top of page
  • Writer's pictureJakob Nielsen

UX Roundup: UX Portfolios | Linkrot | Midjourney User Testing | AI-Sourced Articles | Cheaper AI

Summary: Target UX portfolios at busy hiring managers | Linkrot eating away at Web connectivity | Midjourney to start usability testing | Perplexity AI creates full articles from existing content | AI 8x cheaper per year

UX Roundup for June 10, 2024. (Midjourney)

Portfolio Advice from the Trenches

I have written a good deal about UX portfolios already, but I simply need to share two quotes from Casey Britt’s brilliant observations after he reviewed 600 portfolios for a recent job opening:

  • “If you don’t have time to type or paste a link into the application, I don’t have time to review it.”

  • “Most of the time we're hiring because the team has too much work. Every minute spent reviewing portfolios is a minute not spent on the work. Keep that time-strapped hiring manager in mind when creating your portfolio.”

Most of his recommendations echo what I have said before, but Casey also strongly suggests frontloading the portfolio with the actual design work. Seeing is believing. You don’t have time to scroll a lot when you need to make it through 600 portfolios.

Frontload the actual designs in a UX portfolio to capture the attention of overly busy hiring managers. (Midjourney)

Linkrot Is Undermining Web Connectivity

Linkrot is the phenomenon of hyperlinks on the internet breaking over time, as the linked web pages are moved, deleted, or reorganized. This results in dead links that lead to non-existent web pages, causing frustration for users and reducing the reliability of online content.

The Pew Research Center published a study of Linkrot as of October 2023. Here are some key findings:

  • 35% of web pages that were live 10 years ago are no longer on the Internet.

  • 22% of web pages that were live 2 years ago are no longer on the Internet.

  • Subtraction tells us that an additional 13% of links rotted during the 8 years following those first two years.

  • The initial rotting pace is 11%/year, followed by 1.6%/year later.

  • In other words, early Linkrot is 7 times worse than late Linkrot.

  • Even though the pace of Linkrot declines, the longer a link survives on the Web, linkrot continues unabated year after year. Unfortunately, Pew didn’t look at 20-year-old web pages, but I bet you that less than half of them are still available. (As would be predicted if the 1.6% rotting pace continued from year 10 to year 20.)

  • On news websites, 23% of articles have at least one broken link half a year after publication.

  • On Wikipedia, 11% of all references linked at the bottom of articles have suffered Linkrot, meaning that those references are no longer available.

What do these stats tell me? Is Linkrot a bad problem or a minor problem? On balance, I would say it’s a major problem but not a disastrous problem. Particularly now that we have AI answer engines, one can usually find substitute information for most dead links. Linkrot is highly annoying but doesn’t prevent users from using the Web.

Even though I just declared Linkrot not to be a disaster, I still implore you to fight it. If you put something on the Internet, keep it there. If you need to reorganize your IA and move pages to new URLs, leave a redirect in place so that old links from other websites continue to work.

Web links are rotting, making more and more web pages inaccessible over the years. In a chain, once one link breaks, the entire chain breaks, undermining free web navigation. (Midjourney)

Midjourney User Testing

Alie Jules reported from Midjourney’s recent office hours that they will be starting a round of user testing with new users to improve the usability of the web-based UI for image generation. (Presumably, they have given up on the usability of the Discord UI, but that’s probably the correct decision. The future is clearly the web-based UI and/or a dedicated application.)

Since I have been a harsh critic of Midjourney’s poor usability in the past, fairness dictates that I praise them for this new direction. It’s never too late to embrace usability. The worse your design, the more low-hanging fruit, and even a single round of user testing with 5 participants will immensely improve your UX, assuming that you redesign to fix the worst flaws — which will be seen immediately.

Midjourney has the best image quality of all the image-generating AI tools, and it was number 3 in my study of the AI tools used the most by UX professionals. I’m personally a big Midjourney user, so I am pleased to learn of this new push for usability.

From that same office-hour report, Midjourney also claims that a new version is a few weeks from finishing its training run and that this release will have substantially improved image quality. Since Midjourney is already the best on this parameter, an even better IQ will be astounding. They are also reportedly working on improved prompt adherence, which is currently Midjourney’s weak point relative to Ideogram.

Midjourney will finally do user testing. (Image created with Midjourney!)

Comic strip about Midjourney’s user testing. Ironically, I made this with Story Illustrator GPT, not Midjourney, since Midjourney’s ability to spell is still too bad for a comic strip. (OpenAI has its share of typos, as you can see, but at least we get a feel for the comic in this version.)


AI-Sourced Articles

Perplexity AI has added a feature for users to create full articles the service will host. As an experiment, I made an article about Jakob’s Law of the Internet User Experience.


Perplexity had the advantage that my article was already on the Internet, and indeed, it used my article as one of its sources. (Nicely adding a clickable reference.)

Perplexity uses the term “curating” to describe the human effort involved in creating these articles, which they call “pages.” The supposed analogy is with Wikipedia. Wikipedia currently doesn’t have a page about Jakob’s Law, but only a two-line description in another article. My subjective assessment of this one article on a topic where I am the supreme expert (since it’s my own law 😊) is that the Perplexity pages are of higher quality than most Wikipedia articles.

The “curator” of a Perplexity page can suggest topics for additional sections (which will then be written by Perplexity, complete with references to appropriate websites), change the ordering of sections, add illustrations, and also request edits and rewrites of individual sections.

I think I probably still prefer creating articles on my own, but I can see this new Perplexity service becoming popular as a way for people to get a first article immediately and then gradually improve it.

Perplexity now offers the ability to have AI write full articles under human supervision and direction. (Ideogram)

AI Got 8x Cheaper in a Year

With AI's fast pace, it’s easy to lose track of the bigger picture. We frequently hear announcements of AI prices being cut in half or similar. These price cuts add up.

At the recent Microsoft Build event, Microsoft CTO Kevin Scott highlighted the stat that GPT-4o (released May 2024) has a 12x decrease in cost and a 6x increase in speed compared with the original GPT-4 (released March 2023).

An improvement of 12x over a 14-month period corresponds to 8x in a year.

How can prices drop by 8x in a year? It’s a combo of more efficient software and more efficient hardware, as I discussed in my piece Compute Creates Usability. The third factor that will lead AI to consume fewer resources is still missing: improved AI usability.

We usually read scaremongering reports about the escalating costs of training ever-bigger frontier models. The aptly named Visual Capitalist site has a nice visualization of AI training costs over time. However, I have a more optimistic take: my analogy is with the increasing cost of building cutting-edge semiconductor fabs. Yes, each fab generation is more expensive than the last: a 2nm fab costs about $28B, whereas a 3nm fab only costs $20B. However, the impact on the world’s consumers is indisputably positive: ever-more powerful chips become cheaper and cheaper over time because the more expensive fabs produce cheaper chips.

Similarly, with AI, we have to pay more for each generation, but because future AI is more capable, it will be used more, resulting in lower cost per use. Most of the economic value of AI will be captured as consumer surplus. That is, we all benefit! (But don’t cry for the AI vendors: even though they will only cash in on a tiny percentage of the benefits they create for society, they’ll still make plenty of money.)

Compound growth makes many incremental improvements add up to huge gains in the end. (Ideogram)


Top Past Articles
bottom of page