Summary: Augmented reality glasses help blind users navigate world | Practice with AI to build lasting skills | AI for fuzzy concept searches | AI as a super-thesaurus | Readability levels | Microsoft reducing its spending on AI infrastructure | Comparing two AI video tools’ lip-synching | New ranking of the top AI services | New Chinese agent-based AI
UX Roundup for March 10, 2025. (Ideogram)
I made a short overview video summarizing some of the top stories in this week’s newsletter (YouTube, 2 min.)
Demonstrating GPT 4.5’s Creative Writing Potential
I made a short explainer video about ChatGPT 4.5 (YouTube, 3 min.), where I used the model to write the manuscript about itself, emulating the tone of voice of a classic YouTube UX influencer.
A stereotypical YouTube influencer gets voice through a manuscript written by ChatGPT 4.5 flexing its creative muscles (and turned into a video with the help of ElevenLabs, HeyGen, and Kling.)
The subject matter discussed in the influencer manuscript was based on my analysis of GPT 4.5 in last week’s newsletter. Compare my tone of voice with ChatGPT 4.5’s reimagined version for a YouTube influencer.
Augmented Reality Glasses Help Blind Users Navigate World
I admit to having been mostly negative about the prospects for “metaverse” user interfaces, such as virtual reality (VR) and augmented reality (AR) headsets (for example, the Apple Vision Pro, which indeed has been a flop). However, I also always said that these technologies have their place in specialized applications, even though I don’t view them as “the next big thing” in UX.
Indeed, Meta’s AR glasses made in collaboration with Ray-Ban has found a substantial niche use as assistive technology for blind users, according to a recent Wall Street Journal article (subscription required). The article reports that more than a million units have been sold, and this product is now reported to lead the U.S. smart-glasses market. (Which I’m sure must be very limited for now. Still, being the leading product is impressive.)
A major use case for these VR devices is as assistive technology for blind people. The reported interviewed one user who said that the VR technology helped her “get dressed and prepare meals to sorting laundry and reading books to her 3-year-old granddaughter.” Another user interviewed for the article said “he was afraid to go for a walk alone, fearful of getting lost or hit by a vehicle. That fear is now gone because the tech specs can identify or describe most objects within the camera’s sight.”
The article also reports some failures of the technology, such as misreading a house number. However, since these are still early days for embedding AI within a pair of eyeglasses, I would expect the next generation to be much better. In any case, the competition is not hypothetical perfection, but the realistic alternatives such as guide dogs (who definitely can’t read a house number) or staying home.
A combination of AI and AR technology will probably soon be more helpful to blind people than a guide dog. (Midjourney)
Practice With AI to Build Lasting Skills
Some recent studies have raised a warning flag that students who use AI may not learn the underlying skills sufficiently well. As I’ve said before, this may not be as serious a problem as it’s made out to be because people will be performing virtually all knowledge-worker tasks with AI assistance in the future. So, their ability to work without AI may be as irrelevant as testing students’ ability to multiply large numbers in their heads without resorting to a calculator. Anybody needing to work with big numbers will turn to Excel, not mental math.
A recent paper by Benjamin Lira and colleagues (University of Pennsylvania, Harvard, and Microsoft Research) provides a contrasting example, where learners did retain the skill they were taught with AI.
In this study, 2,238 participants were asked to improve the writing of job application cover letters. First, the participants had a lesson introducing five principles of effective writing. They were then divided into 3 conditions for practicing the skill they had been taught:
No practice exercise.
No AI help, simply completing the exercise of rewriting a sample letter.
Rewriting a sample letter with the help of an AI writing tool that “revised text instantly based on these principles” (that had been taught in the lesson)
After the practice exercise (which was no practice in the first study condition), the participants were asked to rewrite yet another sample letter (which was the first rewrite for people in the “no practice” condition). All these rewrites were done without AI help, so they assessed whether the learners had internalized the writing principles, as opposed to their ability to use the AI tool.
These rewritten letters were later presented to independent judges as pairs of letters from different (non-disclosed) study conditions. For each pair of cover letters, the judge was asked to choose which job candidate to invite for a hypothetical job interview.
The probabilities of having participants having produced a winning letter through their rewrites were as follows:
No practice: 47%
No AI help during practice: 50%
AI help during practice: 54%
The differences were statistically significant at p<0.001.
We see that doing a practice exercise after a theory lecture helps. No surprise. But having practiced with an AI tool that embodied the theory made people better than completing the unaided practice.
The researcher also completed a follow-up study where the learners were asked to perform yet another rewrite one day later. In this condition participants who had practiced with AI continued to outperform the participants who had practiced without AI. In a small surprise, the people who had practiced without AI now scored the same as people who had not practiced at all. The test rewrite probably served as practice for the people in the “no practice” condition.
In a follow-up study, the authors replaced the “no practice” condition with a new condition, where the learners were simply shown a rewritten letter that had been produced with AI, as opposed to working through the exercise of improving a letter with AI help. In this follow-up, these two conditions scored essentially the same. (There was a 1% difference in the win rate, but the p score for this difference was p=0.56, meaning that it was almost certainly pure random chance.)
