Article · AI translation
At last, a starting point you can touch
Jay reporting from Aofei Temple for QbitAI | WeChat account QbitAI
Jay reporting from Aofei Temple for QbitAI | WeChat account QbitAI
The Siri Jobs envisioned may finally be becoming reality.
It looks like this—
Before anything else, an AI product with good design has finally arrived. This UI is genuinely pleasing to look at…
But the big thing is still to come.
This voice assistant is almost unreal.
It really is like having a personal secretary: you can have coffee and bread while listening to it brief you on news worth following today.
The key is how human it feels. Even if you interrupt it midway and give it a task, such as “help me reply to an email,” it does not lapse into awkward silence.
Video source
It is practically Samantha from the film Her.
It is wild. Five years ago, you could only see scenes like this in a movie theater.
The AI world has been battling it out for ages. Every week, the major companies unfailingly have another meltdown, while the frontier keeps producing flashy feats—controlling Blender, bringing robotics into a new dimension, solving a Millennium Prize problem…
But frankly, I have never found any of that particularly exciting.
It is too far removed from ordinary people. For most people, these frontier breakthroughs will not change their real lives in the slightest anytime soon.
But after using this product on and off for nearly a week, all the romantic visions I once had of the AI era suddenly took concrete form.
I even think—
This may be the AI assistant that, so far, comes closest to science fiction films and fulfills all my romantic visions.
Its name is Orilo.
The awkward Personal Agent
The concept of the Personal Agent really has been hyped to the skies lately.
Muse’s growth rate even overtook GPT’s at one point, and Agents have moved decisively from productivity tools for geeks into the mass consumer market.
But honestly, most of these experiences are not much better than the “little lobster” from the start of the year…
I do not know about everyone else, but I only felt a little novelty during the first few days after getting an invitation code. I even went so far as to give it notification permission.
And then, nothing.
Typing is just too painful. Even less human-friendly is its proactive notification system: my pocket suddenly vibrates, and I think it is an urgent WeChat alert. After struggling to pull out my phone on the subway, I find Muse reminding me about a to-do list that should have expired ages ago…
At that point I was… at a loss for words.
So I have always felt that the solutions in the Personal Agent market are still very immature.
The vast majority of Personal Agents are stuck in the same fatal dead end. I call it “one low, two highs”:
Extremely low frequency of use.
Whether it is booking flights or canceling streaming subscriptions, it certainly sounds cool. But how often do you encounter a complex task like that in a month?
For a trivial task that comes up so rarely, I still have to spend ages explaining it to the agent. Why would I not just tap through it myself?
Extremely high interaction friction.
When you are driving, washing dishes, cooking, or running, your hands and eyes are fully occupied.
Is it realistic for your pocket to buzz at that moment and expect you to stop, take out your phone, and type through the logic with an AI?
Extremely high operating costs.
Every time a large model is called in the background, these scenarios require very complex tool calls, and each loop burns real money.
With that cost structure, scaling a consumer product to mass adoption is practically a fantasy.
Technological purism has ruled this industry for too long. But a personal companion product cannot win on model size and benchmark scores alone.
The next generation of Agents that survive must fit quietly and frequently into everyday life.
Voice is the real gateway
That is also what I find most compelling about Orilo.
It has found a new approach that feels very intuitive—
Voice-first, with voice taking priority.
Why do I say voice is so very important for consumer products?
It is the most natural way for people to interact.
Writing was invented long after language emerged, mainly to pass on knowledge. It is not actually essential for everyday communication.
Think about how often it is inconvenient to read text in daily life: driving, cooking, looking after children…
More importantly, when people speak, they unconsciously reveal more personal context (Conversational Context).
You would not just say, “Book me a ticket.” You might casually add, “Book me a ticket to Shanghai for next Wednesday. Somewhere near that hotel from last time is fine. Oh, and the budget is tighter this time.”
None of those casual details can be found in any connector, yet they are key to understanding a person.
But that is still not enough.
There are too many markets where voice could work. A new AI product has to find a killer application.
That is the second clever thing about Orilo, in my view: the team found an ingenious entry point—
The morning.
It is the part of the day when people process information most intensely.
You have just woken up, your head is foggy, and you urgently need to know what has happened in the world and how to arrange everything you have to do today.
Yet that is exactly when you have a toothbrush or coffee cup in your hand.
This is the pain point Orilo primarily targets.
It frees your hands completely: set the device aside and receive information simply by listening.
During that time, if there is anything you need done, just tell the AI by voice.
I could cry, folks. Finally, a team that is not sitting cross-legged in ergonomic chairs imagining users’ pain points.
And as for the UI, I think it really blows the others away.
At last, no more cookie-cutter chat boxes and screens full of toolbars…
It takes the form of a water droplet. It is also connected to the phone’s physical gyroscope: tilt the phone, and it slowly slides across the screen like a real drop of water.
