UX Roundup: Email UX | Sora 2 | AI New Business Capabilities | AI Consumer Surplus | Latent Affordances | Ambient Clinical AI

Summary: Email user experience | New video model Sora 2 compared with older models | AI offering new business capabilities instead of just replacing old jobs | Most of AI’s economic value is consumer surplus | Latent affordances for AI feature discovery | Simple ambient AI scribes help clinicians substantially

UX Roundup for October 6, 2025. (GPT Image-1)

Email Newsletter User Experience

Two videos about the design of email newsletters and why they are the best way to build connections with a loyal audience:

• Music video, with a cowboy avatar (YouTube, 3 min.)

• Explainer video, narrative presentation by a newscaster (YouTube, 9 min.)

Both videos are based on my article Email Newsletters Build Loyal Audiences (25 min. to read). The article obviously provides more detail, but the song is funnier, and the explainer offers a quick, approachable overview. Each of the three media forms has its advantages.

We are still in the early stages of AI video-making, and it’s easy to spot flaws (especially continuity errors) if you watch closely. However, progress has been immense this year, as is apparent from watching the three music videos I made with the same avatar. (Seedream 4)

For the music video, I used the same avatar as I had used for other songs in December 2024 (YouTube, 2 min.) and March 2025 (YouTube, 2 min.). It is interesting to review all three videos: progress in avatar animation and B-roll creation has been immense in these few months. It’s harder to assess progress in music quality, but the new song is probably a little richer, especially in the orchestration.

The explainer video used a new avatar, and I think this is the best I have made yet, in terms of realism. I think HeyGen did better with this avatar because of the simplistic setting: the avatar is sitting behind a desk and not moving around, nor are there other people or animals within the frame. (I don’t know whether this explanation is actually true: unfortunately, current AI is not explainable, so users are left to make up superstitions about how it works.) Compare this week’s avatar with last week’s avatar, which had a more complicated setting and much less natural delivery.

AI avatars are still easily distinguishable from human actors, but they are getting closer every month. (Seedream 4)

New Video Model Sora 2 Compared with Veo 3

OpenAI has finally released its long-awaited Sora 2 video model, after a long hiatus from AI video, since Sora 1 was a sore disappointment.

Sora 1 was immensely hyped before its release, with many stunning clips teased by OpenAI that had been generated using their internal model. However, the release model had been crippled (probably to save compute, or possibly for censorship reasons) and mostly generated unremarkable videos when real users got to try it. A crash-and-burn release! (Seedream 4)

I made a short demo reel comparing a few clips made with the new Sora 2 model (YouTube, 3 min.) and earlier models, Sora 1 and Veo 3 (Google’s current video model). This video also contains an avatar animated with HeyGen Avatar IV through most of the video and Sora 2 at the end, allowing you to compare a different style of video.

Sora 2 is probably slightly ahead of Veo 3 in video quality and prompt adherence, though not by much. If you are a “Pro” level subscriber with OpenAI (costing US $200/month), you can generate 15-second clips with Sora 2, whereas peons at the $20/month level can only generate 10-second clips. Both are superior to Veo 3’s current limit of 8 seconds.

Despite its slow start, OpenAI’s Sora video model has probably overtaken Google’s Veo model. For now, that is. I expect better results from Google soon, as it has been firing on all its AI cylinders recently and likely has more AI compute than OpenAI, due to its use of internally developed Tensor Processing Units (TPUs). (Seedream 4)

8 vs. 15 seconds may not sound like much, but it allows deeper dialogue, going beyond the ultra-short soundbites you will have seen in videos like my own “Error Prevention Explained by Vikings” and “Consistency and Standards: Why People Like This Usability Heuristic,” both of which were made with Veo 3.

Overall, I am disappointed with Sora 2, and I expect Veo 4 to surpass it shortly, not to mention Chinese competitors like Seedream, Wan, and Kling. (Let alone ByteDance’s OmniHuman which will give HeyGen a run for its money if/when ByteDance ever productizes this research.)

Conclusion: After the long wait, Sora 2 is a disappointment. (Seedream 4)

OpenAI’s primary focus with Sora 2 is to build a new social media platform for AI-generated video. I am not sure I believe this will work, since the value in a video is its storytelling and not what tool was used to generate it. Human actors and camera crew or AI? Who cares how it was made when you’re watching. (Even in legacy Hollywood movies, special effects have been all computer-generated for years.)

They emphasize the ability to remix existing videos you come across to put your own spin on them. The original video creator’s rights be damned. Of course, if you create a video with Sora 2 and publish it on their platform, you should be game to have your work repurposed by others — otherwise publish elsewhere.