I would still think that in general, it’s better for students to work through an exercise, but in this very simple case study of rewriting cover letters, simply seeing better writing was enough.
We shouldn’t make too much of this study for two reasons:
The task was indeed very limited in scope: teaching learners the principles for writing a one-page job application cover letter. Not exactly rocket science or UX design.
The assessment of “durable learning” was conducted only one day after the training session. Will learners’ skills still be better a month later or a year later? We don’t know.
Despite these limitations, it’s positive to see more research to tease out the conditions under which AI helps people learn better. Using AI as a teacher is one of the four main metaphors for working with AI, so we need to get better at it.
AI has great potential as a teacher, both for primary education and to upskill business professionals. We still need to discover more about how to best use these new capabilities to instill skills in learners. (Midjourney)
AI Doing Fuzzy Full-Concept Search
When I was writing the previous news item, I could not remember when I had discussed the issue of overly-narrow use of AI in education causing students to retain less information when they had used AI when working on exercises.
It’s almost impossible to search for something as vaguely specified as this. However, AI solved the problem. I asked ChatGPT o3-mini-high with web search the following: “In which of his UX Roundup newsletters on the UX Tigers website does Jakob Nielsen discuss the differences between students using AI to complete an exercise for them as opposed to using AI to assist in understanding an exercise while working on it?” It thought for one second, and bingo, gave me the answer.
(I asked Grok 3 in DeepSearch mode the same question, but it gave me the wrong answer, listing a newsletter that did address education, but not the more precise issue I was after.)
Understanding AI’s ability to grok entire concepts and larger discussions without relying on specific keywords is one of the superpowers afforded to people who learn advanced uses of AI. (Disregarding the point that the Grok product wasn’t as good as o3 at grokking my vague description of my own article.)
AI can search for ill-defined ideas, though you may have to employ multiple AI tools for success if your problem is sufficiently fuzzy. (Midjourney)
AI Super-Thesaurus
Another use of this idea is the use of AI as a super-thesaurus. In a regular thesaurus, we can look up an individual word and get inspired by alternative words that have either the same meaning (synonyms) or related meanings that may actually be better at conveying the author’s intent (connotative variants).
As an example, I pasted the two paragraphs starting “Understanding AI’s ability” and “Another use of this idea” (as they came out of my keyboard) into Claude Sonnet 3.5 with the instruction to write three different versions of the text: one targeting an audience of high-school students, one targeting an audience of graduate students, and one being suitable as a promotional post on LinkedIn to position me as a thought leader in business use of AI.
Here are the 3 rewrites:
For High School Students:
You know how AI can understand what you mean, even when you're not using exact words? That's like having a superpower when you know how to use AI well! Here's a cool example: Think of a thesaurus — you know, that book (or website) where you look up different words that mean the same thing. Well, AI is like a super-smart thesaurus that doesn't just give you similar words, but actually understands what you're trying to say and helps you find the perfect way to say it. It's like having a friend who always knows exactly what word you're looking for!
For Graduate Students:
The semantic processing capabilities of contemporary AI systems extend beyond simple keyword matching to encompass holistic concept interpretation and contextual understanding. This cognitive flexibility represents a significant advancement in natural language processing, particularly for users who can leverage these capabilities effectively.
An illustrative application of this technology lies in its capacity to function as an enhanced lexical resource: unlike traditional thesauri that operate on direct word-to-word relationships, AI systems can analyze semantic intent and suggest alternatives based on both denotative equivalence and nuanced connotative variations, thereby facilitating more precise and contextually appropriate language selection.
For LinkedIn:
🚀 Revolutionary Insight: The true power of AI lies not in keyword matching, but in its ability to comprehend complete concepts and contextual nuances. As a leader in AI implementation strategies, I've observed that this capability transforms how we approach business communication and decision-making.
Here's a powerful example: While traditional tools like thesauri operate on simple word substitution, AI elevates this to strategic communication enhancement. It doesn't just find similar words — it understands intent and context, enabling unprecedented precision in business communication.
This isn't just about better writing — it's about unlocking new levels of business efficiency and effectiveness. Those who master these advanced AI capabilities gain a significant competitive advantage in today's rapidly evolving business landscape.
Readability Levels
I’ve talked about readability levels many times. In the example of using AI as a super-thesaurus to rewrite my text, here are the readability scores (Flesch-Kincaid reading grade):
For high school students: 8th grade
For graduate students: 26th grade
For LinkedIn: 16th grade
It seems the “graduate student” version breaks the Flesch-Kincaid formula, because it can’t truly require 26 years of schooling to understand that paragraph. It is written in insufferable academiaese, but I think a good Ph.D. student could read it easily, despite having been confined for only maybe 20 years in the education system. The “LinkedIn” version would be suitable for people with a bachelor’s degree, which is right on target for a posting aimed at business professionals.
In my own writings for this newsletter, I usually target that same 16th-grade reading level, though I sometimes stray and write at a master’s degree level (18th-grade reading level). Despite being a recovering academic myself, I don’t think I’ve ever breached that 26th-grade level Claude gave me.