(P.S. This is what a consumer product should look like. Take note, professors, and ease up on the executive vibe!!)
As for specific real-world uses, here are the 3 things that appeal to me most—
A morning briefing just for you
Put simply, it is a news digest. Muse has this feature too.
But honestly, I personally do not use it much. It basically pushes stories based on interest tags; at heart, it is just a news feed with a new skin.
Orilo uses a completely different logic.
It looks at what happened and also works out what it means for you.
For example, you might follow Anthropic closely, but from past conversations and long-term memory, it knows that you care more about how the large-model ecosystem’s value chain is being reshaped than about an IPO.
So when it briefs you on the latest funding developments early in the morning, it also offers an interpretation tailored to your perspective at work.
It is like having a top industry analyst standing by your bed, drinking coffee and chatting with you one-on-one.
Keeping a close watch on niche information for you
There are too many things we want to keep up with every day: an obscure startup, a Champions League group-stage match, or a podcast that has not released an episode in ages.
But our attention is extremely limited. Who has time to keep manually checking search results every day?
Now all you have to do is say the word.
It will monitor information sources for you around the clock, 7×24, and proactively report back at the right time as soon as something changes.
Completing a chain of actions with one sentence
Of course, even more possibilities lie in its ability to gather context across different app domains.
Like other Personal Agents, Orilo has an entry point for connectors and includes some standard plugins.
Take email, for example.
While listening to the morning news briefing, you suddenly remember that you have not checked your work email today.
Check my email for me.
Orilo naturally moves into a tool call and can extract the important information in 2 or 3 seconds:
Colleague A says you need to update the PPT for next Tuesday’s presentation.
Colleague A says you need to update the PPT for next Tuesday’s presentation.
Of course, it can also proactively keep track of things.
Say your child’s school suddenly sends an email in the morning saying dismissal has been moved up to 3 p.m. After Orilo catches the email, it casually interjects during your spoken morning briefing:
Your child’s school is letting students out early today. Would you like me to add it to your calendar?
While squeezing out toothpaste, you mumble back:
Add it, and remind me half an hour beforehand.
Add it, and remind me half an hour beforehand.
Done.
From catching important information to carrying out a task across apps, the whole process is completed seamlessly in 2 sentences.
A comparison with current competitors may make this more intuitive.
Ask Muse to look up an email, and the interface often gives you a thumbs-up, then vanishes, only replying more than 20 seconds later.
My reaction was just: ???
My blood pressure shot through the roof.
Of course, since this is a first-generation product, I do have one complaint I have to make.
Why does it only support English? 😭!!!
For the first time, I felt how important learning English is, folks. Only when I actually needed to have a conversation entirely in English did I realize:
I am just an ordinary person who memorized several thousand English words for the postgraduate entrance exam, o (╥﹏╥) o.
I am just an ordinary person who memorized several thousand English words for the postgraduate entrance exam, o (╥﹏╥) o.
Being unable to choose your native language really hurts the efficiency of communication.
Of course, it could also be because Chinese-language voice models are relatively less mature.
How do you get AI to speak up and get things done?
That covers the experience. Now let us talk a little about what is under the hood.
Why can Orilo hold a real-time conversation without ever getting stuck?
In fact, teams building voice Agents today are all being driven mad by the same fundamental physical bottleneck: deep reasoning and ultra-low latency are inherently at odds.
The industry mainly has 2 approaches:
End-to-end native audio: audio goes straight in and straight out. It is extremely fast, with remarkably rich, natural emotion. But its fatal flaw is poor logical reasoning; tasks involving even slightly complex logic bring it to a halt.
Cascaded architecture (ASR-LLM-TTS): first transcribe speech, then let a large model think, and finally synthesize speech to read the answer aloud. It is smart enough, but latency often starts at 1.5 to 2 seconds. That is not a conversation; it is talking over a walkie-talkie.
There is a brutal psychological threshold in human-computer interaction: once audio latency passes 1.5 seconds, if the other party has not responded, the feeling in your head that you are “communicating with an intelligent being” vanishes instantly.
Orilo’s way around this is to separate the “interaction layer” and the “execution layer” completely in its architecture, building a 2-layer network:
Front-end layer: a lightweight, ultra-low-latency voice model handles rapid listening and detects emotion, ensuring an immediate response to any interruption or conversational turn.
Back-end layer: a high-reasoning Agent that thinks more slowly reads emails, calls interfaces, and checks calendars behind the scenes.
The industry is actually pursuing this 2-layer approach already; OpenAI’s GPT-Live is moving in a similar direction. But everyone has stumbled over the same problem:
During the dozen or so seconds the back end is working hard on a task, what is the front end supposed to do—stare at the user?
