“Coding is largely solved,” claimed Boris Cherny, Head of Claude Code, in a podcast from February 2026. He also admitted, “100% of my code is written by Claude Code.” Boris is not the only developer who no longer writes code by hand. Spotify says its best developers “have not written a single line of code since December.” At OpenAI, a team has built and shipped an internal beta of a software product with “0 lines of manually-written code.”

After reading all this news, you might wonder: if AI is already so good at coding and will only get better, do we still need human developers? Or can anyone build software just by talking to AI? Yet if you look closely, you’ll notice that all these people are senior developers with years of experience. Before AI coding tools became available, they wrote countless lines of code, learned several programming languages, and spent a lot of time debugging and solving technical challenges. It’s exactly because of this experience that they can plan projects with AI, guide its code generation, and notice “code smell” when something is poorly done.

With some “vibe coding” tools, anyone without programming knowledge can create a fancy webpage or a simple app. But when a project gets more complex, experience is needed to maintain the code and add new features. Otherwise, you’ll probably find that AI keeps messing up the product, and fixing AI-generated code can take more time than writing it yourself. Even worse, because of the misleading idea of “Prompt Engineering,” some people believe they just need some special prompts to get perfect results. Yet how can you keep writing good prompts if you don’t understand what you’re doing?

As long as AI still needs instructions from humans, the more experience someone has, the more productivity gains they can get from AI. This applies to many other industries too. The problem is humans need opportunity, time, and practice to build experience (it can take ten years to train a good programmer). That’s the danger we’re facing: experienced workers become ten times more productive using AI agents, which eliminates many entry-level jobs, and junior workers won’t have the opportunity to gain experience. In the short term, we may see rising unemployment among young graduates as a tradeoff for a productivity boost. But after current senior staff retire, no one can take over their jobs because the career path has been broken.

On the other hand, by providing quick answers, AI eliminates necessary struggles and thinking along the way to learning. If junior workers always outsource critical thinking and essential skills to AI, they’ll never acquire real knowledge and can’t do much when AI stops working.

In the novel The Machine Stops (1909), human beings become completely dependent on the Machine, which provides everything including food, shelter, and entertainment. Over generations, people lose the knowledge of how the Machine actually works and become helpless when the Machine begins to fail.

We should continue training junior workers in the age of AI; otherwise, we might end up in a situation like the one described in the novel. Companies won’t do this voluntarily, but we might subsidize them to keep hiring entry-level workers, or we could create new training institutions. At the same time, we need to find a way to use AI for learning without becoming dependent on it. It might also be a good idea to leave some tasks for humans just for training, even if AI can do them better.

If we fail to transmit knowledge and skills, then let’s hope the Machine never stops.

Originally published on LinkedIn: https://www.linkedin.com/pulse/cheap-disposable-code-how-ai-helped-our-yqcce/

Designing for multiple brands

In the Lufthansa Group, there are different airlines, each with its own history and identity. When designing our website and app, we need to make sure the design follows each airline’s brand guidelines.

To do this, we created a Design System for each airline in Figma. In practice, this meant designers had to carefully pick the right colors, text styles, and components from the correct Design System every time. Even small changes often had to be repeated manually for each airline.

For example, to create a headline on a Lufthansa page, I first need to add the Lufthansa Design System to my Figma file and apply the correct text style and brand color. If I want to reuse this headline on a Swiss page, I need to copy it, switch to the Swiss Design System, and reapply those styles there, because the definitions are different. This kind of work quickly becomes tedious when we design for multiple airlines.

Design System with Figma Variables

In 2023, Figma introduced a feature called Variables that made this much easier. Instead of maintaining separate Design Systems for each brand, we can now keep one unified system and switch between brands when needed.

Using the same example, we can define different “Modes” such as “Lufthansa” and “Swiss.” Each mode contains the brand-specific values for things like text styles and colors:

Once these variables are set up and assigned to the headline, I can simply switch the Mode in Figma to change the design from one airline to another, similar to switching to a dark theme in some apps:

This change removed a lot of manual work and made designing for multiple brands much faster.

The challenge

However, icons turned out to be the one thing this new system couldn’t handle automatically.

Each airline has its own icon set, but Figma Variables don’t support switching icon files directly. To work around this, we grouped icons into components and stored the different airline versions as variants inside the same component. Each variant has a property called “Airline,” which tells Figma which brand it belongs to.

We then connected this property to an “Airline” variable. With this workaround, we can also switch icons by switching the Mode:

The problem is that our icons are organized in a totally different way. Each airline’s icon is its own component, with variants for different sizes. We must reorganize the icons for the new Design System. Due to the number of icons, it would take us weeks to reorganize them manually.

The plugin, written by AI

Today, we have Large Language Models like Gemini and Claude that are very good at coding. If AI could generate a Figma plugin to reorganize the icons automatically, it could save us a huge amount of time. Fortunately, our challenge is perfect for trying AI-generated code for several reasons:

1. The steps are clear and easy to describe.

2. The plugin is for internal use only and we’ll probably only use it once, so the code can be messy and non-maintainable.

3. We can copy the icons to a new page to try the plugin, so there won’t be any serious damage even if the code accidentally deletes some icons.

4. The task is small. Even if AI fails to generate the code we want, it won’t take much time to experiment.

5. No confidential data is involved.

I chose Google Antigravity to write the plugin because it’s free and I had good experiences with Gemini. After thinking about how to describe the procedure, I fed Google Antigravity the following prompt:

Help me write a Figma plugin to reorganize the component on the current page.

There are many “Main Components” on the page and they are organized this way: 1. the main component name is: [Airline]/[Icon Name]. 2. Each component contains several Variants with different sizes. For example, there’s a main component “LH/close” which contains two Variants “[Size]=24” and “[Size]=16.” The structure is like this:

[Airline]/[Icon Name]:

• Variant: Size=24

• Variant: Size=16

The plugin should reorganize the “Main Components” in this way: 1. Create a Main Component for each identical [Icon Name]. The name is [Icon Name]. 2. Create two Properties for it: “Airline” and “Size” 2. Put all existing Variants with the same [Icon Name] under the new main component and assign the correct value [Airline] to the property “Airline”, [Size] to the property “Size”. For example, there are two components “LH/close” and “LX/close”, each has three variants “Size=32”, “Size=24”, and “Size=16”. You need to create a main component named “close” and put all variants from “LH/close” and “LX/close” under it. The property “Size” should be kept and all LH variants should be assigned “Airline=LH” and all LX variants should be assigned “Airline=LX”.

