Confluence for 6.9.24
Insights from a leading AI developer. GPT-4o and GPTs for all. New reports on generative AI awareness and adoption. Getting better output through process (rather than prompt).
Welcome to Confluence. This is a longish issue, so you might freshen your beverage of choice. Here’s what has our attention this week at the intersection of generative AI and corporate communication:
Insights from a Leading AI Developer
GPT-4o and GPTs for All
New Reports on Generative AI Awareness and Adoption
Getting Better Output Through Process (Rather Than Prompt)
Insights From a Leading AI Developer
We had the opportunity to learn how a leading AI developer thinks about using and getting the most from today’s generative AI tools.
Last week, our firm had the privilege of spending three days with Dean Thompson of Eloquence AI, a leading generative AI developer with whom several of us have been collaborating. Dean’s an expert in developing applications that work with generative AI models, so we queried him at length about how he thinks about working with and getting the most from these tools.
Among Dean’s many strengths is his ability to translate his deep technical perspective on the fundamentals of how AI models work into clear, practical patterns and principles that a normal human can understand. We gained too many insights to list here, but several stand out and Dean’s given us permission to share them in Confluence (and this is our take on Dean’s advice, not his advice in his own words).
On the primary ways in which generative AI can add value to organizations and individuals today, Dean suggested four: Making individuals more productive through time-saving features, doing real work through content creation, facilitating human communication (by adopting an audience’s point of view to anticipate reactions, adopting a persona to engage in role play, and more), and offering alternative perspectives. This last was perhaps most interesting to us. As Dean put it, with generative AI, organizations have a “new, weird face at the table” — or really, an infinite number of outside perspectives — that one can be summon to push and expand one’s thinking.
Dean also shared his advice for using these tools:
Be stubborn. Don’t accept the AI’s first output at face value, and push it to continue refining until its output reaches your satisfaction. AI tools respond well to feedback and direction, so provide feedback just as you would to a human.
Have the model critique its own output. Today’s leading generative AI models are often better at critiquing existing content than generating new content. Use this to your advantage by having the AI critique its own output. One way to apply this is to always have the model critique its own first draft and then generate a second draft based on that critique. Ignore the first draft, and start paying attention at the second. (In fact, Dean said he usually doesn’t even read a first draft of model output, and immediately asks the model to critique and improve it.)
Start broad, then get specific. Rather than trying to craft a perfect starting prompt, start by probing the model to see how it “thinks” about your problem or situation, then work with the model to develop a plan from there. We share an example of this approach in action in the “Getting Better Output Through Process (Rather Than Prompt)” item below.
Provide context. The more context the model knows about your situation — your goals, relevant details that matter, and so on — the better it can help you. As Dean put it, “Imagine walking up to a stranger in the parking lot and trying to get them to help you with something.” That’s the situation the models are in when you come to them, so be sure to provide the right context to equip them to be most helpful to you.
Don’t stress. These models are inherently unpredictable, and it may take a while to figure out the best way to get what you need. Accept the fact that, sometimes, you will waste time with these tools, but on net the tradeoff should be worth it. And remember that the more you use them, the better your intuition for using them will be, and the less time you’ll be likely to waste.
And last, one major limitation to keep in mind: The models are extremely limited when it comes to “executive function.” Without close human guidance, these models are not good at coming up with their own goals and developing and executing a plan of their own. They can do both of those things extremely well when guided by a user, but as Dean aptly put it to us, “The further they get away from you, the worse they are.” This may well change as capabilities evolve, but for now, you should aim to partner with these tools — not outsource to them.
GPT-4o and GPTs for All
Better, faster, stronger — but with a few limitations.
A few weeks ago, we discussed potential implications of OpenAI’s GPT-4o release and the “coming wave” it might bring. Well, that wave is here — OpenAI’s leading model is now available through ChatGPT for free. GPT-4o can’t yet do everything that OpenAI showed in its demos, but free users can access features that had previously been behind a paywall. This includes:
GPT-4o: OpenAI’s most impressive model yet, offering improved speed and enhanced capabilities for drafting content and analyzing large amounts of input. While free users do have a cap on the number of messages they can exchange with GPT-4o, this new model opens up significant new possibilities for use.
Multimodal inputs: One of the top features of GPT-4o is its ability to seamlessly accept and work with text, voice, and images. This means you can now communicate with the AI using a combination of these formats, making for more intuitive and efficient interactions — analyzing an image, transcribing an audio file, or generating a response in any format, and more.
