Summary: Quiz about the 4 AI metaphors | Traditional text entry fields are best | Speaking in Oslo | Usability poster | Interpreting the Elo scores for AI models | Lower cost to orbit
UX Roundup for October 28, 2024. Happy Halloween from UX Tigers! (Midjourney)
Quiz: The 4 Metaphors for Working with AI
Here are 10 questions (plus a bonus question) about my recent article, “The 4 Metaphors for Working With AI: Intern, Coworker, Teacher, Coach,” to help you check your understanding of these important concepts.
The answers are at the end of this newsletter. No peeking!
Q1: What is Jakob Nielsen’s main criticism of the “AI as an intern” metaphor?
A) It is demeaning to interns.
B) It does not accurately reflect the capabilities of modern AI.
C) It is not useful for any tasks.
D) It was only relevant in the 1990s.
Q2: What is the key driver of AI’s increasing capabilities?
A) Moore’s Law
B) The AI scaling law
C) Quantum computing
D) The development of new programming languages
Q3: Which task does Jakob Nielsen believe will likely remain beyond the capabilities of AI until at least 2027?
A) Analyzing qualitative usability studies
B) Mass-analyzing qualitative data
C) Tracking content strategy metrics over time
D) Turning qualitative data into quantitative data
Q4: What are the two forms of tasks AI can perform as a coworker?
A) Teaching and coaching
B) Research and development
C) Independent assignments and collaboration
D) Data analysis and creative brainstorming
Q5: What is the best approach for design ideation with AI?
A) Let AI handle the entire process.
B) Collaborate closely with AI, alternating who takes the lead.
C) Use AI only for generating initial ideas.
D) Use AI only for refining human-generated ideas.
Q6: In the teacher metaphor, what is AI’s primary role?
A) Performing the task for the user.
B) Providing entertainment and distraction.
C) Evaluating the user’s performance.
D) Guiding the user in acquiring new skills and knowledge.
Q7: How does Jakob Nielsen suggest using AI to learn a new subject effectively?
A) Rely on AI to create a single, perfect explanation.
B) Focus on memorizing facts and figures provided by AI.
C) Start with complex explanations and work your way down to simpler ones.
D) Ask for simplified explanations, alternatives, and examples until you understand.
Q8: What historical figure does Jakob Nielsen use to illustrate the benefits of one-on-one instruction?
A) Albert Einstein
B) Alexander the Great
C) Leonardo da Vinci
D) Marie Curie
Q9: What is the primary benefit of using AI as a coach to critique your work, even if its overall performance is inferior to yours?
A) It can help you to automate tedious tasks.
B) AI can often identify errors that humans miss.
C) It forces you to justify your decisions and consider alternative approaches.
D) It can provide you with positive reinforcement.
Q10: What is Jakob Nielsen’s overall recommendation regarding the use of AI metaphors?
A) Avoid using metaphors altogether.
B) Develop your own unique AI metaphors.
C) Choose one metaphor and stick with it.
D) Use different metaphors depending on the task and AI’s capabilities.
Bonus Question: Why did Jakob Nielsen use animals to represent the 4 metaphors for using AI?
Intern: rabbit
Coworker: bear
Teacher: owl
Coach: alligator
Traditional Text Entry Fields Are Best
Text entry fields are probably the single most important interaction design element for any website or application that aims to allow users to make something happen. (For pure content consumption, buttons, hyperlinks, and menus are more important, though search requires text entry, so even these sites need to worry about proper design of text fields.)
How to design text entry fields? That’s easy, and the guidelines have been the same for at least 25 years:
Make it a box. That’s the good, old standard for a reason, because it makes it obvious that this is where you can enter something.
Keep the box empty: no filler or placeholder text inside the box. Empty boxes are magnetic for drawing users’ eyeballs, because people are scanning your layout for where they can do something.
Place a descriptive label just outside the box. Probably above, but could also be to the left (or to the right in languages that read right-to-left).
Optionally, describe any requirements for this field. Usually as small text below the field. For example, “8-20 characters, including at least one digit” for a password.
