Summary: Jakob live on Internet Sept. 19 | AI helps education but teachers are poorly supported | Bill Gates thinks AI is easy | Midjourney visual styles library | Comparing the image quality from Flux Pro vs. Schnell
UX Roundup for September 3, 2024. (Ideogram) Last month, I asked whether you could get a flock of flamingos to stand in a lake such that they spell “UX.” This week, see if you can arrange rose petals to spell “UX Roundup.” Ideogram still has the best typography, but other image generators struggle with such prompts.
Hold the Date: Jakob Speaking September 19
I will be speaking about The Impact of AI on Design at ADPList’s BeMore Festival on Thursday, September 19. Registration required, but in keeping with ADPLists’ mission (and my mission) to take design knowledge to the world, the registration fee is laughably cheap for anybody in a rich country and still affordable for most people in poor countries: only US $15.
Plus, as I mentioned in my newsletter for August 26, the registration fee includes a number of other exciting sessions at this conference. This is s fully remote conference, running Sept. 18-19, with all sessions live on the Internet.
My session is live at 9:00 AM USA Pacific time. Check the time zone calculator for the equivalent time in your location. To save you a click, here’s that time for a few key cities:
San Francisco, CA: Thu, Sep 19, 2024, 9:00 am
New York, NY: Thu, Sep 19, 2024, 12:00 noon
São Paulo: Thu, Sep 19, 2024, 1:00 pm
London: Thu, Sep 19, 2024, 5:00 pm
Paris: Thu, Sep 19, 2024, 6:00 pm
Dubai: Thu, Sep 19, 2024, 8:00 pm
New Delhi: Thu, Sep 19, 2024, 9:30 pm
Singapore, Hong Kong, Beijing, Fri, Sep 20, 2024, 12:00 midnight
Tokyo: Fri, Sep 20, 2024, 1:00 am
Sydney: Fri, Sep 20, 2024, 2:00 am
I’m speaking to the world’s UX freaks on Friday, September 19 (early in the night on Sept. 20 in East Asia/Australia, sorry.) Midjourney.
AI Helps Education but Is Poorly Supported by the Educational Establishment
For a long time, I've believed that AI holds immense potential in two crucial areas — education and healthcare. This potential is particularly significant in developing countries, where AI can play a transformative role.
The Walton Family Foundation (the family behind Walmart, which is the largest shopping chain in the United States) has released a study of the value of AI in today’s classroom, based on a survey of school teachers and students plus university students. The study participants were in the USA, which doesn’t align with my take that the greatest potential for AI to advance humanity is in poor countries. On the other hand, the U.S. is currently the most advanced country concerning AI, so it may be the best place to conduct forward-looking research. This study included slightly more than 1,000 respondents in each of these 4 groups: teachers, K-12 students, undergraduate university students, and parents.
The study's main finding is a stark reminder of the current situation:
While educational AI use has increased from 2023 to 2024,
most schools and colleges lack a clear policy on AI,
are not providing the necessary teacher training,
and fail to meet the demand of students interested in AI-related careers.
This is a concerning trend that is causing our education system to fall further behind.
Teachers are not getting the continuing education they need to build AI skills. (Midjourney)
A related finding is that the AI vacuum in school policy means it is currently used “unauthorized,” while people want policies that encourage AI. Kids, parents, and teachers are figuring it out on their own, without express permission, whereas all stakeholders would rather have a policy that explicitly encourages AI in education.
AI is currently being used at least once a week by approximately half of all the stakeholders in the study: teachers (46%), school (48%) and university (46%) students, and the parents (51%). I doubt much should be made of the small differences in these percentages, but it’s interesting that the parents who probably mainly have jobs in business score the highest, whereas the teachers score lower than their students.
Attitudes toward AI are best among undergraduate students, with 75% having a favorable view of AI. K-12 students are almost as good, with 70% having a favorable view. Interestingly, teachers lack behind, with only 59% having a favorable view of AI. A majority, yes, but could do better.
On the question of whether AI should be banned, allowed, or encouraged for schoolwork, most take the middle view that AI should be allowed, but not encouraged: 66% of teachers want to allow AI, but only 23 percentage points out of those 66% actively encourage AI.
