Reimagining Software Part 1: Thoughtfulness & Personalization in the Age of AI

This will be a 3 part essay digging into how I imagine software (e.g. the application layer) evolves with LLMs. Click here

Click here for part 2 and here for part 3.

It’s only been 2+ years since we officially entered the era of Generative AI. I can barely keep up with the pace of innovation, especially with the frontier models. In the last 6 months, the progress has been mind-blowing: OpenAI announced GPT-o1 on September 12, 2024, GPT-o3 on December 20, 2024, and released a research preview of Operator last week on January 23, 2025. Last week, DeepSeek dropped R1 and made us question everything we thought we knew about model pre-training, scaling laws, and America’s advantage on the path to AGI.

I’m excited for how this keeps playing out. It’s an incredibly dynamic time in technology. But I’ve also been taking a deeper look at how the value of these models trickles back up to the application layer: the everyday software we use. And beyond the chatbots made by the frontier model companies like ChatGPT, Claude, DeepSeek. I believe we’re overdue for big bets on interfaces. So I’m writing a 3 part essay to explore how software might feel:

1. More thoughtful

2. More fluid

3. Like less work with agents

Belief #1: Software will feel thoughtful

A Brief History of Personalization

If you dumb down what building software is about, it’s still about helping people solve real-world problems. Builders often ask “how can we use tools to make our lives better?” We invent solutions, and then abstractions on these solutions to aid with adoption. For example, we created software interfaces as an abstraction to make using computers more approachable. We invented websites and apps to make exploring the connected web more seamless. Every piece of software is an abstraction.

As we invested more into interfaces, we leveraged personalization to also reduce the friction of using these tools. By personalization, I mean software that reflects your taste, preferences, and information back to you. You can trace the evolution within the software industry:

1. Early Recommender Systems (Late 1990s - Early 2000s)

   We deployed recommender systems: a way to predict or suggest what you might find relevant. Companies like Amazon launched collaborative filtering to boost e-commerce sales. Netflix offered $1 million for the best movie recommendation algorithm.

2. Web 2.0 and Social Media (Mid-2000s - Early 2010s)

   We introduced personalized feeds in social media platforms like Facebook and Twitter. We curated content for you based on your interactions, preferences, and connections.

3. AI and Machine Learning Era (Mid-2010s - Present)

   We deployed algorithms to analyze behavior in real-time and serve the most personalized content. Examples include TikTok’s feed or Spotify’s daylist.

We intuitively know that personalization makes using software feel more delightful. It’s part of the magic in technology that leads to trust and strong habits. When done well, personalization can also help you solve your problems proactively: like when Instacart starts predicting what grocery items you might need next.

When Personalization Falls Short

I used to think this level of personalized software was great. I’ve built feed products and studied networks for 10+ years. I assumed we cracked the personalization formula and it’d just keep getting consistently better. But that changed for me about a year ago when I faced a simple problem. I was reading a lot of books and I was frustrated that I couldn’t find better ways to get recommendations on what to read next. I tried everything but by far, the best suggestions came from word of mouth. I was also so frustrated with how bad GoodReads was at recommending books for me.

Around the same time, I went from being AI-curious to AI-obsessed. I began using ChatGPT for a much wider range of use cases, including as a book reading companion. I asked questions, I shared my reactions and thoughts, I expressed as much of my thought process as I could while reading a book. I started building a side project of a GoodReads that would be AI-native, like having your favorite high school English teacher in your pocket.

My core insight was that ChatGPT could give me better recommendations on what to read than most people in my life. This was because it had semblances of topical understanding (trained on a corpus of raw book text like books3 and also web content related to books) and personal context (all my chats asking questions, reacting, and sharing how I felt about a book). The combination of these two things reflected a form of emotional intelligence. And I loved the magic of feeling seen and understood by a piece of software.

I began questioning if that was the new bar for software that feels thoughtful. LLMs that can reason more like humans start convincing us that they can understand us like humans too:

Holy shit.

I used to get a great Netflix recommendation or YouTube right rail suggestion and feel delighted. Now I take them for granted and in fact, find them underwhelming. Here’s why.

Most recommender systems are still crude approximations. We like to believe each of us is unique. It is a fact that there is no one else in this world exactly like you. But the technology we built to personalize software for the uniqueness of each person contradicts this fact. Our software relies on comparing our similarities to each other and getting good at predicting patterns in data sets. We define affinities and approximate where our nearest neighbors are to mimic personalization.

Personalization has come a long way but most systems still rely on this pattern-matching rather than true understanding. This creates a gap between what we really need and what systems can provide.

How LLMs Unlock Thoughtful Software

I am excited about how LLMs give us a chance to go beyond approximations and into hyper-personalized software experiences. I see two key capabilities emerging:

1. LLMs help us remember more about users, their preferences, and how to reason with them using human-like intelligence. This will only get cheaper

2. LLMs are generative, and can tailor the UI in realtime to reflect this knowledge. This creates a feeling of hyper-personalization that has otherwise been unattainable because it’d be too expensive

The result will be be better, more thoughtful software.

This future can be hard to grok without some concrete examples. I went to SPC’s Malleable Interfaces NYC demo night last week and can share relevant explorations that hint at this future.

1. Canopy: exploring how to take brand websites and make them feel custom to you. They demo’d how every person could have a unique Nike.com experience tailored for their preferences (e.g. if you know I’m training for a marathon in the winter, show me assets and imagery that reflect that). In their own words:

   The future of interfaces is here. Canopy transforms static, maintenance-heavy sites into dynamic, AI-ready systems that think, adapt, and work as seamlessly as your team. By blending personalized experiences for customers with structured, AI-friendly endpoints, we're building the next generation of intelligent web platforms.

2. Perhaps Inc: giving you more control over the top of websites. They demo’d taking a list view webpage and turning it into a Tinder-style layout. Now I’m not sure why you’d want to do that, but the point is you can imagine visitors tailoring the UI for their own preferences.

3. Beem Computer: This was a fresh take on a new computing interface. Here’s their demo and according to their site:

   Imagine a computer that...

   Knows the perfect hotel for the group trip

   Invites the right friends to the party

   Finds the perfect flight for that work trip

   Puts on the perfect film for the mood

   Brings you the right files when you need them

   

These demos gave me a glimpse into how we’re imagining the new boundaries of thoughtful, personalized software. Each one raises the bar — and brings us a step closer to software that truly understands us and adapts to us in a way that is seamless and natural.

Redefining the Future of Software

LLMs give us the opportunity to reimagine how software understands and interacts with us. Thoughtful software isn’t just about better recommendations—it’s about creating tools that feel like extensions of ourselves, making every interaction delightful and seamless. And this pushes the arc of abstraction and making our tools easier to use. We’re getting a taste of this with chatbots, but there’s so much potential to bring this across the modern web and app ecosystem.

This is one dimension of how the application layer might evolve. In Part 2, I’ll explore how LLMs can make software feel more fluid and why that matters. And in Part 3, we’ll dive into how agentic consumption further abstract use in the application layer. Together, thoughtfulness, fluidity, and agents represent the future of software. Stay tuned!

A huge thanks to Chris Wang, Will Baine, Ammar Mian, Grace Carney, and Prasad Raje for their feedback and suggestions on this series.