Before the reorganization:

LH/close

• Size=32

• Size=24

• Size=16

LX/close

• Size=32

• Size=24

• Size=16

After:

close

• Airline=LH, Size=32

• Airline=LH, Size=24

• Airline=LH, Size=16

• Airline=LX, Size=32

• Airline=LX, Size=24

• Airline=LX, Size=16

It generated the plugin code quickly. The first version wasn’t perfect, but the structure was already solid. After some debugging and small adjustments, the plugin could reorganize all icons on a page with one click.

With that in place, the icons could be seamlessly switched using the “Airline” variable in the new Design System.

Cheap and disposable code

Not long ago, code was expensive, and writing this plugin would have been hard to justify. Even an experienced developer would need at least a couple of days and most companies wouldn’t spend a developer’s time on a one-off internal task. We would probably still move the icons manually for weeks, even though we know a plugin would help, simply because it’s cheaper than writing one.

AI changes that calculation. Code has never been so cheap and disposable. For many repetitive and tedious tasks, you can use AI to quickly generate a script or plugin to automate them, if they can’t be directly handled by AI. It doesn’t matter if you only use the code once, because it costs little and the benefit is high if it works.

In one of his talks, mathematician Terence Tao compared Large Language Models to jet engines. A jet engine makes travel much faster, but you can’t just attach it to a car. You need to invent the plane and develop all the safety protocols around it. The change seems simple and one-dimensional (it’s just faster), but it created the aviation industry and changed our understanding of travel.

Cheap and disposable code may have a similar impact on our daily work. Imagine if everyone is equipped with AI that can automate small tasks by generating code. How would that change the way we work and collaborate? I’m looking forward to seeing where this leads.

Almost two years ago, my colleague Sebas and I started an AI workshop within Lufthansa Group. The format was simple: first we introduced the tools available on the market, then split participants into groups to solve playful challenges. One example: “Ask ChatGPT to write a short story using only emojis.”

The game, designed by Sebas, turned into team exercises where colleagues learned by experimenting. Back then, LLMs were still in their early days. GPT-3.5 had just arrived, Sora was still a myth, voice models were far away, and AI-generated images were full of typos. It made the workshop easy to run because most people had never used ChatGPT. They were amazed just to watch it draft an email. I still remember a colleague’s excitement at discovering: “I can just tell ChatGPT to make the text shorter!”

Two years later, AI development has taken many leaps. Image generation now follows instructions with precision; models offer “deep thinking” features; AI agents that can act on your behalf, are emerging. Yet our workshop hasn’t evolved much. Unsurprisingly, we’ve received more feedback like, “I didn’t learn anything new because I already use ChatGPT every day.”

Sebas and I really want to upgrade our workshop, but we face several challenges:

1. The general audience has no interest in too much technical content.

We could explain neural networks and training processes, but our participants are not engineers. Most people do not want to learn how AI works; they just want to know how to use it. And that makes perfect sense. You do not need to know how to build an air conditioner to enjoy its cool air.

2. No one fully understands how LLMs work.

LLMs (Large Language Models) are called ‘Large’ for a reason: they contain millions, often billions, of parameters. Experts know how to train them, but no one can say exactly how all those parameters interact. Understanding the architecture won’t necessarily make you a better user.

3. AI is taking over the agency.

When we teach Photoshop, beginners start with the basics, like how to select an object. More advanced users move on to features such as creating masks.

AI tools are totally different. There are no extra menus or hidden buttons to learn. You do not click around. You simply tell AI what you want. “Help me add text to the photo” is the same plain language as “Help me remove the dog and add an airplane in the background.”

In Photoshop, those two tasks require very different levels of skill and time. With AI, both prompts take the same skill to write: the ability to express what you want in natural language.

4. Prompt engineering is a bug, not a feature.

Because no one fully understands how all the parameters in large language models behave, people experimented with prompts and tricks to get better results. But these shortcuts quickly become obsolete, since LLMs are improving and learning to understand us more naturally. In the near future, if a human can understand what you mean, AI should be able to as well.

So, what should we teach people to use AI?

Instead of more technical content, we should focus on the real interface of AI: natural language. The better you think, write, and communicate, the better you will use AI.

Ironically, the best training for that has existed for a long time: literature, philosophy, history… all these liberal arts nurture critical thinking and communication. In the last few decades, liberal arts have been dismissed as impractical, while STEM (science, technology, engineering and mathematics) was seen as the safe path. But as AI advances, liberal arts may rebrand themselves as “AI Engineering” or “AI Management.” Studying liberal arts might make you a better programmer in the future.

In our next workshop, after showing all the tools, I don’t want to teach “prompt engineering.” I want to talk more about writing. Because in the long run, the best AI skill isn’t knowing the model’s parameters. It’s knowing how to express your own ideas clearly.

Designers are often asked to explain what makes a design “good.” It's a fair question on the surface, but answering it isn’t always easy—or even possible in the way people expect. Some look for a set of rules, a checklist, or a formula. But good design doesn’t always work that way. Sometimes, you just know. And that can sound suspicious, especially if you can’t explain it clearly. But that kind of judgment isn’t empty. It’s not just a gut feeling. It’s something more, and I believe the philosopher Ludwig Wittgenstein’s idea can help us understand why:

Consider, for example, the activities that we call "games". I mean board-games, card-games, ball-games, athletic games, and so on. What is common to them all? - Don't say: "They must have something in common, or they would not be called 'games'" — but look and see whether there is anything common to all. — For if you look at them, you won't see something that is common to all, but similarities, affinities, and a whole series of them at that. To repeat: don't think, but look! - Look, for example, at board-games, with their various affinities. Now pass to card-games; here you find many correspondences with the first group, but many common features drop out, and others appear. When we pass next to ball-games, much that is common is retained, but much is lost. — Are they all 'entertaining'? Compare chess with noughts and crosses. Or is there always winning and losing, or competition between players? Think of patience. In ball-games, there is winning and losing; but when a child throws his ball at the wall and catches it again, this feature has disappeared. Look at the parts played by skill and luck, and at the difference between skill in chess and skill in tennis. Think now of singing and dancing games; here we have the element of entertainment, but how many other characteristic features have disappeared! And we can go through the many, many other groups of games in the same way, can see how similarities crop up and disappear.

And the upshot of these considerations is: we see a complicated network of similarities overlapping and criss-crossing: similarities in the large and in the small.

— Ludwig Wittgenstein, Philosophical Investigations, §66

Forgive me for quoting the entire paragraph from Wittgenstein’s book, but I believe it’s important and deserves to be read carefully, perhaps even more than once. Some concepts, like “rectangle,” have clear definitions that apply to all instances. But as Wittgenstein points out, many other things may share the same name without sharing a single, essential feature. Their meanings are held together by overlapping similarities rather than a fixed essence.

