All posts

Looki Doesn't Need to Be an Agent. It Should Be an Agent's Eyes.

Auto-generated diaries are rarely worth revisiting, while proactive reminders often arrive too late. Looki's real role may be to turn life off-screen into observations you can inspect, correct, and give to any agent.

/6 min read/阅读中文版

I've worn Looki for more than four months, and I increasingly find myself wishing it did less.

One of Looki's co-founders came from self-driving cars, and the product pitch follows the same logic: perceive first, then decide. Put a device on your chest and let it watch daily life unfold; it can record, judge, and remind you on your behalf. In that framing, the physical world is the best possible prompt.

It's a compelling story. But the longer I've used it, the more I think Looki doesn't need to own the entire perceive-to-decide pipeline. Its real advantage is in the first half.

Three Products, Three Different Clocks

Early Looki felt like three products crammed into one app. It automatically edited vlogs and made comic strips, like a content tool. It generated diary entries and saved clips, like a second brain. And it presented itself as an agent that could proactively remind you of things and hold a conversation.

Those identities pull in different directions. A vlog can arrive a few minutes late; what matters is whether the final cut is good. A second brain needs to save things reliably and make them findable later. A proactive reminder, however, has to be timely. If it speaks up after the moment has passed, it may as well say nothing.

On the device, Looki samples at intervals. What it captures is only understood after it is uploaded and processed in the cloud. That is a sensible trade-off for battery life, and it works well enough for reviewing a day or organizing raw material afterward. But an architecture like that is being asked to do too much when it is expected to understand the present moment continuously and react right away.

The old app made the mismatch even more obvious. Four tabs held four disconnected feeds. Things you had already seen were hard to find again, and each conversation stopped after one response. If a diary entry got something wrong, I couldn't simply correct it. If a video missed the point, I couldn't ask for another edit. It didn't feel like talking with an agent. It felt like the product occasionally handed in an assignment and walked away.

Looki's recent redesign finally fixed the fragmented feed and made follow-up conversations possible.

A Record Is Not Memory

But that update made a deeper issue clearer. I can now search what Looki has recorded, yet I still cannot see the conclusions it has drawn from those records. The first is a record. The second is memory. If it mistakes an exception for a habit, or simply gets something wrong, I should be able to see that judgment and correct it directly. Memory is not an advanced setting for power users; it is the foundation for every future judgment the product makes.

After the redesign, I stopped worrying about making Looki feel more like an agent. Search and ongoing conversation make it a much better queryable archive. But a smoother app does not make the device any faster at understanding what is happening in front of it. If anything, it confirms my view: rather than forcing Looki to own the final judgment, it should focus on perception.

What I Need Is Observation

I don't need it to wrap my life in more diary entries and highlight reels. Most days are not that novel: I ate a meal; I sat at a computer for several hours. Frankly, even my future self probably will not want to reread it. Nor do I expect proactive reminders to be especially useful. When I have a clear need, I will open the app. By the time it notices something on its own, the moment is often already gone.

What I actually need is something more mundane: observation.

That does not mean recording continuously or producing a nightly summary about how full and meaningful my day was. It means quietly accumulating fragments. Eating dinner at nine on one evening says nothing. Doing it every night for a month may mean my routine has changed. Missing one workout proves little; seeing that my actual activity has stayed below what I imagine for several weeks is useful. One meal out is unremarkable. A month of them may show that I have gradually stopped cooking.

No individual record is interesting, but enough of them can sketch how a person actually lives, rather than how they believe they live. That also fits Looki's existing architecture: interval sampling is fine, delayed processing is fine, and the occasional missed moment is rarely fatal.

Visible, Correctable, and Portable

Not needing real-time responses does not mean Looki should be free to guess. It should retain and show what it observed, when it happened, how certain it is, which raw record supports it, and whether I corrected it later. Otherwise, bad inferences will quietly accumulate until they are difficult to untangle. The raw video can stay out of sight to keep the interface clean. Its understanding of me cannot.

Useful customization is not another set of hydration reminders. I want to be able to say: pay attention to how often I eat out, how much I exercise, and how I commute; ignore the rest for now. Those observations should also be portable, so I can hand them to whichever agent I choose instead of leaving them locked inside the Looki app.

Giving Agents a View of Life Off-Screen

I already sync my time, tasks, and financial data into my own vault. Before asking an agent for advice, I first let it see what I have actually been doing, rather than relying on how I happen to describe myself in that moment. But there has always been a missing piece: life off-screen. My data knows that I scheduled a workout, but not whether I left the house. It knows how many meetings I had, but not whether I spent another four hours at a computer after work.

Looki could fill that gap. It only needs to retain those observations and pass them to an agent better suited to making judgments from them.

I used to think Looki talked most loudly about being an agent while struggling most with the part that idea implied. Now I think that final step may not need to be Looki's job at all. As a bonus, that could save a fair amount of AI compute—and money.

The agent does not need to hang around your neck. The eyes do.