Some Advice for How to Learn ChatGPT
- Accomplishing Specific Tasks
- Going Deeper on AI Models and Current Capabilities
- Using ChatGPT in Your Daily Tasks
- What to Do About Hallucinations
- Learning the Public API and Open Source Libraries
- Discover Related Articles
Something remarkable happened to me today. Over roughly the last 6 months, I’ve been subjecting nearly every single living person I’ve come into close proximity with to an endless barrage of stream-of-consciousness takes on the meaning, implications, and consequences of ChatGPT. Nearly all encounters approximated the following interaction:
But in a shocking turn of events, within the time-span of 24 hours, I found not just one but two individuals who were broadcasting explicit messages that signaled they wanted to learn more about ChatGPT. Most importantly, these messages were broadcast before I talked to either individual. This indicated that (unlike previous individuals) these individuals were consensually agreeing to ChatGPT play.
Having indicated that they wanted to invasively receive my thoughts about ChatGPT directly into their brains, I proceeded to acquiesce. The first individual was my father, who expressed an aspiration to my mother that he wanted to “figure out ChatGPT.” I’ll send him this article once it’s been published (Hi Dad!) to find out whether he found it useful, useless, or neither cause he didn’t actually use it yet.
After hearing that my dad was interested in ChatGPT, I later saw a friend (previously mentioned second individual) tweet a request for ChatGPT learning resources. Knowing that I would likely need to write one for my dad, I dashed off some quick, off-the-cuff thoughts to Dom on Twitter. After this section, the rest of the blog post contains a lightly edited version of the response I gave Dom.
Accomplishing Specific Tasks
The majority of articles and videos out there focus on showing how to accomplish specific tasks with ChatGPT and repeat a lot of the same advice over and over. I thought Adrian Twarog’s video, ChatGPT Tutorial - A Crash Course on Chat GPT for Beginners, does a good job covering the most common use cases with concise examples that show the immediate utility anyone can get today from ChatGPT.
Going Deeper on AI Models and Current Capabilities
For more in-depth information about engineering your own AI models, Noah Hein is building a free course with Swyx that is still in the works but I anticipate has the potential to be very high quality. You can sign up here and follow along as it’s being released.
I’d advise investing in a course like this and steadfastly avoiding all the redundant Twitter threads and Medium articles that promise to teach you 37 prompts which will change your life. Noah’s course is going to dig deeper into a handful of different subjects and questions including but not limited to:
- What is happening when you send a request and receive a response from the underlying models exposed publicly through various API endpoints?
- What are the capabilities of state of the art, non-textual models created by OpenAI such as DALL-E or Whisper?
- How do you fine tune a model with more current data or train a new model totally from scratch?
Using ChatGPT in Your Daily Tasks
In tandem with learning more about the AI models themselves, I’d also recommend picking a couple of things from Adrian’s list of examples and begin to work them into your daily workflows. Tasks I personally use ChatGPT for include summarizing lots of writing in articles, blogs, and official announcements.
I generate lots of small code snippets for example apps with ChatGPT and I also use it to translate code samples between different frameworks and frontend syntaxes. Techniques for working alongside ChatGPT aren’t very well documented anywhere and for now they must be learned mostly through trial and error.
By using ChatGPT more and more, you’ll find ways to prompt it better and recursively. I would especially encourage you to give follow up prompts to ChatGPT’s responses which is what I mean by recursive prompting. Try modifying the length of a summary or tweaking the tone/writing style. Experiment with prompts like:
- Rewrite that but make it half as long
- Rewrite that but talk like a pirate
- Rewrite that but change my voice so I sound happy and like I enjoy my job
What to Do About Hallucinations
It’s especially important to learn how to fix errors of fact and bugs in code. You can prompt it with new content to correct for a hallucination. You can give it relevant information it lacks out of the box about events after September 2021 by feeding it text from a current article containing the new information. Or, if all else fails, you can say to ChatGPT that it’s wrong and should feel bad about being wrong.
In those situations, ChatGPT usually concedes that it has indeed made a mistake in one of its responses. It will subsequently apologize and miraculously, over 95% of the time ChatGPT means the apology sincerely and does not also begin to plot your murder in a covert subroutine running on a hidden process (calm down, there’s multiple open issues and a patch release on the roadmap scheduled for Q4 2024).
Something I think is currently most missing from the current crop of educational content is instruction for developing someone’s ability to improve GPT’s generated output by guiding it through multiple interactions. It’s a difficult skill to learn unless you get in the habit of frequently interacting with ChatGPT on a variety of inputs and outputs.
Learning the Public API and Open Source Libraries
Lastly, I’d recommend getting into the official API sooner rather than later. The API exposes a lot more functionality and configurability not available in the raw ChatGPT interface which gives a better idea of what’s happening under the hood.
Go on, give it a spin. It’s a blast. You know you want to.