The key point of this exercise is that a few different AI rewrites in different styles can give you inspiration for better ways of expressing yourself that reach far beyond the simple word variations you get from a thesaurus. For further ideation, you can also ask for rewrites in drastically varying tones of voice. (When I do this, I usually ask AI to make one of the versions in “Jakob Nielsen’s writing style.” Sometimes it writes better as me than I do myself.)
Ask AI to create multiple versions of your content: it’ll work as a super-thesaurus. (Midjourney)
Microsoft Reducing AI Spending
Microsoft is reportedly scaling back its spending on AI infrastructure by canceling leases on 200 MW worth of datacenters. (As an aside, interesting that datacenter capacity is now estimated by electricity consumption and not by FLOPS.)
My guess is that Microsoft is predicting that most of the value created by AI will accrue to the users and/or the hardware vendors (NVIDIA and the like), and not to the foundation models which are becoming commodities. Last week, the AI Leaderboard announced that Grok 3 had “leapfrogged” GPT 4.5, but only by a single Elo point, which is clearly not enough to make anybody switch. Low prices may be the only way for these models to get customers, since there will be plenty of AIs with almost-equivalent IQ. The problem could also relate to Microsoft’s peculiar relationship with OpenAI, which I am not sure any outsider really understands.
Microsoft closing some AI datacenters. Maybe because they estimate that most of AI’s value will be consumer surplus and not vendor profit. (Leonardo)
Lip Synch Comparison
AI video generation service Hedra is out with a new release that claims to be the first multimodal model “in production” (i.e., that anybody can subscribe to). They handle images and audio and can generate video with both motion and lip-synch. To put it to the test, I generated a few seconds of a song where I had recently lip-synched a segment with Kling AI. See what you think. (The comparison is only for timestamps 2-7 seconds, so a total of 5 seconds can be fairly compared. The rest of my video was animated with other tools.)
Watch the comparison video on my Instagram channel.
I made a short video comparing two AI video tools lip-synching the same song segment.
DeepSeek Takes Silver in New Ranking of Top AI Services
A16Z released an update to its regular ranking of the top 100 AI services. The number-one most used AI tool remains ChatGPT. (It also topped my survey of the AI tools used the most by UX professionals.)
The big surprise is that DeepSeek entered the ranking at the number-two position despite having been released less than two months ago. DeepSeek pushed perennial silver medalist Character (a virtual companion service) into third place.
Rounding out the list of top placements are Perplexity (an answer engine), Janitor (a more spicy virtual girlfriend/boyfriend service than Character), Claude (a general LLM), QuillBot (a writing assistant), and Suno (my favorite music-maker).
DeepSeek is the big surprise. It currently has around 300 MAU (monthly active users), compared with ChatGPT’s 400 MAU. Even more impressively, DeepSeek passed the 10 million user mark in only 20 days — something that took ChatGPT 49 days, which was considered exceptionally fast at the time.
The world’s most-used AI services are ChatGPT, DeepSeek, and Character. (Leonardo)
Talk about Suno, when scrolling to the bottom of my Instagram channel, I came across two songs I had made with Suno in January 2024:
Definition of Usability [Instagram]
Dark Design [Instagram]
Ouch, they’re nasty. Compare this with one of the better songs I made with Suno almost exactly one year later to the day:
Suno has come a long way in only a year. Placing number 8 of the most-used AI services is well-deserved.
New Chinese Agent-Based AI: Manus
Besides being the year of next-generation AI foundation models, 2025 is more specifically the year of AI agents: AI tools that can take action on behalf of a user and not just provide information.
Manus AI is a new Chinese AI product that is being touted by AI influencers for its ability to execute complex, real-world tasks autonomously. Manus has high scores on the GAIA benchmark, surpassing OpenAI’s Deep Research. (This benchmark evaluates AI assistants’ ability to solve real-world problems.)
Manus is only available as an invitation-only web preview, with plans to transition to free access later. Since I’m still on the waitlist, I can’t give you further information, except to echo those rave reports from AI influencers with access. (Invite codes are reportedly being resold at high prices, so I doubt I’ll land one soon.)
Key Features of Manus AI:
Autonomous Task Execution: Manus can independently execute tasks such as report writing, spreadsheet creation, data analysis, travel itinerary planning, and more.
Multi-Modal Capabilities: It processes and generates multiple data types, including text, images, and code.
Advanced Tool Invocation: Manus integrates with external tools like web browsers, code editors, and database management systems to automate workflows.
Adaptive Learning and Optimization: It continuously learns from user interactions to provide personalized and efficient responses.
I like the potential of the example in the demo video, where Manus digests several data sets and creates an interactive dashboard for the user’s further exploration of visualizations of the data. This dashboard is published as a website, so that the user can share the link with others. Vibe coding and vibe design in action!
As Manus points out on its website, “manus” is Latin for “hand,” which an appropriate name for an agent. The evolution of the human hand with its famous opposable thumb allowed us to use tools and become masters of the Earth. Maybe this new AI agent will perform a similar feat. (Midjourney)