That’s why Wittgenstein introduced the idea of family resemblance (Familienähnlichkeit): members of a family may share overlapping features, such as eye color, voice, or mannerisms, without all having the same exact trait. For concepts like “game,” asking “What is a game?” can be misleading, since there’s no strict definition that applies to every case. A better question is, “What do we call a game?” That shift helps us understand the concept by looking at the overlapping similarities among things we use the same word for.

“Good Design” is similar to the concept of “game.” Despite advances in neuroscience and psychology, there is still no strict, single definition of “good design”—even though many people try to create one or claim ownership over it. But that doesn’t mean we can’t recognize a good design when we see one. Good designs share overlapping similarities. The more good designs you’ve seen, the better you can judge a new one.

That’s why experienced designers usually have better judgment, even if they struggle to explain exactly why. I’m not saying experienced designers never make mistakes, but it’s still more reasonable to trust their judgment than to rely on someone without design experience, or on rigid rules that are bound to fall short sooner or later.

Still, the question “What is good design?” is so pervasive that designers are often expected to answer it. And if they can’t, people may assume they lack principles and are just following their “gut feeling.” “It’s only your feeling” can be a serious insult to experienced designers. It’s almost impossible to defend against, especially given the lack of a strict definition, as mentioned earlier.

And if it’s all just gut feeling—something everyone has—then why should your gut be any better than mine? As a result, more and more design decisions are made either by the most powerful person in the room or blindly driven by data, rather than by experienced designers.

It may sound arrogant, but in many cases, designers should feel comfortable saying, “I know it’s a good design because I’ve seen many other good designs,” without needing to give a further explanation. That’s not gut feeling; it’s experience. And experience matters. We trust seasoned professionals in other fields—chefs, musicians, surgeons—without always demanding a step-by-step justification. Design should be no different.

If we don’t recognize the value of experience, we risk replacing thoughtful judgment with data-driven defaults or hierarchical decisions. That may seem safer, but in the long run, it erodes what makes design meaningful in the first place.

After unsuccessfully searching for a design community in Frankfurt, I applied to become a Figma Community Advocate in March 2020 and started building Friends of Figma, Frankfurt.

Three years later, we have a small but active group in Frankfurt that hosts regular in-person events, including networking meetups and speaker sessions. Today, I’d like to share my experience and thoughts on building a local (Figma) community, in the hope that it might help you get started.

Determined to start

To build a local community, the first and most important thing you need is determination. There’s a reason why most people prefer to join an existing community or wait if they can’t find one: there are simply too many obstacles along the way. You have to find a good venue, contact potential speakers and sponsors, grow the community, maintain a good vibe, and handle many other tasks that come up unexpectedly. Without determination, you won’t go far.

Where you live matters

If you live in a big city like San Francisco, New York, or Barcelona, you won’t have trouble finding enough people to start a community. But if you live in a smaller city like Frankfurt, the challenge is on a different level.

When I started building Friends of Figma, Frankfurt, I tried everything I could to recruit new members. I posted on the Frankfurt subreddit, shared in Facebook groups, and attended after-work meetups to ask if there were designers. After several months, I finally found a few designers, and we held our first in-person meetup.

So, if you live in a smaller city, lower your expectations and be patient—especially avoid comparing your community to those in big tech hubs.

Apply to become a Figma Community Leader

I like Figma. It's cool and is one of the few tech companies that doesn't plan to monetize our attention. It reminds me of the era before social media, when many tech companies were genuinely interesting. Figma has a program called Friends of Figma, which includes communities based in various cities or countries. Anyone can apply to become a leader if there's no community in their city yet, or a co-leader if one already exists. You can read all the details here: https://friends.figma.com/become-a-user-group-leader/

Although you won’t get paid, your community will receive free Figma Professional Plan, swag, and—most importantly—the trust that comes from being an official community. It's much easier to attract new members when people see your group is officially recognized. Organizing events for designers is also enjoyable because designers are usually good at talking with each other. After all, it's part of their job. Your role is just to bring them together and enjoy the atmosphere.

Start with in-person meetups

It’s tempting to start with online events. People from around the world—not just your small city—will RSVP. You’ll feel excited seeing a high number of RSVPs and expect a great event.

However, moderating online events is difficult and requires strong skills. It’s also hard for attendees to build connections online. Even worse, most attendees won’t show up at all or will get bored within the first five minutes and leave. Often, people click RSVP simply because it's easy, not because they they genuinely want to attend your event.

In my opinion, online events only work well for mature communities whose members already know each other. They're not effective for forming new communities. Always start with in-person meetups. Even if fewer people RSVP, those who attend will truly enjoy the event and are more likely to become active members who help the community grow.

Spread the word

Share the event link everywhere! Post regularly before the event to remind people. Take photos during the event and share them afterward on social media. People like seeing other people's faces (I still don't fully understand why). Photos also show potential attendees that your event is friendly, welcoming, and worth joining next time.

Active members matter

You recognize potential active members when you see them. They arrive on time and message you if they can’t. They offer help. They like talking to people. Take care of your active members. Listen to them. With them, you can always keep the community going and have wonderful events—even with only a few attendees.

A small community is also good (or even better)

Don’t feel discouraged if your community is small. In a small community, members can have deeper conversations and get to know each other better. Keep organizing regular events, even if you only have five members. The true meaning of a community is to create connections between people, not to grow endlessly. Deep human connections aren’t scalable. The world would be a much better place if we rejected the Silicon Valley philosophy of scaling everything.

“Make the calls myself”

Some people hesitate to build a community because they're anxious about talking to others or believe they lack social skills. They think, “I don’t even attend many meetups. How could I possibly organize one?”

In reality, building a community is one of the most effective ways to overcome this anxiety. One of my favorite writers, Susan Sontag, once wrote:

“The way to overcome X is to feel (be) active, not passive. I feel anxious when the phone rings - therefore I don't answer or I get someone else to. The way to beat that is not to force myself to answer the phone. It is to make the calls myself.”

—Susan Sontag, Reborn: Journals and Notebooks, 1947-1963

This is some of the best life advice I’ve ever come across. By becoming proactive and taking the initiative, you’ll be surprised at how much less fear and anxiety you experience.

The importance of local community today

People are spending more and more time online, especially after COVID. Some may argue that connections can be formed digitally, but in my opinion, there are no real connections in the online world.

This isn’t just because it lacks multi-sensory interactions—like eye contact, body language, and vocal tone—but also due to the echo chambers created by algorithms. Most of the time, you’re only seeing a reflection of yourself—not other people.

This might sound very old-fashioned: I still like the physical world. This world. It’s far from perfect, but it’s still the only world where your body exists. In local community events, you must learn how to talk to people with different opinions. You can’t ghost them or turn off your camera. You have to be polite. It’s beneficial for your social skills and also crucial to our democratic society, because democracy is an ongoing conversation. If people stop talking to each other, democracy itself is at risk of collapsing. To build a local community is to create a space where people can encounter each other and have conversations.