I came across a brilliant Sora 2 video by user JesperAgain that was a pretend BBC news story from 1960 about the launch of Sora 2. I did remix this video as an experiment and redid the exercise with Veo 3. (I changed the setting to be the research lab where I used to work, Bell Communications Research or Bellcore, which was one of the world’s leading research labs back in the day and did much pioneering work in digital video, even though it obviously did not invent Sora 2 or Veo 3, as it says in my spoof news stories.) Watch my newsreel about the “invention” of AI video by my old lab in 1960 (YouTube, 1 min 32 sec.).

For now, the true beauty of Sora lies in its high usability for remixing existing videos: specify a one- or two-word change (e.g., “make it Henry VIII” to alter a video of Genghis Khan giving a TED Talk). It’s always easier to imagine a change than to imagine something completely new from scratch. This ease of iterating creates a flood of variations, most of which are so-called “slop,” but every now and then, it takes 19 tries to strike gold. (Seedream 4)

Software Eating the Labor Market

A perceptive presentation by Alex Rampell, General Partner at Silicon Valley’s leading venture-capital firm A16Z (YouTube, 26 min.) discussing the TAM (total addressable market) for AI being many times bigger than that of traditional software. AI will eventually replace all current human labor, freeing humans to focus on the 3 ways in which people will rule AI: Agency, Judgment, and Persuasion.

Technology has already devoured entire industries, but we’re now witnessing something far more profound: software isn’t just transforming work anymore, it’s replacing workers.

This transformation targets a prize so vast it dwarfs the entire software industry. While the global SaaS (Software as a Service) market hovers around $300 billion annually, the US labor market alone is worth $13 trillion. Software is no longer competing for IT budgets; it’s competing for headcount.

For decades, software companies thrived as glorified digital filing cabinets. Salesforce digitized the Rolodex, SAP digitized inventory, and Epic digitized patient records. These tools moved information from paper to screens, creating trillions in market value. But they didn't eliminate the human worker, they just changed their medium. A human still had to read that digital record, interpret it, and act on it.

Next-generation AI breaks this paradigm entirely. It doesn’t just store customer data; it calls the customer. It doesn’t log overdue invoices; it initiates collections. The software performs the job.

This shift is shattering traditional business models. Consider Zendesk: a company with 1,000 support agents pays roughly $75 million annually in labor costs but only $1.4 million for their software seats. When AI can answer customer queries autonomously, that company needs zero agents and therefore zero seats, wiping out Zendesk’s revenue. The alternative? Stop selling tools and start selling outcomes: handle all customer support for $5 million, saving the client $70 million while tripling revenue.

Yet framing this purely as cost-cutting misses the bigger story. AI delivers superhuman capabilities: instant scaling for demand surges, tireless execution of demoralizing tasks, perfect compliance with inscrutable regulations, and instant multilingual support in dozens of languages. These aren’t just cheaper alternatives but new capabilities that were previously impossible.

AI agents can perform the work of myriads of humans by running in parallel, making it possible for companies to get tasks done that used to fall by the wayside. Humans are still needed for oversight and to handle edge cases. For simple tasks, the ROI is already there, and more complex tasks become feasible with every AI upgrade. (GPT Image-1)

Last year, I presented my analysis of the same situation, albeit slightly less aggressively, as I was writing based on one year less progress in AI: “Service as a Software.” For a shorter version of that article, watch my explainer video ( YouTube, 4 min.) or my music video (YouTube, 2 min.). It’s interesting to compare these two videos, which I made only 10 months ago, with my current videos. The avatar animation was so bad that I completely abandoned lip synch for the music video and relied solely on B-roll. The song is great, though, and is based on jazz violin, which is a rare style.

Epoch AI has estimated that OpenAI’s current compute capacity (56 trillion tokens per day) translates into the equivalent of 7.4 million digital workers at an 8-hour workday, or slightly less than one million working round the clock, as AI can and should do. This is obviously far too little to take over all current human-performed tasks, let alone all the many new tasks that become possible, as just mentioned, which is one of the reasons to be incredibly bullish about the build-out of AI compute.

Should we worry that AI is taking over jobs that were once performed by humans? I say no, because most of those jobs were not that great to begin with. The more AI does, the more it frees up humans for new things. (GPT Image-1)

AI Value = Consumer Surplus

Sam Altman was on a world tour last week to raise trillions of dollars to build out OpenAI’s AI compute infrastructure. As we saw in my previous news item, we need this extra capacity urgently, and it will be worth many more trillions to humanity in improved standards of living. I expect that most of the value created by the AI buildout will become consumer surplus and not captured by the builders.

Sam Altman traveled the world to drum up more GPU and high-bandwidth memory (HBM) compute for his next-generation AI models. (GPT Image-1)

Consumer surplus is the economic term for the difference between the highest price a consumer is willing to pay for a good or service and the actual price they pay in the market. If you would pay up to $100 for an AI tool but pay only $30, your consumer surplus is $70: the extra value you get, but do not pay for.