Optionally, add just-in-time help, usually activated by hovering over a small (i) icon at the end of the label. Requirements can be relegated to the help popup if they are usually not needed.
That’s it. Deviate from these guidelines at your peril. If you’re an ecommerce site, you’ll lose several percent of sales if using any other design in your checkout flow.
Boxes good. Help icons good. Placeholder labels bad. (Nuts.com)
Adam Silver, who’s one of the world’s best designers of simple user interfaces, wrote a longer piece about text-field design that’s worth reading. His main point was to warn against three alternative designs that have lower usability:
Placeholder labels and help text placed within the text box: terrible for two reasons: (a) scannability is reduced when the box isn’t a gaping open area that draws eyeballs, and (b) placeholder text within the text field disappears as soon as the user starts entering something (which is the entire purpose of a text entry field), meaning that people can no longer read what they need to know to complete the field.
Float labels that start out as placeholder text but then move up above the field as soon as the user starts typing. This design is better than pure placeholders, but suffers from other problems, such as the need to limit the label length to fit within the field.
Google’s Material Design that shows text entry fields by an underline instead of a text box. (The official Material Design website seems to acknowledge the usability problems with the original text field design, because it now also mentions boxes as an alternative design. They still list the bad option first, though.)
As Silver points out, many people have followed the first option in Google’s design docs under the assumption that Google must know what they are doing. No, Google doesn’t know what it’s doing when it comes to usability, except for the design of the search results pages. (And even there, the presentation of non-organic links borders on dark design.)
Google’s enterprise products have terrible usability. It pains me to say this since I was on Google’s advisory board during its startup years. But back then it focused on search and on creating the best user experience possible. They left that vision behind long ago and I no longer have anything to do with the company.
As with many usability issues, the good old design works best. If it ain’t broke, don’t fix it.
When faced with a wall of information on a web page, eyeballs are drawn to open text boxes that the user can act upon. Similar to the open doorway in this wall. (Midjourney)
Speaking at the Y Oslo Conference
I will be live on stage at the Y Oslo Conference in Norway this Wednesday, Oct. 30. Before the event, I hope to get time to visit the iconic “Scream” painting at the art museum. (Midjourney)
Usability Poster
I continue in my failed attempts to make posters for design and usability. Here’s my latest, using a variant of a prompt posted by OscarAI (a great creator worth following). In general, I recommend following several creators and getting inspiration from their work. The AI community is admirably willing to share.
Usability leads to happy users. (Midjourney)
Interpreting the Elo Scores for AI Models
The AI Leaderboard publishes Elo scores for AI models, which is one of the main ways of assessing the prowess of new models. Unfortunately, Elo is a number that makes no sense.
Peter Gostev published a nice chart that relates Elo scores to the average win rate in pairwise competitions between AI models. For example, 100 extra Elo points roughly correspond to a 25% improvement in win rate, whereas 40 Elo points correspond to 10% higher win rate.
The concept of Elo ratings is borrowed from chess. Elo ratings are a method for calculating the relative skill levels of players in competitive games, particularly zero-sum games like chess. Named after its creator, Arpad Elo, this system provides a numerical representation of a player’s ability compared to others within the same player pool. The system first calculates the probability of a player winning based on the rating difference between opponents. Then, after each game, both players’ ratings are adjusted based on the outcome. Unexpected results (e.g., a lower-rated player defeating a higher-rated one) lead to larger rating changes.
The AI Leaderboard is constructed by having two different AIs solve the same user-provided question and asking the user to select the best answer. Similar one-on-one battle as a game of chess, so the same rating methodology is used to keep track of the outcomes and rate the players. (Midjourney)
Lower Cost to Orbit
You have probably seen videos of SpaceX’s marvelous achievement of using a “chopstick” tower to catch the returning heavy booster rocket after the Starship launch a few weeks ago. While it made for great video footage to snatch a descending 71 meters (233 ft) tall rocket out of thin air, what happened afterwards was more important: after the catch, the booster rocket was lowered back to the launch platform. The plan is to achieve a turnaround time of one hour between launches: the rocket goes up, the booster lands, the booster is refueled, a new payload is mounted on top, and up it goes again.
