OpenAI's GPT-3 may be the biggest thing since Bitcoin

It depends. At the moment if I see a reddit post saying product X was really appreciated by a user, most of the time I'll believe that was an actual human appreciating that product. But if modern mass marketing is going to be the injection of seemingly sincere product recommendations into reddit threads that will obviously lose value - the bot comment is lower value than the human comment, and because I can't distinguish them all such comments lose value. Similarly for political statements of support.

I'm sure there is potential for extremely useful bots (e.g. such as article summarization bots on reddit) which increase information. I guess it really depends on who decides to set up a bot, and their goals and implementation.

Many people have no clue that automation has come this far and will judge every comment they read online as sincere. If they're actually not, and many are driven by political and commercial agendas, I think that's a bit dangerous, because people will act on them.

My understanding is that GPT-2 was actually trained on a dataset that was designed to avoid those pitfalls. They followed all the links posted to Reddit that had more than a couple karma, under the theory that the content was at least slightly interesting to some actual humans, as opposed to a giant blob of search keywords or what have you.

More or less. I have tools I wrote that takes sampling settings and a string and dispatches it across a cluster of machines. Via cluster-ssh I see the each of them expanding the text in a different random sampling.

This is custom stuff I've wrote that is tied to my environment. If you'd just like to play around, I can recommend https://bellard.org/nncp/gpt2tc.html in text generation mode as being extremely easy to get going. (the default model however is pretty small and dumb). The paradigm you're thinking about is exactly what you get from gpt2tc.

At any point I can abort a job, tweak the sampling settings, or the text I'm expanding. E.g. one operation is that if I see one sample seems to be on a good path but has made an error, I'll abort it and restart all of them from a fixed version of that sample.

Often I'll end up with my prompts in text files because they get a bit long at times, also escaping quotes and linebreaks on the commandline can be a pita.

I have some aspirations of integrating this into a text editor, so that as I type future text is just appearing ahead of me and I can just hit a cursor to accept parts of it. But in my experience GPT2 isn't good enough where looking only at one continuation is enough or where I don't get a lot of advantage in having it work from modified text.

GPT2 has preconceived notions about what kind of text you're writing based on the words you use. So it can be useful to alter your input text to replace persons/places/things names with different ones that get in into the right context and then back substitute them.

To give a concrete example, if I wanted GPT2 to show me example bio blurbs for my partner (always a pain to write but easier if someone generates examples), it works better if I change her name--Kat-- because it either turns her into a man or it resists talking about her being a lawyer and a board member and instead makes her into an artist or a dancer.

One thing to watch out for is that when GPT2 makes a benign error, like switching the gender of a pronoun mid-stream it often trashes the quality of the later output in unexpected ways (like causing it to output nonsense). Changing my SO's name to something it's not unsure about saves me time having to abort completions that have gone off the rails.

Perhaps I shouldn't use a gendered example. GPT2 isn't sexist its everything-ist. Every word has 1001 hidden meanings that subtly bias its behaviour, many of the biases are actually the point- they exist in the world and they're what makes the output useful-- others are weird and unexpected and are just a training/corpus artifact. Good use requires a degree of anticipating and exploiting these biases.

I even have to change my own name, because GPT2 knows very well that "Greg Maxwell" has something to do with Bitcoin and it readily lapses into Bitcoin conspiracy theories if my name is used. ["Gessh, even the machines are harassing me!"]

Another class of biases is language that trips it into fantasy land or 'silly' writing. Any word that is commonly used in writing for children is at risk. If there is another common word that only adults use it might be a better choice if you want serious text, or you setup a context that makes the meaning more clear.

It looks like for GPT3 you don't have to use subtle hints as much. E.g. that you can just tell it more explicitly what kind of thing you're doing and won't get tripped up as much by spurious correlations.

OTOH, since it doesn't look like they're going to release the GPT3 model it's likely that I'll never get an opportunity to use it in my workflow.

On a re-read, I'm not sure why I thought that, sorry. Context: I don't know much about machine learning and when I was scanning through comments, doing text generation one character at a time seemed silly and I must have been in the grips of "everything could be GPT!" hysteria. My robot detector needs work, clearly. Need to get educated.

> it doesn't know how to look at what it's written and decide if it matches its intent, or whether it'll break consistency or get in the way later.

And we can build other models specifically for this. We don't need to add this stuff to GPT-3; GPT-3 can literally act as a part, a component. GPT-3 can serve the role in a larger model that "imagination" does in a human brain—being fed inputs; having corresponding outputs scavenged through by the rest of the model; and then being "fed back" with input that relates to the scavenged outputs.

One thing I'd be very curious to see tried, is to get a system consisting of GPT-3 as "writer", and some other (summarization?) model as "editor", to attempt to dramatize or adapt into prose fiction, a machine-readable sequence of events (e.g. a machinima recording of a stage-play enacted within an MMO game.)

We already have models that turn machine-readable sequences of events directly into prose; see e.g. baseball news reporting. Such models can work just as well in reverse, summarizing in-domain prose back into machine-readable facts.

So if you take such a prose-to-factual-assertions "reading comprehension" model, and feed it GPT-3's output; and then measure the distance between the set of events comprehended by the "reading comprehension" model from GPT-3's output, and the source data (which is also in the form of a set of factual assertions), then you can iterate GPT-3 — maybe even one additional line of prose at a time — to find a story that is a consistent adaptation of the source. In this sense, GPT-3 is acting as a programmer, and the "reading comprehension" model as a compiler — with the compiler reaching out and erasing any line that doesn't compile.

Of course, you're limited in this by the "reading level" of the reading-comprehension model. But this is also true of humans; you can't get out a literary classic if the writer's editor and alpha-readers were five-year-olds.