I once used GPT to help reply to an email. It cheerfully said “Okay,” and then my phone fell completely silent for a full 2 minutes.
I wondered: Come on, are you working or have you crashed?
That long black hole of “the conversation has ended, but the action has produced no result” makes for a terrible experience.
Orilo cleverly solves this with a product-level trick—
It gives the front-end voice model the ability to keep the conversation going while work is underway.
When you ask it to create a calendar event in the background, it directs the back end to carry out the task while continuing to talk naturally about the news you were just discussing.
There is not even a second of silence in the entire interaction. It is like an experienced administrative assistant typing at a desk to book you a ticket while chatting with you about industry gossip, without either activity getting in the way.
While testing it, I also noticed an interesting detail: when it needs a little time to think, it fills the gaps very naturally with humanlike vocalizations such as “hmm…” and “uh…”
After hearing them often, the artificial feel does become fairly obvious, but it is still much better than getting the AI’s “silent treatment.”
You do not have to take my word for it; the data supports this too.
Orilo performs remarkably well in an industry evaluation framework called Full-Duplex-Bench.
Its strongest point is how quickly it starts speaking again after a user interrupts it.
It takes just 1.08 seconds, the fastest in the test. By comparison, GPT Realtime takes 1.50 seconds, while Gemini Live is much slower at 2.62 seconds.
Of course, after trying it myself, I found that the improved interaction comes at a cost.
This two-layer setup, built to pursue responsiveness to the extreme, sacrifices robustness.
Several times, the backend failed to call the API at all, yet the frontend, to avoid an awkward silence, told me with a straight face that “everything was taken care of.”
Then I opened my calendar and found it completely empty 😑.
But looking at the bigger picture, this technical approach is probably the right one; it is just a matter of time.
Baichuan co-founders start a business
Finally, as usual, let’s talk about the team behind it. They are quite interesting, too.
The founder, Jiao Ke, is a co-founder of Baichuan Intelligence.
I don’t know if any of you have heard of it, but last year they released a product called “Laifu Radio.”
At its core, it was an AI radio station offering personalized daily briefs, very much in the same vein as Orilo.
It was doing quite well at the time: with a user base in the hundreds of thousands, it achieved 50% next-day retention and over 10% 30-day retention.
Those are remarkable figures.
So you can see that they have been pursuing the voice use case all along. The clues were there from the start.
Back to Orilo itself—how did the idea come about?
Their experience with “Laifu” helped them validate a product-market fit: the “morning” use case.
But they soon began thinking about the next step: beyond providing information, what else could AI and voice do?
The answer was Orilo.
As Jiao Ke himself said in an interview, they had long wanted to create a “person” who could remember things and reply to emails, just as depicted in Her.
Orilo’s arrival, then, can be called a long-planned shift in paradigm.
We also learned that the team launched the Orilo project early this year.
More than half a year later, all the pieces of the puzzle have finally come together.
One More Thing
Finally, let me add one more thing.
Honestly, judged purely by the experience of everyday life, the previous generation of internet products did a pretty poor job.
They raced to lure users onto their platforms, then, once users were locked in, showed their true colors and bombarded them with junk ads.
But now, the yardstick for the value of a new generation of AI products seems to be changing.
Whether it is Claude Code, Codex, or Muse, these new forms of agents emerging one after another are doing the opposite—
helping users reclaim the time that internet products once took from them.
They let you give your attention back to the people around you.
That is exactly why Orilo chose a voice-first approach.
It has never sought to become an all-powerful gateway to every tool; that is not the mission of an AI product, so it never wants users to keep their eyes fixed on a screen.
If you need something, just say it. Once it is done, it goes straight to sleep.
But that restraint may be Orilo’s most precious quality.
When AI stops trying to intrude on our lives, it may instead become the companion you are happy to chat with for a few moments each morning by your bed, in your car, or in your kitchen.
From “using” an app to “spending time with” an agent.
An entirely new relationship between humans and machines is beginning to take shape.
In 2010, Jobs acquired Siri, a startup incubated by the Stanford Research Institute.
According to Siri CEO Dag Kittlaus, Jobs himself called to arrange a meeting for the next day, then spoke with the three founders for three hours at his home in Palo Alto.
Jobs said he wanted to build Siri directly into the iPhone.
I want to completely transform this phone... You won’t have to keep tapping and typing anymore.
I want to completely transform this phone... You won’t have to keep tapping and typing anymore.
But as everyone knows, LLM technology arrived more than a decade late.
Even more regrettably, Siri has still failed to catch the wave of the new AI era.
And we will never know what Siri might look like today if Jobs were still here.
But now, Orilo has been released—
and the many hopes Jobs placed in Siri more than a decade ago...
finally have a tangible starting point.