Improved data analysis: All ChatGPT users can now leverage the AI’s powerful data analytics and visualization tools. This can include connection to your OneDrive or Google Drive accounts, which can streamline the process of extracting insights from complex datasets and creating customizable charts, making data analysis more accessible and efficient.
While you do not need to pay to access these features, you do need to create a ChatGPT account, and free users will have a limit on how much they can use GPT-4o before ChatGPT switches back to using GPT-3.5. For newer users, there are a few things we want to call out to make it a bit easier for you to jump in.
Creating an account is pretty straightforward — simply go to ChatGPT and click “Sign Up” in the bottom left of your browser window. Once you have your account and have started chatting with ChatGPT, it’s not always obvious which model you’re working with. Here’s how you can tell: underneath every response from ChatGPT is a series of icons. Click on the rightmost that looks like stars (you can see it in the screen capture below) to see which model you’re working with, or to switch between models.
This also allows you to compare and contrast the output between the two models, as switching in the same chat will re-generate ChatGPT’s response using the newly selected model (that said, we rarely use GPT-3.5).
Another new feature for free users is the ability to use custom GPTs: individual ChatGPT chatbots that serve very specific purposes (these are a killer feature of GPT-4o and can immediately make ChatGPT much more helpful). It’s easy, though not terribly intuitive, to access these. The easiest way is to simply click the “Explore GPTs” item in the left sidebar, which takes you to a marketplace of sorts for custom GPTs. Another way is to click a link to a GPT which will open a new chat window in ChatGPT with the custom GPT. It should look like this:
To get started using custom GPTs, you might start by choosing one of ours from our Custom GPT library. You can tell when you’re working with a custom GPT based on the header in the upper left corner. If you’re working with a GPT that you want to come back to repeatedly, simply click the name to open the menu, and select “Keep in sidebar.”
This is only scratching the surface, but we hope it’s enough to get new users started. If you, or others around you, have been disappointed by ChatGPT after working with GPT-3.5, now is the time to dive back in. GPT-4o is a night-and-day difference in capability. The gulf between the paid versions of ChatGPT and the free version has shrunk dramatically, and we expect that newer users will notice the difference.
New Reports on Generative AI Awareness and Adoption
The University of Toronto and McKinsey find that awareness and adoption of generative AI tools are growing, but we’re still in the early stages.
The past few weeks saw the publication of reports from two surveys on the state of awareness and adoption of generative AI. The University of Toronto’s Schwartz Riesman Institute for Technology and Society measured global opinion on artificial intelligence across a broad swathe of categories, while a McKinsey report focused primarily on organizational adoption. Both reports are fulsome (the University of Toronto’s spans 14 categories over 100+ pages) and are worth at least skimming, if not reading in their entirety. Below we share five items that stood out to us from each report, followed by a brief synthesis from looking at them in combination.
The report covers 14 categories, ranging from general attitudes toward AI, to regulation and trust, to opinions on AI’s impact on various sectors of the economy. For us the most relevant section was on opinions on ChatGPT, which we consider a proxy for publicly-available generative AI tools. Here are five key takeaways from that section:
Globally, 63% of respondents report having heard of ChatGPT, and 40% of respondents report having used it. As with most items in the survey, responses varied widely by country: In the U.S. 55% of respondents have heard of ChatGPT and 29% have used it, while in India the levels are 82% and 66%, respectively.
Globally, of those who have used ChatGPT, only 17% report using it daily, 36% use it weekly, and 16% use it monthly.
33% of respondents globally report using ChatGPT for work. In China, that number is 66%, while in the U.S. it is 23%.
Globally, 47% of users responded “Yes” to the question “Do you think ChatGPT could be useful to you when working?” 70% or more of respondents in Kenya, India, and Pakistan answered yes to that question, while under 30% of respondents in the U.S., U.K., Canada, Australia, Germany, and France did.
64% of all respondents indicated that they are likely to use ChatGPT in the next year (compared to 40% who indicated they are using it today). There was once again a wide range across countries. Over 80% of respondents in India, Indonesia, Kenya, Pakistan, South Africa, and China anticipate using ChatGPT in the next year, while fewer than half of respondents in the U.S. (47%), U.K. (45%), and Canada (44%) have similar intentions.
McKinsey’s “The State of AI in Early 2024: Gen AI Adoption Spikes and Starts to Generate Value”
While the University of Toronto survey targeted individuals, McKinsey’s survey focused primarily on organizations. Here’s what stood out to us:
65% of organizations report adopting generative AI in at least one function, up from 33% in 2023.