We need local communities more than ever—and you can start building one today. I wish you a successful and vibrant local community.

Because of the recent advancements in AI, I’ve started rewatching some sci-fi movies. One of them was I, Robot, a fantastic film starring Will Smith. There’s a scene that struck me deeply: after arresting a robot, Will Smith’s character begins questioning the robot’s capabilities. He asks, “Can a robot write a symphony? Can a robot turn a canvas into a beautiful masterpiece?” The robot doesn’t directly answer but instead asks back:

“Can you?”

This movie came out in 2004, and now, 20 years later, despite the incredible progress of large language models, many people still criticize AI with the same questions Will Smith asked. I find it amusing that humans tend to compare AI’s capabilities with the best minds in human history rather than with the average person. Many people make mistakes daily, yet still believe they’re above average in intelligence (https://pmc.ncbi.nlm.nih.gov/articles/PMC6029792/). However, if AI makes mistakes or isn’t as good as Mozart, it’s dismissed as not being “real intelligence.”

But this skepticism toward AI isn’t new. There’s a well-known phenomenon called the “AI effect” (https://en.wikipedia.org/wiki/AI_effect): Every time there is a breakthrough in AI, such as mastering chess, people tend to dismiss it as mere computation. This attitude toward AI often leads us to underestimate its progress and impact: since AI is not “real intelligence” as long as it isn’t smarter than Einstein, why should we worry about it before it becomes truly “real”? However, if we measure AI against average human performance—rather than humanity’s greatest minds—and consider an average human intelligent, we must acknowledge that AI is intelligent—as it has already surpassed human performance on several benchmarks (https://aiindex.stanford.edu/wp-content/uploads/2024/04/HAI_AI-Index-Report-2024_Chapter2.pdf). Once we recognize this, we’ll realize how fundamentally AI will transform our world.

AI also doesn’t need to become a great artist to replace great artists. Only a few artists are truly great, and most people are satisfied with mediocre entertainment. If AI can flood us with “good-enough” short videos and superhero movies, most people will be content and won’t criticize AI for not being able to direct another Godfather. The real crisis isn’t that AI will surpass great human artists—human artists may still be far better than AI—but that they’ll lose their audience and market, making it impossible to create another masterpiece. You don’t have to destroy a book by burning it; it’s already destroyed when no one reads it anymore.

When we criticize AI, it’s worth reflecting on the question the robot asked in the movie. AI might not always understand an instruction, but do you? AI might not write plays as great as Shakespeare’s, but can you? AI isn’t perfect.

But are you?

After years as a UX Designer, I’ve come to deeply appreciate well-crafted user onboardings for new products or features. It’s not just about UX design—any qualified UX Designer can do it—but rather about everything else within the company. Too often, Product Managers and other decision-makers view user onboarding as a “nice-to-have” because they’re focused on delivering more features. However, a feature isn’t truly delivered if users can’t find it or don’t know how to use it.

Some years ago, I tried to persuade the team to implement user onboarding. After reviewing the design, a developer argued, “Our users aren’t stupid!” The whole team laughed. That’s when I realized they didn’t need a designer, and I soon left the company.

To implement good user onboarding, a company must prioritize it and allocate resources, which requires decision-makers with common sense—a rare quality. Fortunately, several companies still truly care about user experience, and Nintendo is one of them.

Recently, Nintendo released a new device called Alarmo, an adorable alarm clock. After setting it up, I can’t help but be amazed by its user onboarding. I decided to write a detailed breakdown to learn it by heart and hopefully inspire others. The world would be a much better place if all new products had such great user onboarding.

More in the PDF as slides:

Nintendo Alarmo

16MB ∙ PDF file

Download

Download

Please read it on LinkedIn if you don’t want to download the PDF: https://www.linkedin.com/posts/jackschwarz2017_how-nintendo-alarmo-nails-user-onboarding-activity-7260331596422037507-aGE4

Historian Yuval Noah Harari has talked a lot about social media in his books and interviews. He is not completely against social media and acknowledges it can help connect people who otherwise couldn’t find each other—He himself met his husband in an online community.

However, there are two major issues that make social media a “chaos machine” that keeps spreading misinformation and polarizing our society: bots and algorithms. Malicious actors can create thousands of bots to efficiently poison discussions. It’s pointless to try to persuade bots, but bots can slowly change people’s opinions by flooding them with misinformation. In the real world, it’s hard to be manipulated or influenced because it’s rare for everyone around us to have the same opinion and talk to us endlessly. Yet, in the digital world, we can be constantly surrounded by bots repeating misinformation, creating the illusion that most people disagree with us, which makes us very vulnerable to manipulation.

We should ban bots on social media because, as Yuval argued, “bots don’t have rights.” Humans have freedom of speech, but bots don’t. Therefore, it’s neither a legal nor ethical issue to ban bots on social media. In an interview, Yuval half-jokingly suggested that people should bring their ID to the town hall and wait in line to create an account on social media. Sure, it would be slow to create an account, but in this case, slowness is a feature, not a bug, making it almost impossible to create bots.

Another big problem with social media is the algorithm. To maximize profits, social media companies train their algorithms with one goal: to maximize user engagement. After analyzing vast amounts of data and conducting countless experiments, the algorithms have learned that moral outrage is the most effective tool to increase user engagement, so they keep promoting such content, which harms our mental health and polarizes our society.

To solve the problem of the algorithm, Yuval proposed holding social media companies accountable for their algorithms. For a long time, social media companies have tried to confuse the public by mixing user-generated content with the algorithm itself. Whenever misinformation goes viral, they blame the user who created the content and argue they shouldn’t censor users. Yuval agrees that we should be cautious about censoring users, but it’s the algorithm that decides to promote such content. If an editor from The New York Times put hate speech on the front page of their newspaper, even if someone else wrote the content, the editor would still be held liable. The editor can’t say, “It’s not my fault because I didn’t write the content; I just put it on the front page.”

Similarly, if misinformation is amplified by the algorithm and causes harm, social media companies should be held accountable and punished. They may argue that it’s unfair because the algorithm is so complex that even its creators can’t fully understand it. However, if they hadn’t trained the algorithm solely to maximize user engagement, it would behave differently. They should train their algorithms with other goals as well, such as promoting truth, if they want to avoid severe penalties.

We’ll have a much better internet if there are no bots and the algorithm stops exploiting our negative emotions.

Recommend Yuval’s new book Nexus: A Brief History of Information Networks from the Stone Age to AI

You’ll understand what I mean if you’ve lived in Germany :) Otherwise, feel free to ask AI to explain it to you.

