Marketing BS Briefing:Prompt Engineering Edition
I am back from vacation. Kids were great traveling through Porto, and a mess traveling through Paris. We all made it hope safe and sound. I have a back-log of links to share (today), and a few half-written essays to finish (in the coming weeks). Be sure to get to the prompt engineering content in the AI section. That stuff is going to be an important marketing skill in the near future. Enjoy:
Marketing
Apple: The WSJ published a scoop showing that in the years leading up to iOS 14.5 Apple was negotiating with Facebook to take a share of Facebook’s revenue generated from ads on iphones. One more data point on the side arguing Apple’s “privacy crackdown” was never about privacy. Privacy was the PR spin used to get acceptance for their revenue grab. Related: Matt Mullenweg (Automattic’s CEO) shared that Apple blocked “boosted posts” on Tumblr until the company agreed to give Apple 30% of the revenue generated. Now Apple is adding more ad placements into their app store. Also related: Amazon is buying iRobot. This has nothing to do with privacy invasion, but that hasn’t stopped journalists on Twitter from sounding the alarm. Amazon really does not care what your home layout looks like… Also: Have these people arguing about privacy ever used a Roomba? “Roomba reporting to Amazon; This house is 2-inches by 3-inches; Either that or I am stuck under the table again; Just to be safe, let’s stop recommending the large packages of toilet paper”.
Uber: Uber is shutting down their “earn and burn” loyalty program (but keeping their membership program). Not surprising. Traditional loyalty programs usually destroy value (even though it is sometimes hard to see), and Dara (Uber CEO), although he comes from a finance background, understands that (In one of my few meetings with Dara at Expedia I walked him through the economics of loyalty programs and how they are over-rated. I expect he did not forget). See my essay on Starbuck’s loyalty program for more.
McDonalds: The QSR is dropping salads from their menu, and many customers are upset. Salad sales at McDonalds were always very small, low margin and operationally complicated. Why did McDonalds have them at all? The original rational was to “stop objections.” Everyone in the family wants to go to McDonalds, but it has nothing for mom, who does not want a burger. McSalads were the answer. Other restaurants have this same problem but chose not to solve it (What if no one likes fried chicken? Or a specific ethnic food?). Solving the “objection” problem only matters if you are trying to significantly grow your TAM. McDonalds got so big they wanted to makes sure they appealed to everyone - so they had a shot at winning any meal occasion with the family. Canceling salads eliminates that ambition and potentially slows down growth (or potential growth). Presumably they are hoping the decreased product complexity makes up for the reduced TAM. The question is, what is McDonalds going to do with their newly freed up capacity?
Lyft: Lyft is building an advertising network. Of course. Why not. Monetize attention. Without much purchase history (other than rides) I expect targeting will be poor, but it is incremental, so why not?
Customers: They really don’t care about you and your brand. I stumbled across this 2013 essay from Benedict Evans that summarized it nicely, “your customers' relationships with you are the only relationships you have as a business and you think a lot about them. But you're one of a thousand things your customer thinks about in a week, and one of dozens of businesses. And they probably have their own ideas about how they want to engage with you (though they wouldn't put it in those words) - assuming they think about you at all.” Maybe someday marketers will internalize this….
Disney: Disney+ is launching ads on December 8th. The new ad-supported service will cost the same as the existing ad-free service ($7.99/month), while the ad-free offering will have a price increase to $10.99. Also: Hulu is now getting more sign-ups than Disney+. This is the lake/river problem. When a service first launches, it sucks up all the existing demand that was “just sitting there for the taking” (i.e., parents that want to give their kids unlimited Disney). After that initial burst, companies are forced to go after the flows rather than the stock (i.e., the new potential customers that are flowing down the rivers into the lake). Disney+ had a bigger lake, but Hulu has more river potential.
Walmart: Walmart has complete negotiations with Peacock to add the streaming service to their Walmart+ subscription service, making it look more and more like Amazon Prime. Recall that good product bundles combine fan but not super-fan overlap. See my essay here.
Shared universes: Disney has shown the power of the Marvel Cinematic Universe. Many other studios are trying to replicate that success with their superhero-like series. But my take is that the most important component here is the shared-universe NOT the superheroes. So why not create film series on things like the making of the atomic bomb, the founding of America or the French Revolution? More here.
Product Placement: Cinnabon is promenently featured in the TV series “Better Call Saul” — and sometimes in a not-so-flattering way. That has not stopped the firm from capitalizing on it. Fast Company has the details.
Marketing to Employees
Vilified Industries: The Economist asks, “why do people agree to work in vilified industries like tobacco and firearms?” They show that employees at these firms are paid a little more than comparable jobs in non-vilified industries, but that in general employees have rationalized that their companies are "doing good”. Some quotes: “But firms under fire are practised at turning the negative effects of their products to their advantage. Energy firms argue that the money they make from oil and gas today enables them to fund the transition to low-carbon energy tomorrow. Diageo, a drinks firm, highlights its programmes to encourage drinking in moderation. Tobacco firms peddle cigarettes even as they endeavour to soften the harm caused by smoking: British American Tobacco says that its purpose is to “build a better tomorrow by reducing the health impact of our business”. and this: “A study by Mr Roulet found that job satisfaction increased at firms that faced disapproval, provided their employees regarded the criticism as illegitimate.”