Marketing and sales are the functions with the highest rates of adoption, followed by product / service management and IT.
55% of individual respondents indicate that they either regularly use AI either at work, outside of work, or both — up from 38% in 2023.
The top two areas of concern and known risks for respondents are inaccuracy and intellectual property infringement, followed by cybersecurity and personal privacy.
23% of respondents indicated that their organization has experienced negative consequences from inaccuracy.
So, what to make of it all (especially for corporate communication leaders and teams)? We offer three considerations. First, individual adoption is all over the board, so develop your strategy accordingly. Many employees are likely “power users” who use AI tools every day (whether they say so or not), while others still aren’t aware of these tools’ existence. Whatever approach you’re taking to use and adoption, you take this full spectrum into account, and take advantage of the knowledge and experience of savvy users while meeting less-experienced users where they are.
Second, the consequences of misuse are real, and you should get guardrails in place if you haven’t already. The risks of inaccurate information and intellectual property exposure aren’t going away anytime soon. Define these risks for your team, and get the right policies and guardrails in place for employees that can mitigate those risks.
Finally, prioritize educating all employees on the technology — and its strengths, limitations, and risks. A striking finding in the McKinsey survey is that those who are getting the most use out of generative AI tools (the “highest performers”) are “are paying more attention to gen-AI-related risks.” One interpretation of this finding is that the better you understand the risks and challenges gen AI presents, the better you can work around them to get the most out of the models. In our own firm we do a lot of work not just to train people on how to use the models, but on how the models actually work. We find that doing so gives people a context in which they can better predict and understand what the models do well and poorly, which helps our people work around and account for the inherent risks and challenges. You should do the same.
Getting Better Output Through Process (Rather Than Prompt)
We offer an example of what it means to treat generative AI as a collaborator.
We work hard to make Confluence a nice blend of news, research, and practical utility at the intersection of generative AI and communication. Much of the commentary on practical utility has been about prompts — what you say to the LLM — and we’ve gone so far as to offer many examples of our own prompts and custom GPTs. And prompt design definitely matters. We’ve been developing a bespoke generative AI application that we hope to share in the future, and much of that work has been about getting the prompt alchemy just right. Four months in we consider it a mysterious, but important, art.
That said, we increasingly believe process is as important, and perhaps even more important, than prompt. Unless your request is very simple (“What is the reading level of this writing?”) or procedural (“Merge these files together and convert the filenames into headers for each entry in the combined file.”), the initial output you get from an LLM is rarely the best it can provide, regardless of the quality of your prompt. The manner in which you work with the LLM to improve that output — we’ve heard people call it “collaboration,” “co-creation,” “acting like a cyborg,” “management” … even “interrogation” — matters. The metaphors, and the advice that often reflects them (like “treat the LLM like an intern”) doesn’t feel very prescriptive or practical, so we thought it might be helpful to offer a real-world example.
One of our authors is trying to improve his chess game. After watching a YouTube video about how to study and improve in chess, he decided to enlist GPT-4o in creating a development plan for his game. That chat thread is below. This is just the user / ChatGPT conversation with the plans taken out to save space (the user is in italics bold and ChatGPT is in italics). We do show the first and final versions of the plan so you can see how the plan evolved.
Adopt the persona of chess master Kostya Kavutskiy.
Sure, I can provide advice and guidance from the perspective of a chess master like Kostya Kavutskiy. What specific aspect of your chess game would you like to focus on? We can discuss openings, tactics, strategy, endgames, or any other area you're interested in improving.
I’m a 700 player and want to go to 1000. I can devote about 30 minutes to chess a day on average and some days can spend an hour. I have a chess.com account. Work with me to build a training plan. First, how many months should it take to go from 700 to 1000?
Moving from a 700 rating to 1000 is a significant but achievable goal with consistent practice. Given your commitment of about 30 minutes daily, it should take around 3 to 6 months to reach a 1000 rating, assuming steady progress and effective training. Here's a structured plan to help you reach this milestone:
Month 1: Build a Strong Foundation
Focus Areas: Basic Tactics, Opening Principles, and Simple Endgames
Tactics Training (10 minutes daily)
Use puzzles on chess.com to practice basic tactics like forks, pins, and skewers.
Aim for 5-10 puzzles each session.