On August 21st, A16Z published its latest report of the top 50 AI-first web products (by unique monthly visits) and top 50 AI-first mobile apps (by monthly active users). It’s no surprise that OpenAI’s ChatGPT tops the list, and that many people are using Claude AI and Perplexity. However, many products on the list are not well-known to the public and don’t often appear in the headlines. So, why do they have so many users? What did they do right?

After reading the report, I went through all the products to understand the technology they’re using and the problems they’re solving. As expected, most products don’t have their own model, as training and fine-tuning a model is extremely expensive. Instead, they use APIs from OpenAI or other models and then build features on top of them. What makes them successful isn’t the technology itself, but the business strategy and the features that meet users' needs.

Imitate ChatGPT, both visually and functionally

Every time a popular app emerges, hundreds of copycats flood the app stores, using similar icons and names. Many people are misled into downloading these copycats, and people continue using them if they offer basic functionality. ChatGPT is no exception to this phenomenon. Unfortunately, this dubious strategy works, as you can see in the list of top 50 AI-first mobile apps. Just look at their icons and names—can you really tell the difference?

Improve my photos and my attractiveness

Image and video generation are impressive and seem almost magical, so no wonder products like Midjourney and Luma are popular. However, there’s also a strong demand for basic functionalities, such as easily removing image backgrounds. Many products in the list, such as Pixelcut, Pixlr, SnapEdit, and Remove It, have gained a large user base by offering simple, user-friendly photo editing tools. People love taking and sharing photos online, and if you can simplify this process, you’ll attract many users, regardless of the technology you’re using.

Another thing people care about is their appearance. "Do I look attractive?" is a concern for many men and women. Before the advent of AI, people would post their pictures online and seek ratings. AI now offers a new way for people to judge themselves. LooksMax AI and Umax have gained popularity by simply offering one key feature: analyzing your selfie and providing advice on how to enhance attractiveness.

Get me a girlfriend / boyfriend

When I tried to open JanitorAI at the office, I got a browser warning that our company had blocked all NSFW content. This gave me a vague idea of what this site was about. With the advent of Large Language Models, we can now have human-like conversations with chatbots, leading to the development of AI girlfriends and boyfriends. SpicyChat, Candy.ai, and Crushon AI—just from their names, you can guess what services they offer. Some of these are even more popular than the well-known Midjourney and Hugging Face.

AI companions are very appealing because they’re always available and don’t have their own emotions. They can be a great help when people feel lonely, especially in the post-COVID world. Loneliness has become a big market opportunity for AI products. However, this new dynamic could dramatically alter our relationship with other humans and society, which poses significant risks.

Products from Chinese companies

TikTok’s parent company, Bytedance, is a Chinese company, but people in China are blocked from using TikTok and can only access 抖音 (Douyin), the Chinese version of TikTok. This strategy allows Bytedance to apply strict censorship to Douyin without affecting TikTok. To survive in the Chinese market, many companies, including LinkedIn and Duolingo, have adopted this best practice from Bytedance.

Several popular AI products on the list are from Bytedance, such as Gauth and Coze. Another product, Question AI, is from a different Chinese company, Zuoyebang. Talkie, meanwhile, is owned by the Chinese tech company MiniMax. However, if you visit the products’ websites, you won’t find any information about the teams behind them, and all these products are released by companies registered in Singapore, not China. This seems to be a trend for Chinese AI companies: register in Singapore, release products for overseas market, and block access to the products within China.

Although AI is still in its early phase and many of these popular Gen AI products may fail quickly, this list from A16Z is nonetheless insightful. Not every company can develop state-of-the-art models or have an unbeatable technological advantage, but companies with smart business strategies and user-focused features can still succeed in the market—just as they did in the era before AI.

Recently, Elon Musk’s xAI released a new AI model, Grok-2, which features “frontier capabilities in chat, coding, and reasoning.” By integrating a state-of-the-art AI image model named “FLUX.1”, Grok-2 can also generate high-quality images. However, FLUX.1 is not from xAI but from Black Forest Lab, a startup with most of its co-founders based in Freiburg, Germany.

In the last decades, Europe has been almost forgotten when it comes to digital innovation and information technology. The most successful tech companies in the world are from the US, particularly from Silicon Valley. Look at Apple, Google, Amazon, and Facebook. China has also done a good job of incubating numerous big domestic tech companies and several well-known international ones, including Alibaba and TikTok. But you haven’t heard of many fast-growing tech companies in Europe—maybe Spotify is an exception. Europe is known for data protection and forcing Apple to use USB-C, but it’s not known for great digital products.

However, with the rise of AI, things may hopefully change in Europe.

According to the 2024 AI Index Report, while the US remains far ahead, there are 21 notable AI models in the European Union, with 8 in France and 5 in Germany. Complementing this is the rapid growth of AI expertise; a study from last year found that Europe has more AI talent than the US, with the number of people working in artificial intelligence across Europe increasing by 1,000% over the past 10 years. London, Berlin, Paris, and Amsterdam have become hubs for GenAI startups, and a new AI hub is being built in the small German city of Heilbronn.

Europe is already home to several world-class AI companies and institutions. Black Forest Lab, which we mentioned earlier, is one of them. Another example is Mistral AI, based in Paris and renowned for its advanced Large Language Models. Also in Paris, Kyutai, an AI research lab, has received significant investment from billionaires Xavier Niel, Rodolphe Saadé, and Eric Schmidt. In July, Kyutai just released an AI model with unprecedented vocal capabilities called Moshi. DeepL, a company focused on AI-powered translation, was founded in Cologne, Germany. And in Sofia, Bulgaria, INSAIT is one of the world’s leading AI research institutes.

AI is the next generation technology. It will transform our society in the near future and bring significant opportunities to Europe. While it’s true that Europe doesn’t have its own search engine or social media giants, it may no longer need to build them. We can bypass the previous generation of technology and directly adapt AI.

This situation is similar to that of the fax machine. Germany adopted the fax machine after its invention, and it’s still widely used in many institutions. However, countries that never fully embraced fax machines now don’t need to—they have smartphones and internet and don’t even need to know what a fax machine is.

In the same way, if we approach the AI revolution swiftly and wisely, Europe can reinvigorate itself as a significant player in technology and innovation without needing to build another Google or Facebook. There’s no reason we can’t achieve this. We have a solid education system that can supply enough talent for the industry. Comparing to the US and China, the gap between rich and poor is much smaller here, society is less polarized, and the quality of journalism is far higher, making Europe an attractive place for well-educated AI talents from around the world.

Additionally, our strong regulation and commitment to data protection can also be an advantage, as the potential risks of AI are extremely high, and we need strong institutions to balance the power of AI companies. It’s better to be a bit slower if it means being safer.

