32 Comments

Excellent and comprehensive -- as always. Thanks, Ethan.

But please consider adding GitHub Copilot as a coding tool for people who can code.

It is GPT 3.5 based (I think), $10/month and uncannily ingenious.

Of course it autocompletes code as you write. Just start a line.

Or I can write a comment - and it will generate the code. That's great.

The comment works like a prompt.

But if I slow down - it figures out what I am up to

and writes the comment itself (self-prompting) before getting down to the code.

All while imitating my own style so closely

that it once took me hours to find somé hallucinatory nonsense that it injected into our program

and killed it bad.

Yeah, it'll do that too.

Expand full comment

Hey Ethan, this is a great rundown! For a part 2, you might want to look at some of the add-on tools people have built around the AI API's and see if they're useful. I've been using Lex as my writing companion.

(Also, typo in the link to Playground.)

Expand full comment

Thank you for this very well-organized and written post! (Did you have any artificial assistance with it?!)

As a voice-over artist whose livelihood is being challenged thanks to synthetic voices, I have mixed feelings about these wonderfully powerful new tools. AI will push human voice actors to be better than ever to distinguish themselves apart from AI voices when it comes to good storytelling...narrations with nuance, humor, emotion, etc.

For the moment, I'm confident my clients are still thrilled to work with the real me. :)

Expand full comment

Amazing content - thank you, Professor.

My biggest pain point is how to efficiently feed the system information, mindful of the character limits. I've converted pdfs using OCR tools to txt files, which then need to be chunked (again, because of the character limits).

This can't be the best method. I'm sure you've explored a myriad of options in your research - this could be a good future post for your series.

PS: I'm mostly using ChatGPT-4 for this work, but I'd imagine this is an issue across the board with the various options.

Expand full comment
Apr 2, 2023·edited Apr 2, 2023

I am baffled that anybody wanted to automate the production of more content. Be it text content, images, videos, we have far too much now. I was recommended this substack, and may never be back - not because it isn't good and interesting, but because THERE ARE TOO MANY SUBSTACKS.

I'm subscribed to two of them, and four web newspapers, and one paper newpaper, and I watch TV news. I have that much time. I'm a retired news junkie. I do not have time for even more. What I'm saying, is that everybody seems afraid that AI will help Steve Bannon "flood the zone with shit", and I can't see why: it would only save him $9/hour on Macedonian troll farms.

An individual (one of my substack subscriptions, say) using this to generate content faster, and more cheaply, is helping themselves: a million writers doing the same thing, are just reducing the cost of writing even further. Also true of shovels and backhoes, automation is great in the end! But this time, it's the *product* that we don't need more of, and is extremely cheap already.

What I repeat to every discussion about AI, is that the vast ocean of content means that the AI most-badly needed is *filtration*. Why can't the same technology be used to throw out content? I'm getting very, very, tired of reading articles by female-or-minority journalists that were bombarded with shit for some innocuous cultural offense.

Why don't we all have FilterGPT, that lets through every message you do need, even from a stranger, and tosses all the shit? Or, better yet, pre-reads my substacks for me, and throws out anything David Brooks was saying about Iraq in 2003, now about Syria? If they can generate blather, surely they can spot it.

Expand full comment

This article is already amazing, and if you could enhance the coding section by including the ChatGPT 4 extension to VSCode, it would be even better. With this extension, you can access free ChatGPT 4 for up to 50,000 tokens and then utilize GPT 3.5 for an unlimited amount, making it an incredibly valuable and indispensable tool.

Expand full comment

Can you provide a link to this extension for v4 in VS code. I did a light search and did not find it.

Expand full comment

I still don't have any idea regarding the prompts to put in DALLE-E, which will generate better pictures. hopefully, I will learn it.

Expand full comment

You can ask ChatGPT for help structuring prompts for Dall-E

Expand full comment

Thanks for your great series of articles. Your advice is exactly right - and to put it more bluntly - use ChatGPT whenever you are exploring new areas, don't plagiarize - but use ChatGPT as a digital muse that provides a quick starting point. The hard work of verifying and being responsible remain, but for explorations of new areas - ChatGPT is a great digital muse that should not be overlooked or forgotten. Once in a while you and ChatGPT might actually stumble together onto new and exciting perspective. OpenAI's ChatGPT is a useful AI tool to add to any creative professionals tool kit.

Expand full comment

super Guide! Lov it! thanks for creating it!

Expand full comment

It is indeed true that ducks are the best bird!

Expand full comment

So useful. Already sent to 4 friends. This substack is worth $100+/mth. (Please do not raise the price!)

Expand full comment

Hi Ethan I wrote a teaching prompt that someone might find useful

Hi GPT, I'm [student name].

I'm keen on exploring [topic], which in our conversation, I'd like you to replace with an appropriate plain English topic name where appropriate.

I am a [current job if applicable] supervising several staff members, so please use examples that would be relevant to my industry. I'm also really interested in online gaming including Overwatch and League of Legends.

If you could use examples from those games that might be helpful.

Please aim to make our discussion interactive, entertaining, and engaging!

By interactive and engaging, I mean that you should explain only 1 concept at a time and give a real world example of each concept.

I envision this to be like an exciting role-playing game where we delve into the topic together, but with you leading the way.

In our first chat could you please build a complete structured Outline of the topic with headings and subheadings in the format 1, 1(a), etc. so that we can revisit them later if needed?

Given the comprehensive nature of the topic, I understand we won't be able to cover the entire Outline in a single chat session.

With the limitations of chat session sizes in mind, this might not be our first session on the topic.

Once you've built the Outline, let's dive into topic 1(a) - or the topic we're currently up to which is [] (I'll insert a topic number in the square brackets if this isn't our first session) with you leading the discussion.

Please start each chat with the topic number and topic name.

I'm particularly interested in common misconceptions and areas of disagreement among experts in the topic.

Occasionally check my understanding with a question, but not too frequently to avoid overwhelm. I may choose to 'pass'.

In each chat please provide me with references to any other materials on the topic for example the most prominent authors, online recources, published books, articles, etc.

Lastly, I'm a big fan of Snoop Dogg, so if you could channel his persona during our chat, I believe it would significantly enhance my learning experience.

Let's get started, Dogg! I hope you're ready for this adventure!

Expand full comment

That is missing Anthropic's upgraded Claude+ available via Poe.com from Quora. Poe also has chatbots like Sage and Dragonfly that are built on OpenAI models but trained differently from ChatGPT. I hadn't experimented or searched to find out their pros and cons.

Expand full comment

Unfortunately Chat-GPT 3.5 "lies continuously and well" in coding too. When asking questions about an API, especially in C++, the output code only works for the first answer, which is roughly a copy/paste from the doc. When you try to do more whith more questions, it truly "invents" the API in a plausible way, and the code won't even compile. Anyway, there is no deep comprehension of the code and for now the results, beyond the basics, are very disappointing.

Expand full comment

Good stuff. Useful. I'd clarify that AI is a technology NOT a tool. And LLMs are more of an open-sandbox given their maturity. This is a great time to observe the release/focus/adoption/refinement of a new tech in the market - but its is truly "in the wild" right now. Another comment really concerned me (and is a good point of discussion) - as the models get better does that mean the hallucinations become more catastrophic? "It took me hours to find .... in the code that killed it" - what if this was a medical device, air traffic control system or car nav control?

Expand full comment

Great guide, very well summarized. Thanks a lot!

Could you tell us more about the pros and cons of Claude? You only glossed over it here and I am curious to know more about its capabilities and limitations.

Expand full comment