Transcription
I'm prompt engineer Michael Taylor. This is prompt engineering.
[Music]
Support at Marites wants to know: serious question, what is a prompt engineer?
Prompt engineering, one of the main things I'm doing every day as a prompt engineer, is A/B testing lots of different variations of prompts. So I might try asking the AI one thing and then ask it a completely different way and see which one works best.
A prompt engineer might be employed by a company in order to optimize the prompts that they're using in their AI applications.
Adam Jones Inc. is asking: does anyone else use "please" and "thank you" when communicating with ChatGPT and Perplexity? I'm hoping that I'll get better responses or be treated slightly better with the AI models.
Ever take over? Specifically saying "please" and "thank you," there is no evidence that that improves the results. But being emotional in your prompts, for example, using all caps, does actually improve performance as well.
So if you say, for example, "this is very important for my career," and you add that to your prompt, it will actually do a more diligent job. It has learned that from reading Reddit posts and reading social media posts that when someone says, "this is very important for my career," the other people that answer actually do answer more diligently.
One thing that we saw last winter was that ChatGPT started to get a little bit lazy. What someone figured out was that when it knows that the date is December, then ChatGPT actually does get lazier because it's learned from us that you should work a little bit less in the holidays.
At Shuffle Upus is asking: do you get better results from LLMs when you prompt it to imagine that you're an experienced astrophysicist? Why would you want them to pretend?
Let's do a little experiment here. Let's write the prompt as an astrophysicist and then write the same prompt as a 5-year-old and see the difference.
So I've asked it to tell me about quantum mechanics in two lines as an astrophysicist, and you can see it uses a lot of big words that a typical astrophysicist would know.
We can then ask it the same thing as a 5-year-old, and now it's explaining quantum mechanics as a magic world. Tiny things like atoms can be in two places at once.
The overriding rule is that you should be direct and concise as an astrophysicist, or you are an astrophysicist. That tends to work better than adding unnecessary words like "imagine."
At VBR is looking for any tips on how to improve my prompts. Well, there are actually thousands of prompt engineering techniques, but there's two that get the biggest results for the least amount of effort.
One is giving direction, and the second one is providing examples. Say, for example, I had this prompt, and I invented a product where it's a pair of shoes that can fit any foot size.
Now, how can I improve that prompt template? One thing I can do is to give it some direction. One person who is famous at product naming was Steve Jobs. You could invoke his name in the prompt template, and you're going to get product names in that style.
Alternatively, if you prefer Elon Musk's style of naming companies, you can provide some examples of the types of names that you really like.
The reason there's two hashtags in front of this is that this means this is a title. It really helps ChatGPT get less confused if you put titles on the different sections of your prompt.
At Pete Mandic is asking a serious question about AI artists: why does the number of fingers on a human hand seem to be particularly difficult for them?
The difficulty in rendering fingers is that it's very intricate, and the physics is quite difficult to understand. These models were pretty small; they didn't have that many parameters, so they hadn't really learned how the world works yet.
We also have a really strong eye for whether fingers are wrong or whether eyes are wrong. It's something that we look out for as humans.
A thing you might try in a prompt is to say, "make the fingers look good." That tends to not work either because everything in a prompt is positively weighted. If you say, "don't put a picture of an elephant in the room," then it will actually introduce a picture of an elephant.
So what you need is a negative prompt. That's not always available; for example, it's not currently available in DALL-E, but it is available in Stable Diffusion.
So we're going to type in "oil painting hanging in a gallery." We're going to hit "dream."
What you can see is that some of them have a big gold frame, but the one on the right doesn't have a frame, and I actually prefer that.
So how can I get it to remove the frames? One thing I can do is if I add up the negative prompt here, I can say "frames" in the negative prompt. It's going to remove that, and now we can see that all of the paintings don't have frames.
At Roberto Digital is asking: what is the weirdest response you've gotten from ChatGPT?
So my favorite one is if you ask it, "who is Tom Cruise's mother?" It knows who it is: Mary Lee Pfeiffer. If you ask it, "who is Mary Lee Pfeiffer's famous son?" it doesn't know that Mary Lee Pfeiffer's son is Tom Cruise, so it will make something up.
