Transcription
ChatGPT launched its most powerful version yet on August 7. Interestingly, a lot of people have gotten mixed results because they're not aware of the fundamental changes and how the model works now. And I've spent the last month digging into OpenAI's own documentation to figure out how to get the most value out of it. And by the end of this video, you'll do, too.
I condensed everything into six topics that you really need to know. And because I respect your time, I'll highlight them on the screen right now so you can jump ahead if you're here for something specific. But if most of this looks new to you, I would strongly encourage you to go through them all. I'm Ali Salam. I currently work as a director in a tech company. And on this channel, I'll help you turn tech and finance into your personal advantage.
[Music]
One of the first things you probably noticed in GPT5 is that the model now auto selects for you. If you ask a simple question, you usually get a fast, lightweight model. But if your request is complex, it will trigger a heavier reasoning model. And honestly, on the surface, that sounds great. The problem is that GPT5 is also tuned for cost efficiency, meaning that in practice you will often get a lighter model than you actually need, leaving the output feeling weaker than what you're used to if you used GPT40. And there are two easy fixes for this. First, you can prompt the model to think deeper. Add a line at the end of your prompt stating, "Think longer about this." And this will trigger a deeper effort in your response. Or the second option is to manually switch models at the top and select one of the dedicated thinking models to force a deeper compute.
Simple prompts generate simple answers, whereas more complex requests generate more complex answers. And what you often times feel is that the answer is trying to cover every single angle of the topic. And sometimes that can be a bit overwhelming. And if you've ever presented to a C-suite executive, you know that you always start with an executive summary before diving into the rest of the details. It kind of eases people in and gives them the big picture first. With ChatGPT, you can and should do the same. Here are three prompts you can use when you want a fast, high-level answer to a complex problem without wading you through the detail: Keep the answers under 100 words. Answer exactly in five sentences. Give me the executive summary first and then the detail. And by the way, if you're a technical person working with ChatGPT on the platform side, this is the same idea as controlling answer length through something called verbosity.
[Music]
If you're asking ChatGPT for something simple, you don't need much of a setup. But when you're asking it to break down complex topics, you need to kind of give it a better steer towards what you're trying to achieve. And more specifically, OpenAI actually calls out four topics. First is identity. Tell ChatGPT who it should be. Think of it like assigning a role. Act as my writing coach. Then instructions. What should it do and what should it avoid? This sets the boundaries. Third, examples. Show it the pattern that you want. Give it input and the kind of output that you expect. It learns the style or the format fairly easily. Lastly is context, where you give it the background. This could be your data. It could be the audience you're writing for, or really anything specific to your case. Together, these four steps make your prompts much sharper and answers far closer to what you actually want.
Quick pause. I have a favor to ask you. If you're enjoying the video so far, you should consider becoming a part of the small but very exclusive group of around 5% of viewers that have actually subscribed. And if you've already subscribed, I just want to say thank you. You're the reason why this channel keeps growing.
[Music]
This is such a free-win technique. It's very easy to implement and it takes almost no extra effort, but it will definitely increase or elevate the quality of your prompts. And the way it works is fairly simple. You're basically asking ChatGPT to, before giving you the answer, have it define what excellent looks like for whatever prompt that you're giving it. Score its own draft against that definition and then revise it until it hits the highest marks. Here's the go-to instruction I use, which you can just copy: "Before finalizing, privately create five to seven criteria for an excellent answer. Draft, self-score against the criteria, and revise until every criterion hits the highest possible score. Only show me the final version. Hide the rubric and drafts." Obviously, you don't need to do this for super simple tasks, but when it comes to more complex requests, it's worth building in the habit because it's such a free win. And I would strongly advise that you add this to your text expander if you're using one, so you can drop it in instantly when working with something important.
[Music]
So in the world of AI, the word "shot" simply means number of examples. And as you move from simpler to more complex requests, it's strongly advised to also move from zero to few shots. And here's what that means. For simple tasks like asking ChatGPT to spell-check an email, you don't need to provide an example. That's called zero shot. For something a little bit more advanced, say an analysis of a spreadsheet, you might want to include an example of how you want the summary formatted. And that's called one shot. And lastly, when you have a task that is super complex or you have a specific output in mind that follows a specific structure, style, or standard, then it's always best to provide several examples. And that is called few shot. So think of it this way: the more you show the model what good looks like, the closer the output will be to what you actually want. And as tasks get more complex, the number of ways that the model could deliver a technically correct answer also increases, unlike simple spell checks where there really are only a few outcomes that would be correct. And that's when it becomes a good practice to guide the model with one or few shots.
This one is brand new, at least at the time of this recording, and it's a real shift. I kind of see it as the beginning of a fundamental change in how we will interact with AI. So, up until now, AI has mostly been reactive. You ask, it answers. And Pulse flips that around by taking a more proactive role, reaching out to you with relevant updates. Here's how it works. Once a day, Pulse will run through an asynchronous research using your past chats, your memory, and if you enable them, connectors like Gmail or Google Calendar. And each morning you'll get a set of visual cards that you can quickly scan, expand for detail, and even curate with thumbs up or down. Unfortunately, right now it's Pro only. I don't personally have it, but if you do, I'm happy for you. However, given how central this could become, I would really expect it to be rolled out more broadly soon. And a couple of like top-of-mind examples that you could use this for would be, for example, trip preparation. So, flag calendar clashes, send top three airport lounges that you're eligible for. You could also use it for fitness or learning plans, allowing you to adjust your training plan if you're out traveling, or you could use it for something like career hygiene, where you get weekly hiring pulses within your industry from whatever sources that you find most credible.
And that's it. Those are the six main points. If you got any value out of this, please hit the like and leave a comment to let me know what was useful for you. Or if nothing was useful for you, let me know that as well. This channel is for you, so your feedback really matters. Also, if you're looking to level up your AI game, I would strongly recommend this video. It breaks down Google's Gen AI course and even shows you how to pass the official certification if you're interested. As always, thank you for trusting me with your time and I'll see you in the next one.