Introducing Quad Labs Research

For most of our history, the research we do at Quad Labs has lived in the margins of our product work: notes in shared documents, experiments that informed a decision and were then set aside, evaluations run once to answer a question and never written up. Today we are giving that work a proper home. The new Research pages on this site describe the areas we study, the questions we are trying to answer, and, over time, the writing and methods we produce along the way. This post explains why we are doing this, what we are working on, and what we hope it becomes.

Why a product company does research

We build products that people use every day. Quad Chat gives thousands of people daily access to more than seventy AI models through a single subscription, alongside image generation, web search and an app builder. Imazho and Annot8 are in development. It would be reasonable to ask why a company in that position should spend any of its limited time on research rather than on shipping the next feature.

Our answer is that the two are harder to separate than they look. Running a product that sits in front of many different models, used by many different people for many different tasks, forces a steady stream of questions that nobody has fully answered. Which model is actually better for this kind of request, and how would we know? Why does a system that performs well in one language stumble in another? When a model is confidently wrong, what should the interface do? These are not questions a roadmap can settle. They need careful measurement, patience, and a willingness to be surprised.

There is also a simpler reason. We think the next decade of computing will be shaped by machine learning systems, and we would rather understand those systems deeply than treat them as black boxes rented from someone else. Understanding comes from doing the work: training models, studying their failures, building evaluations, and writing down what we find. A lab that only integrates will always be a step behind the questions that matter.

What we are working on

Our program is organized around six areas. They are not independent silos; most of our projects touch more than one. But naming them helps us say clearly what we care about.

Speech and text-to-speech. We are exploring how to build speech models that sound natural, carry meaning through prosody rather than just pronunciation, and work well in languages that have received little attention from large systems. Speech is one of the most human ways to interact with a machine, and it is also one of the places where quality gaps between languages are most obvious.

Albanian and low-resource languages. Albanian is personal to us, and one of our offices is in Prishtina. We see every day how much weaker most AI systems are in Albanian than in English. We are working on data, evaluation and modeling approaches that could help close that gap, and we hope what we learn transfers to other languages in a similar position.

Language models and reasoning. We study how language models reason, where that reasoning breaks down, and how to measure it in ways that reflect real tasks rather than narrow benchmarks. Much of this work is informed by what we observe when different models attempt the same problem.

Safety and alignment. We are interested in how models behave for real people in real products: whether they are honest about uncertainty, whether they flatter rather than inform, how they handle requests that sit near the edge of what is appropriate. We think a great deal of practical alignment work happens at the level of product decisions, and we want to study it with the same rigor as any other research question.

Vision and multimodal. Images, documents and interfaces are increasingly part of how people use AI. We are exploring how multimodal models understand visual content, which is directly relevant to products like Imazho, and where their understanding is shallower than it appears.

Efficient ML. Not every useful model needs to be enormous. We are working on techniques that make models cheaper to run, faster to respond and easier to deploy, because efficiency determines who gets access to these systems and on what terms.

How research and products feed each other

The relationship we want between research and product runs in both directions. Products generate questions. When we notice that people routinely switch models for a certain kind of task, or that a feature behaves differently across languages, that observation becomes the seed of an investigation. We try to treat these moments as data rather than anecdotes, and to follow them until we understand what is actually happening.

Research, in turn, changes what we build. An evaluation method developed to answer a narrow question can become the way we decide which models to offer by default. A finding about how models express uncertainty can change how an interface presents an answer. A small, efficient model trained for a specific purpose can make a feature possible that would otherwise have been too slow or too expensive.

We would rather understand these systems deeply than treat them as black boxes rented from someone else.

We want to be careful about one thing. The needs of a product can pull research toward whatever is convenient to measure or quick to ship. We are trying to protect time for questions that do not have an obvious payoff, because in our experience those are often the ones that end up mattering most. We will not always get that balance right, but we intend to keep it in view.

How we will share our work

We plan to share what we learn in a few forms. The first is writing: essays and technical notes that explain what we tried, what worked, what did not, and what we still do not understand. We think honest accounts of partial progress are more useful to the field than polished announcements, and we will try to write that way.

The second is evaluation methods. Some of the most valuable things a small lab can contribute are careful ways of measuring what models can and cannot do, particularly in areas like low-resource languages where good evaluations are scarce. Where it is responsible and practical, we hope to publish the methods and materials we develop so that others can use, critique and improve them.

The third, over a longer horizon, is models and tools. We are not announcing any today. When we do release something, we will describe it plainly, including its limitations, and we will not claim more than the evidence supports.

We should be clear about where we are. This is an early program at a small lab. Much of the work described on the Research pages is ongoing, and some of it will not pan out. We would rather say that now than overstate what we have.

Who we want to hear from

Research gets better when more people look at it. We would like to hear from researchers working on any of the areas above, especially speech, low-resource languages and practical evaluation. We would like to hear from linguists, translators and native speakers who care about how AI handles their languages, because their judgment is something no automatic metric can replace. We would like to hear from engineers who have run into the same hard problems we describe and found different answers.

We are also interested in people who disagree with us. If you think a question we are pursuing is the wrong one, or that our framing misses something important, that is exactly the kind of feedback we want. A lab that only hears agreement stops learning.

Our offices are in Wyoming, Prishtina and London, and we work with people wherever they are. If any of this resonates, we would be glad to talk. The Research pages are where we will keep track of what we are doing. We hope you will follow along, and we hope some of you will help.