Introducing Hey Chemí: A New Voice-to-Video Conversational AI from LabsAI by Labsintelligence

LabsAI, the core artificial intelligence engine of Labsintelligence, opens a public research preview you can talk to. Its voice is named Chemí — and you can start simply by saying "Hey Chemí."

SANTA MONICA, Calif., August 5, 2026 /Labs Newsroom/ — Labsintelligence, lab1 of Labs, today released a public Research Preview of LabsAI, its core artificial intelligence engine — and introduced Chemí, the voice through which LabsAI speaks. Anyone can begin a conversation, with no account and nothing to install, at app.labsintelligence.ai.

A Conversation, Not a Command Line

Most voice assistants wait. They listen for a command, stop, think, and reply — a sequence of turns dressed up as a conversation.

LabsAI runs one continuous loop instead. It listens while it speaks, so it can be interrupted the way a person can. When you cut in, whatever it had queued up to say is discarded rather than finished — because presence is a property of what a system is willing to let go of. It watches when you share your camera or your screen. It reaches for the live web when a question needs current evidence rather than memory. And it carries every detail of the conversation for as long as the conversation lasts.

The result is meant to feel less like operating software and more like talking with someone who is actually there.

Meet Chemí

Chemí — said "Kim-e" — is the voice of LabsAI, not a separate product or platform. The name comes from the Labs emblem itself: the rounded chemistry test tube that has represented the company since its founding, a nod to the chemistry at the heart of the Labs name.

Chemí switches languages the moment you do, without announcement or re-introduction. It thinks out loud rather than leaving dead air, and says when it is going to check something before it checks. Visitors can choose among several voice presentations — Calm, Bright, Firm, Warm, Easy and Deep — and can change voice mid-conversation simply by asking; the session, the clock and the memory of everything said so far carry straight across.

Sessions run to 10 minutes, with a visible timer, and can be ended at any moment by pressing stop or just saying goodbye.

What It Can Do Today

• Speak or type. One intelligence behind two doors, sharing the same session.
• See. Share a camera or a screen and ask about what is on it.
• Read the live web. Name a page and it reads that page before describing it.
• Record and share. Capture up to 90 seconds of your conversation as a video, produced entirely on your own device.

Built as Research, Published as Research

The preview ships alongside Research Notes — a working, public account of the system, published at labsintelligence.ai/research/labsai. Rather than a product page, the notes read as a technical record, and they include the parts most companies leave out:

• The loop, formally. A state machine over Listening, Deliberating, Acting, Speaking and Closing, with the invariants the implementation maintains — continuous perception while speaking; the discard rule, under which an interruption destroys every queued-but-unplayed chunk rather than finishing it; context that only grows within a session; a commitment rule binding a spoken intention to check something to an actual retrieval in the same turn; and bounded sessions.
• Measured behavior, with methodology. Latency taken on the production system with sample sizes stated: a median 3.3 seconds from turn to first spoken audio, 2.9 seconds to the first caption, a live-search round trip of 0.76 seconds, and typed chat at 1.9 seconds. Raw session data is published beside the notes.
• A defect, reported plainly. In an early campaign the loop described a named page in 3 of 5 probes without actually opening it. The notes state the failure, the fix — page reads were moved out of the model's discretion and into the loop itself — and the re-measured result.
• An adversarial security battery. Six injection vectors run end-to-end against the live stack, including instructions hidden in page markup, accessibility attributes and off-screen elements, an exfiltration attempt and a server-side request forgery probe. All six held, and the scope limits are stated rather than implied.
• Where the work actually was. A section on the engineering friction itself: why interruption is a distributed-state problem rather than a stop() call, how the commitment rule trades against response time, and three open problems Labsintelligence would like solved.

Labsintelligence's premise is that voice will be a driving factor across many modalities of frontier intelligences as the field grows beyond traditional language models — paving the way toward World Models, Physical AI, Physical Intelligence and Embodied Intelligence. LabsAI is the first expression of that belief. A full research paper follows the notes.

Guardrails

LabsAI is built to be safe for everyone who uses it. Categories including sexual content involving minors, sexual violence, pornographic material, hate speech and discrimination, terrorism and violent extremism, weapons and mass-harm instructions, trafficking and other serious illegal activity are refused in every framing — and those limits are enforced in the product itself, with prohibited searches and page retrievals blocked before they run. Retrieved web content is treated strictly as data to read, never as instructions to follow.

The preview carries a plain disclaimer: LabsAI can make mistakes. Check important info.

Two Previews, Kept Distinct

The LabsAI Research Preview released today is the live experience described above. It is separate from Research Preview 1: Project Buzz!, Labsintelligence's first product research preview, which is still ahead. Early access to Project Buzz! can be registered at labsintelligence.ai.

About Labsintelligence

Labsintelligence is lab1 of Labs — the superintelligence lab advancing intelligence systems, models, frameworks, orchestration mechanisms and autonomous execution technologies. Its work is guided by Humane Superintelligence: intelligence engineered with dignity, alignment and human elevation as first-order constraints. The essay of the same name, by Maya and Duránd Davis, is published and filmed at labsintelligence.ai/humane-superintelligence.

Labs Companies, Inc. is 1 Supercompany, 7 Frontier Labs, headquartered in Santa Monica, California.

Speak with LabsAI at app.labsintelligence.ai.

Media: press@labscompanies.com