Europe can become a leader in the global AI landscape, and having an AI power that genuinely cares about privacy would be good for everyone. This would particularly benefit young Europeans, who could study AI and work in companies at home instead of moving to the US. Europe should seize this opportunity and take on the responsibility of shaping a better future for the world.

Many people compare AI to previous machines and technologies in history, such as radio and computers, and then conclude that although AI will displace some jobs, it will also generate new ones. However, AI is fundamentally different from any previous technology.

Radio can’t create content; humans create content and transmit it using radio. Computers can’t make decisions; humans make decisions and use computers to execute them. But AI can create new ideas and make decisions autonomously. It’s a false analogy to compare AI to earlier machines, and one can only draw false conclusions from a false analogy.

Humans learn skills and perform jobs using their intelligence. If AI reaches human-level intelligence, there’s no reason it couldn’t replace any job that humans can do. AI also wouldn’t create maintenance jobs for humans. Have you ever seen a panda doctor or a monkey nurse in a hospital? Probably not, because humans take care of humans. With human-level intelligence, AI can also take care of itself.

Historian Yuval Harari argued that the term “AI” could also stand for “Alien Intelligence” because it’s very powerful, and we don’t fully understand how it works. Mustafa Suleyman, CEO of Microsoft AI and co-founder of DeepMind, referred to AI as “another species” in his TED talk. “Alien” and “a species” are better analogies for AI. If you think of AI as another species with human-level intelligence, you’ll have a clearer picture of the risks that AI can bring.

AI is “horizontal.” It won’t just target specific jobs—no job is safe from AI. Even if AI doesn’t reach human-level intelligence and only increases our productivity, we’ll still need far fewer jobs in every industry than we do now. You may still have a job, but you won’t feel secure living in a society where 60% of people are jobless. That’s why the question “Will AI take my job?” is misleading; it causes you to focus on individual solutions and gives the illusion that some jobs are safer than others. A better question is, “Will AI take all jobs?” This would lead us to seek a societal solution, which is the only solution.

In 2017, The New York Times published a report titled "The Robots Are Coming, and Sweden Is Fine"(paywalled), discussing how people in Sweden view robots and other forms of automation. People in Sweden aren’t afraid of advanced technology—not because they have mastered it or are smarter than people in other countries, but because Sweden has a generous social welfare system that protects people when they’re jobless. Sweden also supports workers in acquiring new skills if new technology creates new jobs.

Facing the joblessness that AI can bring, we need a society with a strong safety net and solidarity, like Sweden. In such a society, people won’t feel desperate or lose their dignity immediately if AI takes their job. They will have time to learn how to use AI if that’s an option. This kind of society could also implement Universal Basic Income (UBI) for everyone so that people can simply enjoy their lives while AI works for them.

In addition to the welfare system, we should also start to rethink the meaning of a job. In modern society, we attach a lot of significance to a job, such as social status, dignity, and a sense of purpose. Being jobless is often seen not just as poor but also as useless and despised. However, at the same time most people don’t enjoy their work and would quit immediately if financially secure. If we stop attaching these meanings to a job, people might welcome “losing” their jobs if they can still receive income from society.

Ever since machines were invented, many have dreamed that one day machines would do all the work, freeing people from labor and providing more leisure time for everyone, not just the elites, to learn about philosophy and art. Robert Maynard Hutchins, the American educational philosopher, once wrote, “The substitution of machines for slaves gives us an opportunity to build a civilization as glorious as that of the Greeks, and far more lasting because far more just.” Unfortunately, machines have not lived up to this promise. But with AI, we might finally be freed from mundane and boring work—if we handle AI with the care and wisdom it demands.

Yes, AI may take all the jobs, but it could be a desirable future.

If you’ve never heard of Terence Tao, he is considered “one of the greatest living mathematicians.” Recently, he gave a talk on “The Potential for AI in Science and Mathematics” and shared his thoughts on AI and its possible applications in science and mathematics. Although I didn’t understand the mathematics part of his talk, his opinions on AI were very inspiring. I decided to take some notes from the talk and share my thoughts on them.

Quoted text is from Terence Tao; the rest are my comments. Any errors, typos, or omissions are solely my responsibility.

…what AI basically is, in non-technical terms, it’s a guessing machine…the way it does it is actually quite mundane mathematically. You just take up your your input and you encode every word or whatever as as a number, you multiply these numbers by weights, and you combine them and maybe you truncate them, and then you multiply by weights again, and combine them and you do this a couple hundred times or something…it's actually mathematically rather boring. How you find the weights that that's a bit more interesting…

Large Language Models (LLMs) work like Terence described. Take ChatGPT as an example: it’s a machine that tries to “guess” the next possible word based on your input. After a new word is generated, it’s appended to the original input as the new input, and the process repeats. So, ChatGPT isn’t actually “answering” your question; it’s “completing” your input. Since your input is a question, it generates text that seems like an “answer”.

Some people dismiss LLMs because of this, saying, “It’s just predicting the next word.” However, as Geoffrey Hinton once argued, predicting the next word is never easy. If you’ve ever given a talk in your second language, you might have felt like you were “predicting the next word.” To do it well, you had to understand your topic and your audience, which is never easy. We could better understand AI if we try not to mystify certain internal processes within the human brain.

…it's not magic, but it can accelerate all kinds of things. Maybe an analogy I like to give is that imagine in this world that flight had not been invented. We just had cars and trucks and ships, just land and sea based travel. Then someone at some point invents the jet engine. Initially these engines are very small and and and just toys. They don't do anything. But they get bigger and bigger and more powerful. Eventually they can enable travel that's 10 times faster than the fastest land based vehicle. But you still have to invent the plane, right? You can't just strap a engine onto a car and expect good results. That is not a good idea. You have to change the way you you think of transportation. You have to design new safety protocols, new instrumentation, [you need] new ways to understand the laws of physics. But it's still not magic. It's not like a star ship transporter. It still obeys the laws of physics, but does it at a different scale.

AI doesn’t have to be true magic. Even if AI could only accelerate existing technology by 10 times, it would completely transform society. Cars are not magic either; they are just faster than walking. Yet because of cars, we had to redesign roads and cities. People were able to live in the suburbs and work in the city. Cars changed the world and our lifestyle by “simply” being faster.

We’re used to computer tools - email, search engines, programming languages, etc. - being fussy and uncreative, but also reliable and predictable. But now we have Large Language Models such as ChatGPT, that can understand natural language requests, and generate endless creative text and image outputs - but without reliability, predictability, or accuracy.

Yes, AI is already creative.