How do teachers currently use AI? The main use is for creativity, echoing the general finding that AI is a superb ideation tool. Here are the top 7 uses:
Helping with creative ideas for classes (37%)
Lesson planning and preparing instructional materials (32%)
Creating worksheets or examples for students (32%
Creating quizzes or tests for materials taught in class (31%)
Detecting students’ attempts to cheat or plagiarize (25%)
Building your own background knowledge for lessons and classes (23%)
Writing notes to students, parents, and colleagues (23%)
Many teachers use AI to help them generate creative ideas for classes or make lesson plans. (Midjourney)
The top answers given by the teachers when asked what positive impacts AI has had:
Students learn more and faster (35%)
Student engagement (10%)
Makes teaching easier (10%)
Improves creativity (5%)
Percentages are given relative to those teachers who are using AI in their classrooms. I am disappointed at the low scores for improving creativity. Since AI does have great creativity potential, this indicates to me that these teachers have not received proper training in AI. In fact, only 24% of public school teachers (and 25% of private school teachers) say they have received training or professional development on using AI in the classroom.
AI is currently underused for sparking student creativity. (Midjourney)
The students gave their extremely high ratings to “AI’s ability to provide explanations to complicated concepts I struggle with in school,” with 80% of K-12 students agreeing and 84% of college students agreeing with this statement. There’s a slight difference between the two levels of students regarding their main use of AI. K-12 students are most likely to have used AI chatbots to write essays and other assignments (56%) and study for tests and quizzes (52%). The university students swapped the rank of these two uses, with 61% using AI while studying for tests and quizzes and 56% using it while writing essays 56%.
Students mainly use AI to help with homework and to study for tests. (Midjourney)
Bill Gates Thinks AI Is Easy
I am unreasonably happy to have found a reason to disagree with Bill Gates. It’s one of the regrets of my career that I never got a chance to debate him. While I’m sure that Bill Gates has a higher IQ than me, in my dreams, I would still have prevailed in a debate on whether fancy features or UI simplicity is better. In his heydays, Gates constantly pushed for more features at the cost of UI complexity and lower usability, whereas I would have advocated for simplicity.
I have had less reason to take issue with anything Bill Gates has said after he retired from Microsoft. He’s become more reasonable with age. (Or maybe I have become less likely to take offense at other people’s statements.)
Now, finally, a reason to quarrel. The “Next Big Idea Club” published an interview with Bill Gates about his views on long-term trends in AI (YouTube, 57-minute video). Honestly, most of what he says is great.
However, when asked about difficulties and possible delays in the broad uptake of AI, he marginalizes the usability concerns. Gates basically says that since you prompt AI with natural language, anybody can do it with minimal learning.
Yes, it’s true that you don’t need weird syntax when talking to AI. It’s also true that with voice I/O you don’t need to type — or even be able to read — to use AI. It’s even true that minor languages are being added at a reasonable clip, so that people who speak, for example, Icelandic (about 300K native speakers) can talk with AI in their native language.
This makes AI easier than DOS, but it doesn’t make it easy.
The user still needs to know what to say and how to articulate the prompt to achieve what he or she wants. There’s a substantial articulation barrier in talking to AI. This is particularly noticeable for low-literacy users who struggle to formulate complex thoughts in prose. But even for a highly literate person like myself, it’s often hard to put specific words to ideas that go beyond traditional thinking. We want to use AI for new things, so we must be able to use words that cover new domains. For example, it’s quite hard to prompt music-generating AI services to create the style of songs you want — unless you’re a music theory major or a professional musician, but the promise of AI-supported music creation was to bring music-making to the masses.
It's even more challenging to prompt video-generating AI to create great movies. We must start thinking about positioning and moving the camera in 3D space. Yes, professional Hollywood cinematographers and movie directors have the vocabulary and 3-D spatial reasoning skills to communicate these concepts. But the average user surely doesn’t.
The UI for any advanced AI service must be dramatically more straightforward to use than today’s offerings. Otherwise, we won’t get broad adoption.
How do we direct a movie-making AI to capture the film we want? There’s a serious articulation barrier in these prompts for anybody who’s not a professional Hollywood director. (Ideogram)
Midjourney Visual Styles Library
While on the topic of the articulation barrier, a tough barrier for many is prompting for visual style. Yes, those with a decent classical education can use keywords like “impressionist oil painting,” but that’ll only take you so far.