I think the last one that I got was John Travolta. So the reason why this happens is there's lots of information on the internet about Tom Cruise and who his mother is, but there's not that much information on the internet about Mary Lee Pfeiffer and who her son is.
Hallucinating is when the AI makes something up that's wrong, and it's really hard to get away from hallucination because it's part of why these LLMs work. When you're asking it to be creative, creativity is really just hallucinating something that doesn't exist yet.
So you want it to be creative, but you just don't want it to be creative with the facts.
At Schwarz Child is asking: I'm not an expert in AI, but if an LLM is trained on biased data, then won't that bias come through in its responses?
Well, you're absolutely correct because AIs are trained on all of the data from the internet, and the internet is full of bias because it comes from us, and humans are biased too.
But it can be pretty hard to correct for those biases by adding guardrails because by trying to remove bias in one direction, you might be adding bias in another direction.
A famous example was when Google added to their prompts for their AI image generator service an instruction that they should always show diverse people in certain job roles.
What happened was people tried to make images of George Washington, and it would never create a white George Washington. In trying to do the right thing and solve for one bias, they actually introduced a different bias they weren't expecting.
There is a lot of work in the research labs, like Anthropic, which has a whole safety research team that has figured out, you know, where is the racist neuron in Clo, which is their model? You know, where is the neuron that represents hate speech? You know, where is the neuron that represents dangerous activities?
And they've been able to dial down those features.
At Cara John wants to know: how much of the conversation context does ChatGPT actually remember? If we chatted for a year with information-dense messages, would it be able to refer back to info from a year ago?
So when you open a new chat session with ChatGPT, it doesn't know anything about you unless you put something in your settings specifically.
They do have a feature, which is a memory feature that is experimental, and I don't think it's on by default.
So one trick that I tend to use is I will get all of the context in one thread for a task. I'll just ask it to summarize, and then I'll take that summary and then start a new thread.
Then I've got the summary, more condensed information. It will get less confused by all of the previous history it doesn't need to know about.
At Brandon White, does customizing your settings in ChatGPT and providing your bio, SLP personal info, help better results?
Yes, I find that you get wildly different results when you put some information in the custom instructions. You have two fields: custom instructions, which is what would you like ChatGPT to know about you to provide better responses, and the second box is how would you like ChatGPT to respond.
I use ChatGPT a lot for programming, so I tell it what type of languages I'm using, what type of frameworks. I give it some preferences in terms of how I like my code to be written.
The second box is really anything that you get annoyed about when you're using ChatGPT. You could put that in the box.
So, for example, some people put "quit yapping," and then it will give you briefer responses.
At Travis Media asks: what makes a prompt engineer an engineer?
A prompt engineer is designing that system of prompts that are being used in the application and making sure that they're safe for deployment.
Make sure that they work again and again and again reliably. That's the same sort of thing that a civil engineer is doing with a bridge, right? Like they're designing the bridge, and they're making sure that when you drive over it, it's not going to crash into the river.
At Igan Blade is asking: do you think we can draw parallels between large language models and human brains?
LLMs, or large language models, are actually based on human biology. They're what happens when you try to make artificial neural networks simulate what our biological neural networks do in our brain.
So there are a lot of similarities, and a lot of the things that work in managing humans also work in managing AIs.
So if you've heard of transformer models, which is what the LLMs are all based on, the breakthrough there was figuring out how to make it pay attention to the right words in a sentence in order to predict the next token or word in the sentence.
So that was the really big breakthrough that was made by Google and then used by OpenAI to create ChatGPT.
At This One Optimistic wants to know: what are tokens?
So let's say I started writing the sentence, "LeBron James went to the..." What word could come next?
Well, an LLM looks at all the words on the internet and then calculates the probability of what the next word might be.
So this is the token "Miami," which has a 14% chance of coming next, and we have "Lakers," which has a 133% chance of coming next.
We also have the word "loss," which is just the beginning of the word "Los Angeles." Here we have the token "Cleveland," which only has a 4% chance of showing up, but the LLM will sometimes pick this word, and that's where it gets its creativity from.
It's not always picking the highest probability word, just a word that's quite likely.
