Three Internships, One Question

When people look at my resume, they usually see three different internships. One focused on large language models. One was about multimodal understanding. The most recent one was recommendation systems.

At first glance, they don't seem to have much in common. For a long time, I thought so too. Only recently did I realize that every project I had chosen was quietly pulling me toward the same question:

How can AI truly understand humans?

Not generate better text. Not rank better recommendations. Not recognize more objects. But understand people.

We don't interact with AI through labels.

My first exposure to this idea came during my internship at TCL, where I worked on an LLM-powered speech understanding system for smart home devices. On paper, the task sounded straightforward: given a user's voice command, predict the correct intent and execute the corresponding action.

Traditional NLP treats this as a classification problem.

"Turn on the AC." → Intent = TurnOnAirConditioner → Execute

The system works surprisingly well — until people start speaking like people.

"It's getting a little warm."
"Could you make the room more comfortable?"
"I'm about to sleep."

None of these utterances explicitly says turn on the air conditioner. Yet humans immediately understand the underlying goal. That made me realize something simple but important: intent isn't expressed directly in language. It has to be inferred.

Large language models impressed me not because they produced fluent text, but because they began to infer the latent goals behind language rather than merely mapping sentences to labels. Language, I realized, is only one window into human understanding.

Humans communicate far beyond language.

That idea stayed with me during my next internship at TikTok. This time, the problem wasn't language at all — it was vision. I worked on improving open-world visual understanding for large-scale recommendation systems, helping models recognize concepts that had never appeared in the training data.

At first, this looked like a classic computer vision problem: improve recognition, increase precision, handle long-tail concepts. But as I spent more time with the data, I started asking a different question:

People don't upload pixels. They upload experiences.

A video isn't merely a collection of frames. It represents a memory, a mood, a hobby, an identity, or sometimes a life event. The challenge wasn't simply recognizing what appeared in a video. It was understanding what that video meant to the person who chose to share it.

Models like CLIP and later open-world vision systems represented an important shift: from recognizing predefined categories toward understanding richer semantics. But even semantic understanding felt incomplete. Knowing what someone posted still isn't the same as understanding who they are.

People are trajectories, not snapshots.

This realization became much stronger during my internship at Pinterest. Unlike content understanding, recommendation systems observe users over months or even years. Every search. Every save. Every click. Every notification opened — or ignored.

For the first time, I wasn't modeling content. I was modeling behavior.

Modern recommender systems have become remarkably good at predicting the next click. Architectures evolved from Deep Crossing and Wide & Deep to DIN, DIEN, Transformer-based sequential recommenders, and increasingly sophisticated user representations. Yet I found myself wondering:

Is predicting the next click really the same as understanding a person?

Imagine two users who both search for running shoes. One has just signed up for their first marathon. The other is recovering from an injury. Their next click might look identical. Their underlying intentions are completely different.

Behavior is evidence — not identity. Sequences reveal more than isolated actions, but they still only provide observations. The real objective isn't predicting behavior. It's inferring the latent preferences, goals, and evolving intentions that generated those behaviors in the first place.

That was the first time I became genuinely interested in user modeling rather than recommendation itself.

Toward models of humans.

Looking back, I no longer see these internships as separate experiences. Each one explored a different observable signal — language, vision, behavior — and each offered a partial view of a human. None is sufficient alone.

Increasingly, I think the future of AI will depend less on building larger models and more on building richer models of people. This is one reason I'm excited about the recent direction of generative recommendation. Traditional recommender systems largely retrieve and rank items from a fixed candidate set. Recent work — including Generative Retrieval, TIGER, and Recommender Systems with Generative Retrieval — suggests a different paradigm. Instead of asking:

Which item should I rank highest?

we may begin asking:

What internal representation of this person best explains their past — and predicts what they will need next?

That shift feels subtle. I think it's fundamental. Recommendation is gradually becoming less about ranking items and more about modeling humans.

An open question.

I don't know what the final answer looks like. Maybe it involves long-context memory. Maybe it involves world models. Maybe it requires agents that continuously learn from interaction.

What I do know is that every project I've worked on has quietly pointed in the same direction. Language helped me understand intentions. Vision helped me understand experiences. Recommendation taught me to think about people over time.

Looking back, none of those internships was really about speech, vision, or recommendation.

They were all different attempts to answer the same question: How can AI build a richer understanding of humans?