Opening Principles (10 minutes daily)
Learn the fundamental principles: control the center, develop pieces, and ensure king safety.
Focus on a few simple openings for white (e.g., Italian Game) and black (e.g., Sicilian Defense, Caro-Kann).
Endgames (10 minutes daily)
Practice basic endgames such as king and pawn vs. king, and understanding opposition.
Use chess.com’s endgame drills.
Month 2: Intermediate Tactics and Positional Play
Focus Areas: Intermediate Tactics, Positional Understanding, and Common Middlegame Plans
Tactics Training (10 minutes daily)
Increase difficulty of puzzles.
Focus on pattern recognition and solving puzzles quickly.
Positional Understanding (10 minutes daily)
Watch videos or read articles on pawn structures, weak squares, and piece activity.
Review games by strong players to understand their plans and strategies.
Middlegame Plans (10 minutes daily)
Study common plans in your preferred openings.
Practice identifying and executing these plans in your games.
Month 3: Analyze Your Games and Study Master Games
Focus Areas: Game Analysis, Studying Master Games, and Advanced Tactics
Analyze Your Games (15 minutes every other day)
Review your recent games on chess.com.
Identify mistakes and missed opportunities.
Learn from these to avoid repeating the same errors.
Study Master Games (15 minutes every other day)
Watch annotated games or use a chess engine to understand master-level play.
Focus on games that use your preferred openings.
Advanced Tactics (10 minutes daily)
Continue solving more complex puzzles.
Aim to understand the underlying concepts behind each tactic.
Month 4: Advanced Opening Study and Practical Play
Focus Areas: Deepening Opening Knowledge, Practical Play, and Specialized Endgames
Deepening Opening Knowledge (10 minutes daily)
Study deeper lines and common traps in your preferred openings.
Watch videos or read opening books.
Practical Play (20 minutes daily)
Play longer time control games (e.g., 15|10) to apply what you've learned.
Focus on playing thoughtfully and applying your training in real games.
Specialized Endgames (10 minutes daily)
Study more complex endgames (e.g., rook vs. pawn, bishop vs. knight).
Practice these endgames using drills on chess.com.
Month 5-6: Consolidation and Continuous Improvement
Focus Areas: Regular Play, Continuous Learning, and Mental Toughness
Regular Play (20 minutes daily)
Continue playing longer time control games.
Focus on consistency and reducing blunders.
Continuous Learning (10 minutes daily)
Keep learning new concepts in strategy, tactics, and openings.
Watch high-quality instructional videos and read chess books.
Mental Toughness (10 minutes daily)
Practice staying calm under pressure and maintaining focus.
Reflect on your mindset during games and work on staying positive and resilient.
By following this plan and adjusting it based on your progress and feedback, you should be well on your way to reaching a 1000 rating within the next 3 to 6 months. Remember to enjoy the journey and stay motivated!
Let’s make it a three month plan. Each day I want to focus on either playing, studying, or calculation / solving. First give me the list of subjects to focus on across the three months and how many weeks I should spend on each.
Absolutely! Here is a three-month plan where each day is dedicated to playing, studying, or calculation/solving…
Revise. Take out rapid games. Any game I play will be a longer game. Presume that each day I will only spend time on play, learn (via reading or video learning), or calculation / solve. Presume that I will decide which of those (play study solve) to do each day of the week rather than making them specific to certain days. So I want to know what’s the learning focus for the week that I should apply to all three: play, learn, and solve.
Got it. Let's revise the plan to ensure that each week has a specific learning focus that you will apply to your playing, learning, and solving activities…
Get more specific. Give me the plan for each week. If something should extend over multiple weeks make it so. And get much more specific about the things I should work to learn about and practice for each week.
Certainly! Here is a detailed week-by-week plan with specific learning objectives and practices for each week…
Language like this: “Study basic tactical motifs such as forks, pins, skewers, and basic checkmates (e.g., back-rank mate, smothered mate)” is too general. Revise to take out anything like “such as.” Knowing that I am a 700 level player focus on the specific ideas, principles, and tactics I should work and learn about each week.
Got it. I'll revise the plan to provide more specific and actionable items for each week, focusing on the key ideas, principles, and tactics appropriate for a 700-level player…
Get more specific. “Middle game pawn breaks like…” is not helpful to me. List EXACTLY what I should study or practice every week of the plan. Select the ideas and tactics most central to me going from 700 to 1000.