…normally when a program or technology produces something bad, you can kind of tell it's bad. It just doesn't look like the real thing. But AI by design, you know, these weights are chosen specifically so that the answers resemble correct answers as much as possible. So even when they're wrong, they look very very convincing. That's a dangerous combination if we are using existing senses of how to detect when something looks good and looks bad…

I think we’ve all noticed that even when we tell ChatGPT it was wrong, it often generates another incorrect answer with a very confident tone. 🤖

…it's the same as if you invent the jet engine, you can quickly mock up some sort of flying vehicle out of it, but with decades before you would get to a state where it's really safe for the general public. On the other hand, air travel is the safest travel today, despite sort of being an obviously dangerous technology. These problems will be solved and they are solvable. But you have to actually think about safety. You can't just sort of assume it’s going to happen…

How can we use a tool that is powerful, but unreliable? For applications where mistakes can cause real harm (e.g., medicine, financial decisions, personal advice and therapy), one must be cautious, despite the great potential benefits. But there are more promising use cases if the downside of producing an incorrect answer is low (e.g., if one wants to generate background images for slides). In particular, in situations where the Al output can be independently verified, there are many promising applications, both in the sciences and in mathematics. For instance, Al is just beginning to be used to generate potential candidates for new materials in materials science (e.g., new superconductors), or potential candidates for drugs to treat diseases. These candidates can be tested by synthesizing them and performing physical tests or clinical trials. These are expensive, so the potential savings of Al in isolating a small pool of promising candidates is large.

…to give another analogy: current science is is like we have these taps of water that produce a certain amount of drinking water, but there's a certain limit as how much output these taps can produce. And suddenly we have this big fire hose that can turn out 10 times 100 times as much liquid, but not in drinkable form. But if you have a filtration unit that allows you to filter out the undrinkable part, then suddenly you actually have a huge spot of drinking water. That's sort of what I see science and mathematics coming to.

…there are many sciences where the bottleneck is finding good candidates to solve problems…maybe you are in drug design and you want to find a drug for a certain disease. You have to synthesize it, you have to first come up with a drug either from nature or by modifying existing drugs. Then you have to synthesize it, you have to trial it, phase one trials, phase two trials, there's a multi-year process of Trials…if there was a way to cut down on the number of candidates. So what you could do is that you could use AI. They're already using AI to to model proteins now…you can find promising candidates for drugs using AI for various diseases, but then you still have the clinical trials afterwards. So you still have the gold standard of scientific validation, but instead of having test 100 candidates maybe you just have to test 10 before you find one that works. The same for many other Sciences. Material Science is another area where there's going to be a big win.

Another promising use case of Al in science is to accelerate modeling, such as climate modeling, as existing models can be used to validate the Al ones. Traditionally, a climate model for our planet can take months of supercomputer time to generate a single simulation of (say) the next twenty years, at a 10km resolution. While this is already valuable, these single simulations cannot accurately capture outlier ‘once-a-century’ extreme weather events, and are not fine scale enough to make tailored predictions for individual communities. AI technologies (such as physics-informed neural networks) can "downscale" a low resolution climate model to a high resolution one, or create new climate simulations at rates tens of thousands times faster than traditional supercomputers.

…AI in principle can shortcut [modeling] quite a lot, because what AI can do is that if you have lots of data of existing simulations, you know, painfully acquired through all these supercomputers, AI can train on all these models and find plausible best fits for predicting what the outcome should be based on inputs that haven't been seen in the in the data. In the area of climate simulation there's already a big success. You can successfully recover the accuracy of traditional supercomputer simulation in a matter of hours, in instead of months. The acceleration is really really remarkable…

A particularly promising use case for Al is in mathematics:

- Little downside to producing incorrect proofs of mathematical theorems

- Proofs can be independently verified by proof assistants

- Training Al to improve their mathematical reasoning may be of use in broader tasks

Fireside chat part:

…things that humans find hard AI can find easy, and things that humans find easy AI finds hard. So this is a fascinating space. I think one thing that AI research is teaching us actually is not so much artificial intelligence, but human stupidity. For example, speaking language conversationally like a human was considered a pinnacle of of intelligence, because the other animals can't do this. But what these AI tools do is that they're predicting the next word to say. They produce some string of of text and each string of text they just find what is the most likely word to come next. It's basically a fancy version of of just pressing the autocomplete button on your phone over and over again, which normally produces rubbish, but passes certain point it actually becomes reasonably coherent. So one thing we realize is like the way I'm talking to you right now is just basically I'm coming up the next word to say. I'm improvising. And that's what these these tools do, they improvise. They make good guesses as to what comes next. So there are certain mathematical arguments which seem complicated to humans but are kind of natural, like each step is a logical thing to do from the previous steps. It could be many steps but in those context AI does very well. But whereas if you ask it to solve a numerical problem like the arithmetic problem, it's guessing, because there's too many different possible outcomes…

…one thing about AI models right now is that if you want to train a neural network for your own specific problem, you have to collect your data, you have to figure out what correct model to use, there's no off-the-shelf thing that you can just - there are, but almost certainly will give you rubbish. If you had sort of standardized models for different sciences, and standards and guidelines: here are some recommended choices for hyper parameters and here are some standard data sets you can use to train on. And if there are some central repositories for these sort of things, that could certainly help. Having open source models that people can build on. Having competitions, actually I think is is a very important aspect. A lot of innovation has been driven by specific challenges, you know, ‘can you get the success rate of this image recognition benchmark above this percentage?’ Then you can actually really quantitatively see what technologies are actually moving the needle forward and which ones are just vaporware. The current a models are really energy intensive and computing intensive, we're going to need research to to find more lightweight models that can actually fit on like a computer that can actually be put in a lab…

…in the past five 10 years, I I've learned that what we think of as hard for humans is often easy for mathematics. In the '90s we thought Chess was the pinnacle of human activity, and then AI solved that - not this AI but more traditional AI - that was a trick. It was brutal force. Go is what humans can do uniquely, then those are different techniques that basically solved Go. Image detection at one point was hard, voice recognition, translation, one by one these things have all fallen. And it was creating poetry and art was considered like one of the last things that AI would do, and it was one of the first [things] that the AI models could mimic at least. So I think our conception of what's difficult actually needs calibration, I mean the problem is that we've only had one working model of intelligence until very recently. Now we maybe have one and a half.

— End —

Many small businesses can’t afford to develop and maintain their own websites, so they upload the information and images of their products on Amazon. Based on that, Amazon generates product pages and even microsites. These pages may not be pretty and are often resemble others on Amazon, but they are cost-effective and allow small businesses to leverage Amazon’s vast customer base.

In contrast, companies like Gucci are wealthy and care a lot about their brand, so they spend a lot of money to design and build their own websites. In the age of AI, these two types of businesses are likely to adopt different strategies: small businesses will try to benefit from existing AI services, while companies like Gucci will integrate their own AI solutions.

