61–70 of 87 posts
Maybe I’m too naive here, but I’m not seeing the potential malicious usage of this model. People will generate text, and then what?
It will accelerate the development of a fact checking AI, which means lying and manipulating people won't be so profitable any more, I guess.
Earlier quoted context omitted.
That’s not correct. What you are saying is that there is no plausible organized effort that could stop or slow the creation of signal processing models that will have pronounced negative impacts. The error is on two levels: you are using too much analogy with other technologies. And you are writing off the possibility of stopping ai when it’s still not clear that it can’t be stopped. This isn’t something that can be…
Correction: You are in a car that doesn't have any brakes.
Imagine a thing "we should not do", let it be creation of really powerful language model, genetic engineering in order to produce smarter, stronger children, or whatever. Whatever you are opposing to, really. The simple fact is that if you (and by "you" I mean any entity you associate yourself with, be it literally you, or your company, or a group of researchers in your country, or the government of your country) can do something, anybody can. Even if it will be a couple of years later. You are not unique, nor alone. You can stop "yourselves" as long as you want, but there are other people, companies, research groups, governments, and they don't give a fuck about what you think "we" should do.
So, in the end, the car really doesn't have any brakes. Even if it truly means the end of it all, it's just unfortunate, but really, really unavoidable.
Notably, the 345M model (1.5 GB on disk) is big enough that it's pushing the limits of conventional GPUs, and an alternative method of finetuning the model ( https://github.com/nshepperd/gpt-2/commit/47df6da611716b4826... ) has to be used to prevent the GPU from going OOM. I'm working on tools to streamline GPT-2 text generation: I'm currently porting the code above to gpt-2-simple ( https://github.com/minimaxir/gpt-…
I haven't heard of gradient checkpointing yet, thank you for the link! Do you know how it compares to gradient accumulation? The latter basically reduces the batch size, but takes the sum of multiple gradients before actually performing an update, thereby having the same effect as the original batch size. The generated titles are great! You can put them into hncynic ( https://github.com/leod/hncynic ) to get closer t…
If I had to compare them, I'd say that accumulation is about working on a minibatch datapoint by datapoint and faking being able to run an entire large minibatch in a single shot, while checkpointing is about working on a model layer by layer and faking being able to run an entire model in a single shot.
The problem with GPT-2-335M and why nshepperd had to mess with gradient checkpointing is that the GPT-2-335M model will literally not fit in your standard 11GB GPU (and from Twitter comments about people trying it on the new 16GB Google Colab instances, it's unclear if 16GB would be enough either!). You can't even run minibatch n=1. It doesn't fit. It OOMs.
The model itself is only a gigabyte or so, the problem is that the self-attention layers, when run, use up a huge amount of memory for their intermediate steps, which must be stored in order to trace everything backwards through each step for the backprop part of training.
(Right now I believe nshepperd's code punts on doing gradient accumulation simultaneous with gradient checkpointing, so we've just been reducing the learning rate, which is sort of similar to faking large minibatches with gradient accumulation.)
Fortunately, because the self-attention layers are so small and cheap to compute, they work well with gradient checkpointing. They're cheap to recompute on the fly, so it's more important to save memory and allow training at all. (This is also how OpenAI is training the Sparse Transformers which are enormous; they haven't said either way, but I assume this is how they trained the larger GPT-2s like the 1.5b parameter version, because I can't imagine what hardware would fit even a single GPT-2 1.5b without tricks.)
Earlier 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.
Earlier quoted context omitted.
This thing really does scare me and fascinate me at the same time. There's so many nefarious things you could do with this... Generate fake news. spam google. etc. One valuable use could be to generate comedy and parody. You could also make it to sabotage others too. You could set it lose on nazi forums and have them argue with bots constantly.
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…
Consider the issue of fake reviews: sometimes fake reviews are really obvious. Sometimes they aren't. Often the best way to pick out the fakes is to analyze all of the other reviews the user has written. That is going to become harder.
For very technical topics, where the reader comes with a strong background knowledge in that topic, picking out the fake material isn't too difficult. I suspect for the hazier things where the writers are more or less stating opinions, like politics, it is going to be incredibly difficult (for a human reader) to separate the bots from real people.
Earlier quoted context omitted.
> of a juicy acquisition OpenAI is a non-profit. They are not looking to get acquired.
Is it still a non-profit? I think they changed their structure to some fudged thing recently.
Earlier quoted context omitted.
I haven't heard of gradient checkpointing yet, thank you for the link! Do you know how it compares to gradient accumulation? The latter basically reduces the batch size, but takes the sum of multiple gradients before actually performing an update, thereby having the same effect as the original batch size. The generated titles are great! You can put them into hncynic ( https://github.com/leod/hncynic ) to get closer t…
Gradient accumulation and gradient checkpointing are orthogonal. You might want to use them simultaneously. If I had to compare them, I'd say that accumulation is about working on a minibatch datapoint by datapoint and faking being able to run an entire large minibatch in a single shot, while checkpointing is about working on a model layer by layer and faking being able to run an entire model in a single shot. The pr…
If I understand nshepperd's code correctly, it uses a constant and small learning rate. Do you know if this works better than the learning rate schedule that is usually used for Transformer models (https://www.tensorflow.org/alpha/tutorials/text/transformer_...)?
Earlier quoted context omitted.
Someone is going to invent this model sooner or later, simply because it is possible. There is not much sense in trying to stop it. We just have to adapt.
That’s not correct. What you are saying is that there is no plausible organized effort that could stop or slow the creation of signal processing models that will have pronounced negative impacts. The error is on two levels: you are using too much analogy with other technologies. And you are writing off the possibility of stopping ai when it’s still not clear that it can’t be stopped. This isn’t something that can be…
No. What I am saying is that it's impossible to centrally control the actions of 7.7 billion free humans. A lot of them will disagree with your position (and any other position as well).
By trying to "put a stop to it" in a central manner, you are only making it harder (but not impossible) for some subset to learn about this phenomenon, to improve on it and to understand its strengths and weaknesses.
I am unconvinced by your argument that we could reliably detect, much less stop, attempts to train a large and useful machine learning model.
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?
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 impl…