In the standard supply and demand model, consumer surplus is the area under the demand curve but above the market price, forming a triangle that measures society’s unseen gain from voluntary exchange.

In a supply–demand chart, the x-axis indicates the price of a product or service. The blue line represents the demand curve, showing the quantity that customers are willing to buy at any given price. If the price is low, customers will buy a lot, but as the price increases, demand declines. The red line represents the supply curve, showing the quantity that producers can profitably produce at any given price. If the price is low, few will produce the item, but as the price increases, more becomes available. The yellow dot indicates the market price where supply and demand meet. (If the price were higher, unsold products would sit on the shelf, causing producers to lose money. If the price were lower, many customers would go home empty-handed because not enough could be made at that price.) The green area represents the consumer surplus, indicating cases where customers derive a higher value from their purchase than the market price.

AI capabilities unlock enormous practical value for users to save time, enhance creativity, and empower productivity. However, for several structural reasons, most of this value accrues to end-users, not AI firms as profits:

1. Marginal Cost of AI Scales Near Zero: Once an AI model is trained, serving another user or task is cheap. Since many firms compete, prices are forced down toward this marginal cost, and much of the value spills over to consumers as surplus.

2. Rapid Competition & Replication: Since most foundational AI advances diffuse quickly, rival firms can offer similar tools. This keeps prices low and makes it hard for companies to maintain high profit margins for long, again raising consumer surplus.

3. Price vs. Willingness to Pay: Research has shown that the collective consumer surplus from generative AI in the U.S. (the gap between what users would pay and what they actually pay) already dwarfs the total profits or revenues of leading firms. For example, in 2024, Americans enjoyed roughly $97 billion in consumer surplus from generative AI tools, but total U.S. revenue for OpenAI, Microsoft, Google, and Anthropic was about $7 billion.

4. AI as an Enabler, Not a Tightly-Gated Product: Like search engines or email, AI often becomes a widely available, sometimes free, platform. This openness ensures much of the value lands with users, not providers.

Since we’re all consumers, but few of us own AI foundation model companies, this is mostly good news, as long as the AI companies still realize enough profits to keep operating, which I suspect they will in the long run.

Latent Affordances for AI Feature Discovery

My old friend Don Norman famously advocated the use of perceived affordance in user interfaces: people should be able to see what they can do. Strong affordance visibility is still beneficial for most designs, but since AI can perform almost infinitely many tasks, we can’t make them all visible.

Enter latent affordances: a subtle clue that only appears in the user interface when it’s useful. Obviously, this modifier, “when useful,” was not possible in traditional systems, but an AI system that monitors the user and understands his or her behavior and intent can implement latent affordances.

Latent affordances are capabilities that are present in the system but not uncovered and made salient in the user interface until the user needs them. (ChatGPT Image-1)

Consider these three design options:

1. Loud UIs with strong and permanently exposed perceived affordances cause feature blindness along the lines of banner blindness as people ignore the fireworks.

2. Hidden AI without perceived affordances in the UI causes capability blindness (people never discover it, because normal people don’t go exploring in a UI for things they don’t see).

3. Latent affordances thread the needle: discoverable when relevant, invisible otherwise.

Many current products use a “loud” design to promote their AI features, but users have learned to tune out such aggressive design, leading to feature blindness. However, the solution is not hidden AI, as found in the standard open-text-field prompt box, where you can type anything, but don’t know what you can do. Latent affordances present a “Goldilocks” middle ground that has become available to designers through AI’s own abilities. (GPT Image-1)

Traditional UI design equates discoverability with exposure, making features more visible through buttons, banners, or tooltips. Latent affordances, which are discoverable when relevant, invisible otherwise, rely on precision of timing for the work that persistent visibility once did. When users notice help right when they need it, they internalize its value far more quickly.

The timing for revealing a latent affordance can fall into three moments along the user’s intent curve: pre-intent, mid-intent, and post-intent.

• Pre-intent cues appear before a user commits to an action. They act as gentle previews or prompts. For example, if a writer pauses before typing a title, a faint chip might appear, reading “Generate five title ideas.” It’s an invitation, not an interruption.

• Mid-intent cues occur during an active task. These are inline assists that respond to what the user is already doing, such as a greyed-out autocomplete that suggests the rest of a sentence, or an icon that appears over a table header offering “Auto-categorize.”

• Post-intent cues show up after the user finishes an action. They’re reflective or educational, helping users understand or refine what the AI just did. A document editor, for instance, might display a Teach-back card saying, “I tightened passive voice in four spots—undo any?”