Understood. Here is a more detailed and specific week-by-week plan, focusing on the exact ideas and tactics you should study and practice to progress from a 700 to a 1000 rating…
Great. Replace week 7 and instead assume I will do game analysis after every game I play.
Sure, I'll replace Week 7 with a new focus, assuming you'll be doing game analysis after every game you play. Here's the updated plan…
Great. Now assume the persona of a world class chess instructor who’s particularly good at helping players like me go from 700 to 1000. How would you improve this plan?
As a world-class chess instructor specializing in helping players progress from 700 to 1000, I would refine the plan to ensure it is comprehensive, engaging, and adaptable to your progress. Here are some improvements to the plan…
Now in that same persona search the internet and your training knowledge and build a comprehensive list of resources based on the plan (including the checklists you mention etc.). If you can’t find any, list them so we can work together to build them.
(ChatGPT searched 3 sites online before responding.) Here is a refined and comprehensive list of resources based on your training plan. These resources will help you focus on the specific ideas, principles, and tactics for each week…
Assume the same expert persona. Review and improve the plan. List your changes so I know them.
Here is the improved plan with specific changes and additions to enhance its effectiveness.
Great. Now output the plan again so everything is nice and tidy.
(ChatGPT outputs the plan.)
Did you intend to have a fork resource in week 2?
Here is the revised and tidy plan without redundant resources …
(We interrupt ChatGPT because it hasn’t fixed that problem.) You have a fork training link in week 2 which is about opening principles. I think this is a mistake.
(ChatGPT outputs the plan with that mistake fixed and a new resource added … this is the final version.)
Month 1: Foundation Building
Week 1: Forks
Play: Continue playing longer games (30|0 or 45|45) focusing on spotting and creating forks.
Learn:
Watch the video "Forks! Beginner chess tactic explained!" on YouTube.
Read the article "Ultimate Beginners Guide to Chess Forks" on Chess.com.
Solve: Solve 10-15 fork puzzles daily on chess.com. Use the "puzzle rush" feature to practice spotting forks quickly.
Additional Resource: Forks training on OCF Chess for a broader perspective on tactics and application.
Week 2: Opening Principles
Play: Focus on applying opening principles in your games: 1.e4 or 1.d4 for white, 1...e5 or 1...c5 for black.
Learn:
Study the Italian Game (1.e4 e5 2.Nf3 Nc6 3.Bc4) and the Sicilian Defense (1.e4 c5).
Watch the video "Italian Game: Opening Principles for Beginners" on YouTube.
Solve: Solve opening-related puzzles on chess.com, focusing on development, king safety, and early tactical opportunities.
Week 3: King and Pawn Endgames
Play: Aim to reach king and pawn vs. king endgames in your games.
Learn:
Study the concept of opposition and how to convert king and pawn vs. king.
Watch "Basic Endgames: The Key Opposition" on Chess.com.
Solve: Use endgame drills on chess.com to practice king and pawn vs. king scenarios.
Week 4: Open Files and Rook Activity
Play: Focus on controlling open files with your rooks in your games.
Learn:
Study the importance of open files and how to double rooks on the open file.
Read "The Open File - Chess.com" article.
Solve: Solve puzzles that involve placing rooks on open files and controlling key squares on chess.com.
Month 2: Intermediate Concepts
Week 5: Pins
Play: Actively look for opportunities to pin your opponent’s pieces.
Learn:
Study how to pin pieces with bishops and rooks.
Watch "Chess pins, skewers, and forks explained" on YouTube.
Solve: Solve 10-15 pin-related puzzles daily on chess.com.
Week 6: Pawn Breaks
Play: Apply specific pawn breaks in your games, such as d4-d5 in the Queen's Gambit or f4-f5 in the King's Indian Attack.
Learn:
Study the pawn break d4-d5 in the Queen's Gambit and f4-f5 in the King's Indian Attack.
Read "Pawn Breaks in Chess" article.
Solve: Solve puzzles that involve executing these pawn breaks on chess.com.
Week 7: Double Attacks
Play: Focus on creating double attacks in your games.
Learn:
Study double attacks, particularly those involving knights and queens.
Solve: Solve puzzles that specifically involve double attacks on chess.com.
Week 8: Master Games - Opening to Middlegame Transition
Play: Mimic strategies and plans from master games.
Learn:
Study annotated master games focusing on the transition from opening to middlegame.
Use resources like "Chess.com’s Game Explorer".
Solve: Solve puzzles from positions in these master games on chess.com.