If this can be made to work, fast reuse of expensive launch components will cause the cost of space launches to drop dramatically. The cost of sending something to low Earth orbit (LEO) used to be an inflation-adjusted price of $75K per kg in the age of NASA’s space shuttle program during the 1980s. Now, using the SpaceX Falcon Heavy rocket, the cost is $1,500 per kg, or 50 times less.
When the price of something changes this dramatically, the way it’s used changes as well. The most striking current example is the Starlink Internet service, which is now the cheapest and best way to provision Internet service outside densely populated areas. Starlink works by transmitting signals to/from 7,000 orbiting satellites. The large number of satellites provides high-quality service globally, but it would have been completely cost-prohibitive to launch that many satellites with NASA’s rockets.
If Starship can be refined to support those hourly launches, the price per kg to orbit will drop to around $100. This will make it possible for any high school science club to launch its own spaceborne experiments. But more importantly, it’ll enable an entirely new set of services. For one, the next generation of Starlink satellites will give customers 10x the Internet bandwidth, relying on the adding booster power of Starship to launch heavier satellites.
Tomas Pueyo published a great analogy with Earth-based transportation. The famous Silk Road was properly named: it was based on transporting silk from China to Europe. Because camel-based transport was expensive, it was only feasible to export materials that were simultaneously lightweight and high cost: silk!
Today, we ship everything from China to Europe (and the Americas, which were unknown to the ancient Chinese), even cheap and heavy stuff. Why? The cost of transportation per kg has dropped to almost zero with container ships.
Think of space going from camels to container shipping in about a decade. That’s the revolution we’re about to experience.
The cost of space launches is dropping dramatically, thanks to SpaceX’s approach to reusing expensive rocket boosters. Lower prices enable many new applications. (Midjourney)
Answers to the AI Metaphors Quiz
Here are the answers to the quiz in the beginning of this newsletter. How many did you get right? If you failed some questions, review my full article, “The 4 Metaphors for Working With AI: Intern, Coworker, Teacher, Coach,” to learn more.
How did you do on the test? Post your score in the comments. (Midjourney)
Q1: What is Jakob Nielsen’s main criticism of the “AI as an intern” metaphor?
Correct answer: B) It does not accurately reflect the capabilities of modern AI.
Q2: What is the key driver of AI’s increasing capabilities?
Correct answer: B) The AI scaling law
Q3: Which task does Jakob Nielsen believe will likely remain beyond the capabilities of AI until at least 2027?
Correct answer: A) Analyzing qualitative usability studies
Q4: What are the two forms of tasks AI can perform as a coworker?
Correct answer: C) Independent assignments and collaboration
Q5: What is the best approach for design ideation with AI?
Correct answer: B) Collaborate closely with AI, alternating who takes the lead.
Q6: In the teacher metaphor, what is AI’s primary role?
Correct answer: D) Guiding the user in acquiring new skills and knowledge.
Q7: How does Jakob Nielsen suggest using AI to learn a new subject effectively?
Correct answer: D) Ask for simplified explanations, alternatives, and examples until you understand.
Q8: What historical figure does Jakob Nielsen use to illustrate the benefits of one-on-one instruction?
Correct answer: B) Alexander the Great, who had Aristotle as his personal tutor
Q9: What is the primary benefit of using AI as a coach to critique your work, even if its overall performance is inferior to yours?
Correct answer: C) It forces you to justify your decisions and consider alternative approaches.
Q10: What is Jakob Nielsen’s overall recommendation regarding the use of AI metaphors?
Correct answer: D) Use different metaphors depending on the task and AI’s capabilities.
Bonus Question: Why did Jakob Nielsen use animals to represent the 4 metaphors for using AI?
Answer: I used animals to represent the four AI metaphors because I didn’t want to draw a robot four times. Robots are usually the best way to depict AI in illustrations; see my article on how to visualize AI.
Happy Halloween
If you made it this far, you deserve an extra Halloween image:
Happy Halloween. (Midjourney)