Deciding what latent affordance to show depends on the right trigger. Three categories capture most cases: content, context, and behavior.

• Content triggers rely on the properties of the material the user is working with. For example, if a document contains many acronyms, the AI might softly suggest: “Generate a quick glossary?”

• Context triggers react to the situation around the user: device type, network conditions, or policy settings. A mobile app on a slow connection could quietly offer a “lightweight draft mode.”

• Behavior triggers arise from what the user does. Repeated deletions, hesitations, or loops can all signal that help is needed. If someone backspaces the same title three times, the system might offer: “Would you like to see alternative titles?”

Each trigger can pair with any timing, creating nine possible combinations: the “9 cells” of latent design.

Latent affordances are an emerging design pattern with few exemplary implementations to date. Here are two ideas:

Quiet ghosts are the gentle placeholders that whisper potential without demanding it: a faded hint like “Try ‘Summarize in 5 points’,” or a ghosted command that fades once used. Unlike onboarding tours or modal pop-ups, these are invitations rather than instructions. Their presence suggests capability without creating noise, and over time, they teach users what’s possible through repetition and disappearance.

Progressive hinting is a system of learning through gentle correction. The first time a user fails at a task, the interface offers a minimal nudge (e.g., “You can ask me to ‘extract dates’ next time.”). There’s no punishment, no training mode, just quiet feedback that turns failure into awareness. One hint per session is enough; subtlety builds mastery better than floods of guidance.

Both examples demonstrate the importance of the AI’s ability to predict what the user is trying to accomplish closely enough for the latent affordances to be useful, rather than annoying. The AI doesn’t need to be perfect in its predictions, but it can’t be wrong too often. One strategy to enhance usefulness is to surface two or three likely options instead of a single guess as to what the user needs in that moment. (You can’t show too many options, or the design crosses into that “loud” territory where overkill causes users to ignore UI promotions.)

Watch: Explainer video about Latent Affordances (YouTube, 5 min.).

Ambient AI Scribes Reduce Doctor Burnout

I am most excited about the ability of AI to totally transform the economy by making new things possible that were never done before through completely redesigned workflows or fully new workflows. My story in this newsletter about “AI Eating the Labor Market” is an example of the radical change we need to uplift humanity and improve living standards for billions of people.

However, many of these trillion-dollar opportunities require immense organizational change, which doesn’t happen fast. The bigger the company, the more it resembles the Giant Sloths of the Pleistocene epoch.

The Giant Sloth (Megatherium) was big (weighing about 8,800 pounds = 4 metric tons), but known to take it easy, meaning that it was easy prey for humans once they arrived in what we now call Latin America. Today’s enterprise companies are the same. (Seedream 4)

While we wait for revolutionary transformations, smaller gains are welcome and can improve users' everyday lives. A great example is a new paper by Kristine D. Olson and many coauthors from Yale School of Medicine and a number of medical clinics. The researchers investigated the impact of ambient AI scribes in clinical use.

Currently, physicians spend more than half their workday updating the electronic health record (EHR) system, and only a quarter of their time is spent face-to-face with patients. The research documented substantial changes in this abysmal picture.

Ambient AI scribes record provider–patient conversations during office visits and use speech recognition and large language models to transform the audio recordings into structured clinical notes. These notes integrate directly into the EHR, streamlining workflows for documentation, billing, and prior authorizations.

Ambient AI scribes record the interactions between a clinician and a patient and transform them into notes, freeing the clinician from the heavy burden of documenting the encounter and allowing him or her to focus more on the patient than on the electronic health record. (Seedream 4)

Before using the AI scribe, 52% of clinicians (mostly physicians) indicated symptoms of burnout, whereas with use of the AI scribe, burnout dropped to 39%. (Statistically significant difference at p < 0.001.) Clearly, this one AI tool doesn’t solve all the problems in modern healthcare, but considering how simple it is, the burnout reduction for critical staff is a clear gimme.

Other statistics that improved substantially with the AI scribe:

• Note-related cognitive task load (significant at p < 0.001)

• Ability to focus undivided attention on patients (significant at p < 0.001)

• Reduce time spent documenting after hours (significant at p < 0.001)

• Create notes that patients can understand (significant at p = 0.005)

• Ability to add patients to the clinic schedule if urgently needed (significant at p = 0.02)

Happy Leif Erikson Day

October 9 is the annual Leif Erikson Day. This gives me an opportunity to try out HunyuanImage 3.0, which just topped the AI leaderboard for text-to-image. This is a well-rendered, dramatic image of a dragonship, but maybe HunyuanImage took too many cues from my dragon token when it added those sea monsters. Of course, this scene is probably what Leif thought he would be facing on his way to explore unknown lands in North America.