AI
Prompt Engineering: It is becoming more and more clear that, as powerful as the new text and image generating AIs are, there is a real skill in crafting prompts. The difference between a “good” and a “bad” prompt can dramatically change the output. A great example is Joy Zang’s attempt to use Dall-e to create a Llama dunking a basketball. Her first prompt was simply, “llama playing basketball” and resulted in clipart images like this:
Through many iterations (and $15 spent on Dall-e credits), she worked her way to the prompt, “Dramatic photo of a llama wearing a jersey dunking a basketball like Michael Jordan, low angle, wide shot, indoors, dramatic backlighting, high detail.”:
In the process she found a number of common “mistakes” Dall-e makes if the prompt is not properly designed, including:
Composition: Dall-e knew a dunking llama needed to have a ball, a basket and a llama, but was generally terrible at knowing where to place the three objects
“Photoshop”: Unless Dall-e’s database has good existing examples of the type of image you are generating, it tends to default to cartoons, and even adding prompts like “realistic” just creates realistic images “photoshopped” into place
Faces: Dall-e struggles to create realistic faces. This may have been to prevent abusers of the technology from creating “Deep Fakes”, but the result is that even llama faces end up getting mangled
Angles: You can ask Dall-e to change the angle of the “camera” that creates the image, but it often gets confused and ignores the request. Joy exmplains: “No matter how many variants of ‘in the distance’ or ‘extreme long shot’ I used, it was difficult to find images where the entire llama fit within the frame.”
Spelling: Dall-e can’t spell at all, which means that any image that need text needs to be edited by a human (although it knows enough to put letters where they belong,
Over prompting: To create something great usually requires a fairly long, detailed prompt. But often if you add too many qualifiers Dall-e gets confused. When Joy started asking for a “llama wearing a jersey” some resulting images removed the llama entirely. Adding the prompt “fluffy llama” seemed to “break” the output entirely. Joy’s conclusion: “In working with DALL·E 2, it’s important to be specific about what you want without over-stuffing or adding redundant words.”
Thumbnails: Don McKenzie at Deephaven used Dall-e to generate images for 100 different blog posts, and summarizes what he learned:
Prompt Engineering is hard. It takes a lot of tries to get what you are envisioning
Experience matters: He estimates it took him 6-7 tries to get something acceptable when he started. After creating a hundred blog images he now believes he can generate something he likes with 2-3 tries (at $0.13 per attempt, that means ~$100 or so in spend to double or triple his effectiveness)
Style modifiers are essential: Without them most results are relatively bland. He recommends, “still from [visually stunning movie”, a visual aesthetic [list here], or a famous artist.
Use Reddit: He recommends the Dall-e subreddit which has many examples of images and the prompts used
Use Photoshop: If you get something close, fix the details with photoshop - especially any gibberish text that Dall-e decided to add. You can also upload the Dall-e created images back into Dall-e and ask it to make modifications
Accept the limitations: Dall-e is not good at math. If you ask for 1-3 things the tool usually does fine, but more than that and you are unlikely to get what you want. Accept that you can’t get Dall-e to create exactly twelve turkeys.
Disruption: Based on the currently limitations of the tool he argues that artists and designers are not going away any time soon. But stock-images sites might be. If you need a generic picture of a shark, the old way to do it was to buy it from Getty. Now you can just ask Dall-e to do it and you can iterate through a dozen different prompts for under $2.
Slide Decks: Killer app- using Dall-e to create images for slide deck presentations. How long before it is built into Powerpoint?
Prompt Engineering Guides: You can learn anything on the internet and prompt engineering is no exception.
PromptBase is a marketplace for prompts. For $1.99 you can buy a prompt that generates “predictable” results on specific AI platforms (the startup takes a 30% cut)
Prompt Book: A guide to using DAll-e
Simple guide: A Google doc, simple guide to using Dall-e prompts
Self-improving: Davis Blalock shares a paper on Twitter about how AI models can learn to code better. The general way to create code with language models is to prompt it with a natural language request. However programers have found that you can instead just prompt with a “programing puzzle” and then check to see if the resulting code solves the puzzle. What is powerful about that technique is that AI can also generate the puzzles. So instead of being limited by training data and human assessments, AI can now create their own puzzles, solve the puzzles, and then assess how effectively they solved the puzzle in order to make iterative improvements on their coding. This should allow for rapid improvements in performance that are not possible when “humans become the roadblock”.
Advertising: AdWeek asks, “Is AI Generated Art Good Enough for Ads?”. The links above kind of answer this. One thing I will add: Often advertising is about signaling. Historically high end images were expensive. Putting high end images into an advertisement signaled product quality (and a company’s commitment to the product and brand). If high end images get cheap and easy to make, firms will need to find other ways to signal that quality commitment. There are always ways to do that…
Careers
Legibility: I wrote about the upsides and downsides of legibility last year (I will get to Part 3 one of these weeks). Adam Mastroianni in his Substack writes about two kinds of intelligence, the ability to solve “Well defined problems”, and the ability to solve “poorly defined problems”. He argues that our IQ tests solely focus on measuring the first kind of intelligence, but it is the second kind of intelligence that makes us “happy” (i.e., “how to I live a fulfilling life” is a poorly defined problem). I am not convinced on the happiness angle, but his essay is worth reading as “well defined” and “poorly defined” are very similar concepts to legible and illegible, and “making things more legible” is an important macro-trend we are all living through right now. related: Russ Roberts has a new book on how to solve “Wild Problems” (i.e., illegible ones)
Travel: New paper shows that employees (or at least basketball players) are more productive when they travel east (vs west). Do with this as you will.
Fun
How things are manufactured: My new favorite Twitter account (link is to thermoforming of suitcases, but there are hundreds of old videos that are mesmerizing)
Keep it simple,
Edward