News

Perspectives · Sep 2026 · 7 min

Superintelligence should make us better

If AI systems eventually exceed human ability in many domains, the most important question will not be how capable they are. It will be whether they leave people more capable, more informed and more free.

Research · Aug 2026 · 7 min

Teaching machines to speak Albanian

Albanian is a language with its own branch of the Indo-European family, two major dialect groups and very little speech data. Here is what that means for building voices and listeners that actually work.

Research · Aug 2026 · 6 min

Why we're building our own text-to-speech models

Speech APIs have never been better, so why would a small lab train its own voices? Because the things we care most about, like latency, prosody and languages the big systems serve poorly, are the ones you can't adjust from outside.

Perspectives · Jul 2026 · 6 min

Albanian deserves frontier AI

Making the best models work well in smaller languages is usually framed as charity. We think it is one of the harder, more revealing problems in the field, and Albanian is where we have chosen to work on it.

Engineering · Jul 2026 · 6 min

Latency is a feature: the engineering of real-time speech

A voice system can sound perfect and still feel broken if it answers a beat too late. Here is how we think about the milliseconds between someone finishing a sentence and hearing a reply.

Research · Jun 2026 · 6 min

How we evaluate synthetic voices

A single listening score can tell you a voice sounds pleasant and still miss that it mispronounces every surname. Here is how we try to measure the things that actually matter.

Safety · Jun 2026 · 6 min

Voice, consent, and the rules we set for ourselves

A voice is part of who someone is, and the technology to copy one is now widely available. These are the commitments we hold ourselves to as we build speech models, and the trade-offs we are still working through.

Safety · May 2026 · 6 min

Alignment is a product problem, too

Whether AI behaves well for real people is decided not only in training runs but in defaults, interfaces and feedback loops. What running a multi-model product has taught us about safety in practice.

Research · May 2026 · 6 min

Tokenizers are a language policy

Before a language model reads a single word, its tokenizer has already decided which languages are cheap and which are expensive. That decision is rarely discussed, and it has consequences.

Research · Apr 2026 · 6 min

What it takes to measure an AI in Albanian

Translating an English benchmark into Albanian gives you a number, but not necessarily a measurement. Here is why, and how we are approaching an evaluation suite built natively in the language.

Engineering · Apr 2026 · 6 min

What we learned putting 70+ models behind one interface

Building Quad Chat meant making dozens of different models feel like one coherent product. The abstractions that held, the ones that leaked, and what it all taught us about building models of our own.