Hacker News Reader

AI text content generation is indeed a legit industry that's still in its nascent stages. It's why I myself have spent a lot of time working with it, and working on tools for fully custom text generation models (https://github.com/minimaxir/aitextgen).

However, there are tradeoffs currently. In the case of GPT-3, it's cost and risk of brushing against the Content Guidelines.

There's also the surprisingly underdiscussed risk of copyright of generated content. OpenAI won't enforce their own copyright, but it's possible for GPT-3 to output existing content verbatim which is a massive legal liability. (it's half the reason I'm researching custom models fully trained with copyright-safe content)

I would like to think the consumer would merit a thought, too.

Fiction might be one thing; if it is entertaining, that's enough. But if I'm reading something supposedly nonfiction that is generated by a machine, I want to know provenance.

In the alternative, it should have a human's name attached to say that they've verified it is correct information, and take the reputation hit if it isn't. Given the above discussion of copyright, it seems reasonable enough - if you want to profit from AI output, you should stand behind it.

On the topic of copyright, it has a ton of stuff memorized. The first thing that came to mind when I checked this many months ago was the first chapter of harry potter, which it knows verbatim. For whatever that’s worth.

UPDATE: These got a fair amount of traction, and I removed them out of an abundance of caution around deployment regulations that OpenAI enforces. Also cost considerations. I don't want to hijack the thread away from OpenAI, but you can also build stuff with Cohere and AI21 on Mantium, AI21's J1-Jumbo has pretty good performance, and Cohere just put out some significant updates for their models.

UPDATE 2: I couldn't help myself. I think this stuff is pretty fun. So here's a biology blog post generator using 2 chained Cohere prompts :) https://11292388-8f03-42d2-8a68-7039b24fcc2e-i.share.mantium...

However, the thing that is interesting to me about OpenAI, AI21, Cohere, and all the other LLM providers is that they are broadly useful, and often helpful.

The thing I find interesting about GPT-3 and company is that they do say things that are "often helpful" but that "often" doesn't necessarily translate to "broadly useful".

For example, lately, I've been some car repair - I'm a strict amateur. Suppose I asked GPT-3 how to do X. If there's a 75% chance it gives the right answer and a 25% chance it gives an answer that could damage my car or injure me, I'd say the thing could 0% useful, despite being quite impressive.

I like your use case for your blog. I dabbled with something similar last month: I have had a sci-fi mini-book that I have been slowly writing for a long time. I tried using their text completion API on long snippets of my own prose. I got some interesting ideas. If I end up using ideas and generated text, I have no idea how to license the book, probably some form of Creative Commons license (I was the featured Creative Commoner about 20 years ago, BTW, and use one of their licenses for my AI related books). To be fair, the Copyright would belong to me and everyone who contributed any text in the web that the model was trained on. I wonder what Laurence Lessig’s take on this will be.

EDIT: I just asked Lessig on Twitter what his opinion is on this.

I just tried your biology blog post generator and the second paragraph of the generated text, also the second sentence, is "Transcription is the process of converting audio into text." Obviously, the generator is confusing audio transcription with biological transcription like DNA transcription. Is this a common occurrence? Or did I make some mistakes in using the generator? I just pressed the "Execute" button.

This is, in my opinion, one of the biggest challenges with generative models right now. I'm not sure if this is the industry-adopted term, but I call them hallucinations. This is why I don't just pipe it straight into my blog, but rather use it as inspiration for a blog post that I write myself. It is easier for me to edit and expand on something that is already written, though.

FYI, I get a 404 for both of them.

I use Mantium and have had a great experience so far generating company marketing material

Good to see Cohere.ai mentioned in your comment