The “MidLibrary” has more than 5,000 visual styles documented. Usually with at least 4 images generated with Midjourney after specifying that style. Most of these styles I had never even heard of, but they make for interesting experimentation. Here are four illustrations of a library with leatherbound books, using styles from MidLibrary:
Clockwise from top: Intaglio, Risograph, Distressed Photo, Stencil Graphic. I have to say that the point of a “distressed photo” is that the photo is distressed, not that the subject itself is distressed, so I don’t think Midjourney interpreted the prompt correctly. But it did generate an interesting alternative view of the concept of a collection of leatherbound books. (Midjourney)
Flux Pro vs. Schnell: Image Quality Compared
The new image-generation tool Flux has become extremely popular since it launched only a month ago. It is now offered by a very wide range of AI services, from Perplexity to Grok. Unfortunately, many of these services don’t disclose which of the 3 Flux models they use: Pro, Dev, or Schnell. Pro is the biggest and most expensive to run, but also generates the best images, as we shall see. Dev is in the middle, and Schnell is a fairly small model that’s fast (thus its German name) and cheap to run, but also generates less-good images.
Umesh (the creator behind the Story Illustrator AI service I have used to make several comic strips for this newsletter) has published a series of benchmark comparisons where he used the same prompt for Flux Pro and Flux Schnell.
Reviewing these images makes it abundantly clear that you should use Flux Pro to create great images. Some people call Flux the “Midjourney killer,” which I don’t actually think it is, but Flux Pro does generate images that are close to Midjourney 6.1 in quality. Flux Schnell is closer to Midjourney 5 — in fact, the Schnell images would have been considered quite good a year ago. But because of the fast progress in AI image generation, Schnell is no longer acceptable.
I used Poe (which offers all 3 Flux models with impeccable disclosure) to generate two different images with both the Pro and Schnell versions of Flux. Poe charges 1,250 credits for each image made with Flux Pro and only 75 credits for each Schnell image generation. Assuming these charges are fair, it’s almost 17 times as expensive to run Pro than Schnell.
The first pair used the same prompt I’ve been using to test Midjourney: a group of 3 K-Pop idols dancing in a TV production. For the sake of the experiment, I also reused one of Umesh’s benchmark prompts: “A vibrant, cartoon-style illustration of a whimsical forest filled with colorful, anthropomorphic animals. The creatures have exaggerated, expressive features, such as big eyes and wide smiles. The trees and plants are stylized with bold outlines and bright, saturated colors. The scene should be playful and lively, with a focus on the character design, dynamic poses, and a fun, cheerful atmosphere typical of cartoon illustrations.”
Comparing the low-end Flux Schnell model (left column) with the high-end Flux Pro model (right column). Top row: my prompt. Bottom row: Umesh’s prompt.
The Pro image is substantially more compelling for the cartoon animals. The difference is less evident for the K-Pop performance, but I do think the Pro version seems more, well, professional. In the Schnell version, compare the leg positions of the left and the right dancers: certainly, JYP would never have let an unsynchronized dance like that represent his company. On the other hand, I prefer the straightforward lighting in Schnell’s version over the more “artistic” backlighting in Pro’s version. (For better lighting, either reissue the prompt a few more times, or change the prompt to specify front lighting.)
On balance, this small experiment confirms Umesh’s more significant benchmark exercise: Pro is better. Is it 17 times better? Probably not, so if somebody had to generate millions of images on a budget, I can see why they might use Schnell. Poe’s credit points cost $16.67 per million credits, meaning that each Pro image costs 2.1 cents, whereas each Schnell image costs 0.1 cents, so you do the math, depending on your application.
As a final point, kudos to Flux for getting the number of K-Pop idols right in every image, even the cheap one. Midjourney still often gives me more than 3 dancers, even when the prompt explicitly states that I want a 3-member group. (It’s an open secret that these image tools are trained on pictures from the Internet, and 3-member K-Pop groups are rare: Perplexity lists 12 three-idol groups, of which I had only heard of one. Thus, the training set will include overwhelmingly more images showing four or more dancers.)