whole purpose of OpenAI: acknwoledging that we are at point when anybody with enough resources can produce something, let's say, interesting with ML, and striving to give everyone more or less equal possibilities by serving as a more effective academic organization for the world, before Facebook or Google takes over the world completely.
Plus, it's not as if this thing is more "potentially harmful" than, well... basically anything of use, like electricity, internet, fire, less perfect language models. In fact, it isn't even anything new, it's just (possibly) less broken language model than what we already have.
Admittedly, it would be quite problematic to use the full model with today's mainstream GPUs, so I'm not that much saddened by them hoarding on it. It just seems curious to me.
Re: OpenAI releases larger GPT-2 model
#58Earlier quoted context omitted.
They said within 6 months, however by that time there will likely be something better.
> They said within 6 months, however by that time there will likely be something better. There has. BERT has come out and is better.
Erm, isn't a big part of this whole story that GPT-2 > BERT -> GPT?
Re: OpenAI releases larger GPT-2 model
#59Maybe a dumb question - how does the model which is trained to predict the next words answer questions, as shown in the reading comprehension example? Do you just feed it the question and watch it generate the answer, or is something else going on?
you add a linear classifier at the top to predict start and end positions of the answer span. The augmented model is trained on a qa dataset like squad to actually learn how to answer questions.
hugging face has a simple implementation that augments bert in this manner and you can see the code there. their bertqa model get like an 84 F1 on squad 1.1 which really strong performance. you can augment the thier gpt2 implementation similarly.
Re: OpenAI releases larger GPT-2 model
#60Maybe I’m too naive here, but I’m not seeing the potential malicious usage of this model. People will generate text, and then what?
Especially what kind of usage that could not already be achieved by asking a human to write a text.
It makes it cheaper, faster, and effectively infinitely scalable. You'll run out of man hours and people to generate this stuff by hand than you can launching a ton of resources in the cloud.
With the main concerns being troll army/fake news type stuff, I don't think this makes a difference. We seem pretty sure there are state level actors behind a lot of that stuff, and I think it would be silly to believe they can't recreate something at the level of GPT-2, especially with the underlying principles out there and understood, competitors like BERT available, etc.
I think their heart is in the right place, but also incredibly naive.