The reason they use tokens instead of words is it's just more efficient. When you have a token, which is a little part of a word like "loss," that can be more flexible, and it can be trained to be used in different contexts.
At Edam Test Mo, for what is the best LLM in your opinion?
For me, it's Claude 3. Opus, I agree. Anthropic, who makes Claude 3, is doing a great job.
I'm going to test this against ChatGPT and then Meta Llama, which is an open-source model, and show you the difference in results.
So the prompt we're using is: give me a list of five product names for a shoe that fits any foot size. We're testing the model's creativity here.
And you can see that we have "Unifit Shoes" as one idea, "Adaptic Shoes," which is pretty creative, and "One Size SS," which is my personal favorite.
I'm just going to copy this prompt to Claude, and with the same prompt, we get different names. We have "Morfit," "Adap Toep," "Omnis Shoe." That's my new favorite.
Now we're going to test it on Llama 3, which is Meta's open-source model, and you can see it comes up with really different names: "Fit Flex," "Size Savvy," "Adjust a Step," "Universal Fit."
It comes with this text at the beginning, and then it's describing each name as well. That's not what I asked it to do.
Personally, it's subjective, but I like the Anthropic Claude response best.
At G Monster 7000 is asking: what is the simple task that an LLM has done that has changed your life?
For me personally, it's been the programming ability that I get from using ChatGPT and Anthropic's Claude. Those models are so good at writing code and explaining what that code does that I have really lost my fear of what I can build.
So if we pop over here to Claude, I've made up a fake product, which alerts you if your baby is choking. I'm trying to build a landing page for it because my developers are busy, and it's actually going through and just writing that code for me.
Say, for example, I don't understand what this section is doing. I can just copy that and then paste it at the bottom and say, "What does this do?" and it's going to give me bullet points on what that specific code is doing step by step.
And that's the way that you learn with programming. I find that I just never get stuck when I use this.
One of the coolest things I've done, a little automation or a little life hack that I use every day, is that I set up an email address that I can email with any interesting links I've found.
That will send those to AI, summarize them, and then put them all into a spreadsheet for me to look at later.
Can of Goldson is asking: what is prompt chaining?
If you wanted to write a blog post, you wouldn't get great results just by asking it to write it all in one step.
What I find works is if you ask it first to write an outline, do some research, and then when I'm happy with the outline, come back and fill in the rest of the article.
You get much better results, and they're comprehensive and fit the full brief of the article that you wanted to write.
Not only does it make the thought process more observable because you can see what happened at each step and which steps failed, but also the LLM gets less confused because you don't have a huge prompt with lots of different conflicting instructions that it has to try and follow.
At Automation Ace is asking: how can you automate AI?
That's what's called an autonomous agent, where it's running in a loop, and it keeps prompting itself and correcting its work until it finally achieves the higher-level goal.
Microsoft Autogen is a framework for autonomous agents, an open-source framework that anyone could try if you know how to code.
And I think that's really the big difference between ChatGPT, this helpful assistant that we're using day-to-day, versus having an AI employee in your Slack that you can just say, "make me more money for the company," and it will go and try different things until something works.
At Mr. Drov is asking: how would you prompt the LLM to improve the prompt?
There's actually been a lot of really interesting research here, techniques like the automatic prompt engineer technique, where the LLM will write prompts for other LLMs, and this works really well.
I actually use it all of the time to optimize my prompts. Just because you're a prompt engineer doesn't necessarily mean you're immune from your work being automated as well.
At Bahoo Prompts is asking: how long until prompt engineering or a future field related to this becomes a degree? Will it be a standalone field or part of every field being taught?
That's a really great question because some people I talk to say that prompt engineering isn't going to be a job in five years. It's not even going to be something we practice because these models are going to be so good we won't need to prompt them.
I tend to disagree because, of course, humans are already pretty intelligent, and we need prompting. We have an HR team, we have a legal team, we have management.
So I think that the practice of prompt engineering will always be a skill that you need to do your job. But I don't necessarily think that in five years we'll be calling ourselves prompt engineers.
So those are all the questions for today. Thanks for watching prompt engineering support.
[Music]