gpt-3-experiments

The problem with ChatGPT as a replacement for Google is that it was not designed to produce accurate facts, and it shows. This model cut its teeth writing articles about the discovery of unicorns in the Andes[0] for goodness sake! It's a language model, and a very impressive one at that, but language is used to express falsehoods and fiction just as regularly as it is used to express truth. This doesn't mean that it can't produce accurate facts, most of the time it does! But when it does produce nonsense, it does it in exactly the same tone of authority, so if you don't already know the answer you may well walk away believing an AI hallucination. And the trouble is it doesn't really matter if everyone here thinks "well, I would follow up each request with research to verify the answer", because most people won't! This is like the Google a

See also my experiments with GPT-3 on sane prompts, which have wildly varying quality even after generating them in bulk: https://github.com/minimaxir/gpt-3-experiments Creative writing hasn't been one of the super-hyped use cases by OpenAI for the OpenAI API outside of AI Dungeon, surprisingly. For just random generation, the necessary curation can detract from the time-savings advantages. (as an aside, the API is also extremely expensive for long-form content to the point I'm not sure how the economics work for these startups even with charging monthly fees). I'm more bullish on small bespoke models for a given use case, which is what I spend my time researching. (example: a ~1M parameter GPT-2 Magic: the Gathering card generator: https://colab.research.google.com/drive/1VOt090UzvltoBgMdUZm... )

Thanks, didn't twig onto the fact that you linked a subtree of the whole repo. Weird that even with the nonzero temp the AskReddit prompt went a bit loopy. > https://github.com/minimaxir/gpt-3-experiments/blob/master/e... Oh my goodness that is absurd in the most delightful way. Thanks for sharing that.

Those are the temperatures (0.0, 0.7, etc.) that control the craziness. 0.0 means the model makes the optimal guess for each token. Unfortunately, both GPT-2 and GPT-3 have a tendency to enter loops. But it's not always bad: https://github.com/minimaxir/gpt-3-experiments/blob/master/e...

I ran an experiment through GPT-3 with just that frame: https://github.com/minimaxir/gpt-3-experiments/tree/master/e...

The title is very misleading: GPT-3 generated the titles, but as usual, they were manually curated on the "picked the title I liked the most" criteria. There's more to getting to the front of Hacker News than just a good title (although it helps); ironically this framing undersells the writing capabilities of the author by attributing success to the AI only. There are ways in theory for AI models to generate "better" titles, which is something I'm actively researching (rough example: https://github.com/minimaxir/gpt-3-experiments/tree/master/e...).

I have a demo for /r/WritingPrompts/ titles (https://github.com/minimaxir/gpt-3-experiments/tree/master/e...), albeit not the generated output from that title.

The 0.7 unicorn example is doing absolutely nothing to temper my expectations. Quite the opposite in fact. (https://github.com/minimaxir/gpt-3-experiments/blob/master/e...)

After spending a lot of time working with GPT-3/the OpenAI API (https://github.com/minimaxir/gpt-3-experiments ), one notable part of GPT-3 is the high signal-to-noise ratio in generated output. When finetuning GPT-2, only about 5-10% of the generated output is usable/coherent. But with GPT-3, easily 30%-40% of the generated text is usable/coherent, which is a big boost in quality.

I released a large number of GPT-3 demos yesterday: https://github.com/minimaxir/gpt-3-experiments