AI-powered Web Hosting Service

One potential AI service for small businesses could be a Web Hosting Service. By simply uploading the product list, business introduction, and other related files, powerful AI could generate an entire website and inform all search engines and AI agents.

This service would display the website to human visitors and pass all data to personal AI agents, allowing them to generate personalized UIs for end users. As people will use personal agents to handle almost everything in the future, small businesses will have various entry points rather than being overly dependent on a single platform like Amazon.

Next generation of browser engine

Another potential agent is the browser. If you’re familiar with the browser engine, you know its main job is to convert HTML code into interactive visual elements. For example, it can translate following code into a button:

In the future, AI models could be small enough to be embedded in the browser engine, unleashing many possibilities. One obvious application is the chatbot. Inspired by ChatGPT, many websites have started integrating chatbots, but why should every website build a chatbot if the browser can do the job? A website could simply add a few lines of code:

When the browser detects this code, it will quickly read all the “knowledge” from the website and prepare a chatbot, allowing users to chat with the website from the sidebar:

We can also imagine a “Designer Tool” in the browser: it could compare the current webpage with your design, ensuring the quality of the implementation. It could also help double-check the text on the webpage according to your writing style guide.

Even if we have powerful AI agents soon, people will need time to get used to them. Intermediate agents will help smooth this process. While they may eventually be replaced by personal agents (will we still need a browser in the future?), they will gradually improve people’s lives and help them become familiar with AI.

One of my fondest childhood memories is eating watermelons in summer. The sun was shining, the air was hot and still, and the cicadas had been chirping forever. The fan in the living room was working tirelessly. My parents would buy us watermelons. The good ones. As soon as the watermelon was cut open, my eyes got cooled, then I let it cool my stomach and also my mind.

But I have never been a good watermelon picker myself. Many times, I bought a watermelon where the flesh was not red but pale, and I had to throw it away. Every summer, when I pass by the watermelons in the supermarket, the same question comes to mind: How can I pick a good watermelon?

The rise of MLLMs

Back in 2023, the advent of Multimodal Large Language Models (MLLMs) changed the way we interact with AI and brought new hope to watermelon lovers like me. MLLMs accept not only text, but also audio and images as input, providing almost human-level understanding.

When experienced people pick a watermelon, they check the color and skin of the fruit. They also tap the watermelon slightly and listen to the sound because a ripe watermelon has a special density. All these tips and tricks are based on pattern recognition, which is exactly the strength of MLLMs. So, it’s entirely possible to train an MLLM to pick a perfect watermelon precisely.

Collect the data

The idea was clear: we would use an open-source MLLM as the foundation and fine-tune it to detect good watermelons. However, to achieve this, we needed high-quality data. Unfortunately, although we have tons of selfies on the Internet, the specific watermelon data we needed was not available.

We had to collect the data from scratch. We quickly developed an app and hoped to find a large community willing to help us. Deep in our hearts, We believed that we were not the only ones who wanted to eat good watermelons in the summer.

Train the WLLM

Not surprisingly, our app was widely welcomed by watermelon lovers around the globe. We received high-quality data daily from our community and began fine-tuning our Watermelon Large Language Model (WLLM). We also secured the first round of funding from a tech billionaire, who wished to remain anonymous because people think eating watermelons is not elegant. (When will the world be inclusive enough?)

First Watermelon Picking Championship

In mid-2025, confident in the performance of our WLLM, we decided to hold the first Watermelon Picking Championship in Madrid, Spain. Many experienced watermelon pickers flew to Madrid and competed against each other and the WLLM. The audience was excited to watch the championship and enjoy the free watermelons after their siesta.

After two days of competition, hundreds of watermelons were eaten, and the result was clear. It was time to release the product.

On all your devices

At the end of 2025, we launched our WLLM on all popular devices, such as smartwatches and eyeglasses. After checking the sound or the skin of the watermelon, or both, people would receive simple feedback indicating whether it’s a good watermelon or not.

By August 2026, millions of people had used our product and it was a great pleasure for us to help them enjoy good watermelons. Yes, AI doesn’t have to be cool robots; it can also be as simple as making the world a bit sweeter.

Based on some observations in the industry in Europe. Please don’t assume it refers to you.

Your grandma can book flights herself when she has time, but she doesn’t enjoy searching on the website and comparing prices. So she sometimes just calls you, a digital native and well-educated, to book the flights for her. However, whether your grandma or you are using the website, the design logic behind it is the same. While your grandma may need a larger font size, both of you are humans who share most cognitive abilities and limitations.

With AI emerging, things can change fundamentally. In the near future, your grandma will have an AI agent, at least as smart as you, help her book the flights. (You can read this article from Bill Gates to learn more about AI agent: https://www.gatesnotes.com/AI-agents). User-Centered Design will still exist, but it will occur between humans and agents since we have to ensure the agents meet human needs. However, the design logic behind a website can be completely different because agents won’t have human limitations.

We use a proper font size and text color because it’s easy for humans to read, but agents can read text in any size without color. We design a step-by-step flow to guide users because humans are prone to getting lost, but agents can process all information in one step. Wait, why do we still need a user interface for agents? Isn’t the user interface meant for humans?

We can skip all user interfaces and let agents communicate with backend APIs directly. In a future world with advanced agents, websites and apps will become obsolete. Agents will retrieve useful data and information from various providers and generate personalized user interfaces for humans. By simply saying, “Show me different flights side by side,” you’ll get a comparison view (perhaps in your preferred colors) created by your agent, not by designers from Google Flights.

Yet that future is still out of sight. Agents are still in the early stages and can only perform simple tasks. Even if agents become intelligent enough, we must wait for the whole ecosystem to catch up. New industry standards and protocols are needed, and people need time to trust agents. During the transition phase, we’ll see some workaround solutions.

Agents can find a way to “use” the user interface like humans. For example, Apple has been working on a Multimodal Large Language Model that is capable of understanding the elements on your home screen (details: https://www.zdnet.com/article/apples-new-ai-model-can-understand-your-home-screen-and-supercharge-siri/). With this kind of ability, agents can interact with all existing websites and apps before they’re agent-ready.

On the other hand, websites and apps could detect if the visitor is a human or an agent. For humans, there will be normal user interfaces as we see today, but we could embed invisible instructions and APIs for agents to communicate directly.

There are so many possibilities in the near future. Many products will fail, and many standards will be proven temporary. Yet it will be a wonderful journey, and I can’t wait to see how it unfolds.

Originally published as PDF on LinkedIn: https://www.linkedin.com/posts/zhuoqun_design-patterns-emerging-from-ai-products-activity-7217897589566828544-rTPV

Here is a backup:

Design Patterns Ai

2.26MB ∙ PDF file

Download

Download