This has been my second opinion piece about ChatGPT and AI more broadly. It certainly won’t be the last, but next time I’ll be taking a break from think pieces to finish a massive post next week I’ve been writing for months called A First Look at OpenAI.
I’ll leave my opinions aside in favor of an in-depth, technical tutorial. A GitHub repository of code will accompany the walkthrough. The tutorial will provide a deep dive on OpenAI’s API endpoints and the DX achieved with the
openai-node open source library.
Remaining ChatGPT Contentions
For the final section of this article, I would like to directly address three significant points of contention I hear frequently from AI skeptics but haven’t spoken about yet:
ChatGPT is useless, its answers are terrible, and the output it produces is unusable. At best it can only achieve some basic parlor tricks, simple stuff like translating the entire Bible into any language you can think of and providing detailed historical and textual analysis to every verse.
I don’t want to give up writing, I won’t give up writing, and you can’t make me give up writing! A subscription to ChatGPT might as well be a prescription to chop off my hands with a meat cleaver.
What makes you so damn sure you’re right, anyway? All the AI experts quoted in the NY Times say this thing is useless. According to them, all the billions of dollars going into developing language models is a waste and that should be going to their AI labs (such truly selfless souls to make a sacrifice like that to rescue the future of AI).
ChatGPT Is Useless and Words Aren’t the Point
If you don’t like the way ChatGPT says something even when the thing it’s saying is consistent with the question or request you asked, then your entire argument amounts to, “I am a snob and my taste is more important than objective truth.”
Before ChatGPT, the most common argument against using AI for serious work was the widespread belief that AI wasn’t intelligent yet and couldn’t generate useful work, novel ideas, or anything approaching a tool with free world utility.
From the 1940s to December 2022, that was pretty much true across the board for all AI models. But it’s almost impossible in 2023 for anyone to make that claim today with a straight face (although somehow Gary Marcus continues to do so).
Tens of millions of people now pay money to use ChatGPT and produce large amounts of work with its assistance that the users are then monetarily compensated for. This can be motivated by the wish to save time by finishing work faster or the wish to make more money by producing a larger output of work.
For as long as the term AI has existed, we’ve never truly had useful AI models that could accelerate someone’s work to a larger degree than the already existing productivity tools and techniques that developed during the automation, computation, and information revolutions of the 20th century.
Today, the most common arguments I hear as push back against the argument that ChatGPT has any usefulness or utility:
- I don’t like this output for totally subjective reasons.
- I used ChatGPT wrong, got bad output, and didn’t know how to fix it.
- I don’t understand how ChatGPT works or what it is so I think it’s the equivalent of tapping the “generate next word” button on an iPhone thousands of times in a row.
But I Love Writing and Would NEVER Give It Up
Whoa, whoa, whoa, hold on! Who ever said anything about giving up writing? The second most common push back I hear comes from people who write for a living and thus stand to gain the most from adopting ChatGPT as a writing aid.
They’ll claim that they don’t want to use AI because they don’t want to give up writing or they think they’ll become depressed when a program is able to do all their work for them better than even they could do it.
This is much trickier to counter especially because of how counterintuitive a conclusion it is for lovers of ChatGPT. If you’re a person who loves ChatGPT and is using it frequently you are probably writing more than you’ve ever written in your entire life.
Since ChatGPT came out I’ve written more in the previous 6 months than the amount I wrote in the 3 years before that. All of that writing is then also augmented by having all of the additional text written by ChatGPT that is available to use.
Writing is something writers value as an activity itself and using ChatGPT does not stop you from writing your own thoughts when you are thinking them! I didn’t use ChatGPT to write a single word, sentence, paragraph, or idea contained in this blog post.
My Thoughts Are Constantly Evolving and Yours Should Too
Last week, I published on my blog my first article containing some of my recent musings on the intersection between ChatGPT and DevRel. I’m currently working through my thoughts and ideas in real time by writing about them in public as frequently as possible.
The complexity, unpredictability, and speed of progress around all this stuff creates an imperative for consistent vigilance. We must all strive to resist deception and misconceptions from others and most of all we must resist the ever present temptation to deceive ourselves.
The first principle is that you must not fool yourself. And you are the easiest person to fool so you have to be very careful about that.
After you’ve not fooled yourself, it’s easy not to fool others. You just have to be honest in a conventional way after that.
Richard Feynman - Caltech 1974 Commencement Address
With that quote in mind, I want to stress that everything I’m saying at this point about the current state of AI and thus my corresponding recommendations are based on a brief snapshot in time.
My thesis hasn’t fully cohered and is constantly in flux. The only through-line I feel confident will remain is the thesis that everyone should use ChatGPT because everyone stands to benefit from using it.
How a specific person will specifically benefit from ChatGPT then becomes the perennial question. I welcome reactions, feedback, critiques, and any other thoughts you might have. Even if your thoughts are, “nuh-uh, you’re wrong!”