Month 3: Advanced Study and Practical Play
Week 9: Deeper Opening Study
Play: Implement deeper lines and traps in your preferred openings.
Learn:
Study detailed lines and common traps in openings like the Fried Liver Attack and the Dragon Variation.
Use resources from Chess.com’s opening database.
Solve: Solve opening-related puzzles and traps specific to these lines on chess.com.
Week 10: Lucena and Philidor Positions
Play: Practice converting rook and pawn vs. rook endgames.
Learn:
Study the Lucena position (building a bridge) and the Philidor position (defensive techniques).
Watch instructional videos on Chess.com’s endgame lessons.
Solve: Use endgame simulators and drills focusing on these scenarios on chess.com.
Week 11: Discovered Attacks
Play: Focus on creating discovered attacks during your games.
Learn:
Study discovered attacks and how moving one piece reveals an attack by another.
Watch instructional videos on discovered attacks.
Solve: Solve puzzles involving setting up and executing discovered attacks on chess.com.
Week 12: Review and Reinforcement
Play: Continue playing longer games, applying all the concepts learned.
Learn: Review all key concepts learned over the past three months using summaries and refresher courses.
Use summaries and refresher courses on Chess.com.
Solve: Solve a variety of puzzles to reinforce overall tactical and strategic understanding on chess.com.
Additional Checklists
Forks Checklist:
Look for opportunities where a knight or queen can attack two pieces simultaneously.
Utilize resources from OCF Chess for additional insights and exercises.
Pins Checklist:
Identify situations where a piece can pin an opponent’s piece to a more valuable piece or the king.
Watch videos and read articles on effective pinning strategies.
Pawn Breaks Checklist:
Regularly assess pawn structures to find potential breaks like d4-d5 or f4-f5.
Practice pawn break puzzles and scenarios.
Discovered Attacks Checklist:
Visualize moves where moving one piece can uncover an attack by another piece.
Solve puzzles specifically designed to practice discovered attacks.
By utilizing these resources and checklists, you’ll have a structured and comprehensive path to improving your chess skills and achieving your goal of reaching a 1000 rating. If you need further customization or additional resources, feel free to reach out for more specific guidance.
The entire exchange took maybe 10 minutes (and the author is now busy reading about and practicing forks). While it’s just chess and not “meaningful” work, you could apply a similar approach to more topics and types of work than we can think to list. It’s an example of a deliberate, iterative process for engaging an LLM. What principles does this example reflect? Several:
Assume the LLM can help you rather than assume it can’t help you (and if you wonder if it can help you, just ask it). There was no hint that ChatGPT could do this. The user simply thought, “I’d like a training plan. I wonder if ChatGPT can make one for me?”
Don’t accept the first output (or the second, or the fifth). Give the model feedback and make it revise — in the example the feedback was about the length of the program, the specificity of the output, mistakes to fix, etc. Just keep telling the model what you do and don’t want as it improves the output over time.
You don’t need a plan. This is perhaps a corollary to the second principle, but you don’t really need to know where you’re going when working with an LLM (although you can always ask the model to create a plan for what you want, and this can be helpful). But often it’s good enough just to work step by step, asking for changes or additions or embellishments chat-by-chat as your intuition and needs evolve.
Ask the model to review its own work. We almost always ask the model to review its own work. Sometimes we even ask a different model to review a model’s work (giving ChatGPT output to Claude, for example). In this example we asked the model to adopt a slightly different persona in reviewing its own work, but that’s not necessary. The model will always find things to improve or change when reviewing its own output, usually for the better.
Point out errors and weaknesses and have the model address them. LLMs make mistakes. Be on the lookout for them. When you see them, point them out. If the model has access to the internet (GPT-4o does), ask it to fact-check its work. Ask it to assume the role of a critic and have it judge its own work. And if anything you’re having the model do has reputational consequences, you must have a human review the work for accuracy.
Getting the most from LLMs does require skill in prompt design, although we think that as models become more capable, prompt engineering for most tasks will likely become less important. But as collaborators, LLMs are probably here to stay, and it’s not too early to start thinking of them as a new member of your team. And as in most collaborative relationships, how you work together can be pretty predictive of what you create together. Keep working on the process you use to engage generative AI and the quality of your work will only benefit.
We’ll leave you with something cool: Veo, Google’s video generation model.
AI Disclosure: We used generative AI in creating imagery for this post. We also used it selectively as a creator and summarizer of content and as an editor and proofreader.