What we believe about AI

At AR Labs we think conversation will become a main way people work with computers, that AI should augment people rather than replace them, and that a person and an AI working together can do what neither does alone. We also think much of the risk of AI lies in how it is used: chasing speed, handing over judgement, and letting AI agents (language models run in a loop with tools, so they can act, not only reply) act without limits anyone wrote down. This page sets out each position, our reason for it, and where it shows up in our work.

Conversation as the next operating system

We think devices and applications will be controlled more and more through voice and dialogue. We want software that communicates naturally and understands what a person means, not just what they say, and we see a path from text, to voice, to interaction that mixes voice with pictures. Speech alone is a poor way to return complex results, so we pair it with visual structure; the VOX page explains why.

Where it shows up: Salesbot, a live browser voice agent, and the voice interface designs behind VOX OS, our design for an operating layer over agent work.

Augment people, do not replace them

This is one of our four core beliefs. The reason comes from our founding framework, cognitive ecology, which describes a mind as a small conscious attention that frames the work of a large, fast, pattern-completing unconscious, and judges actions by their wider consequences. We read a person working with a language model the same way: the person supplies intent, framing and judgement, and the model supplies fast pattern completion.

The essay series on our blog takes the same line in public. It calls advertising the replacement of half a team with AI something to be ashamed of, and asks who will become a senior developer if AI replaces the junior ones.

Where it shows up: our coding pipeline keeps a person at the design, the plan and the merge, and our code tutor has the person write the code while the agent teaches.

Intelligence should be accessible to everyone

We think AI capability should not belong only to corporations. We want to make intelligence accessible to individuals and small teams, putting power in people's hands rather than in corporate entities, and we test that by building tools a single person can run.

Where it shows up: we run our own work as a one-person company operated with AI agents; we teach the second brain workshop and publish guides for small businesses.

What AI is good for

We see three sound uses:

• Thinking alongside a person. Agents that hold memory and context, plan, and show their working, so that a person's productivity and creativity grow.
• Grounded answers. Language models produce confident errors; connecting them to real knowledge sources, so that answers can cite where they came from, is our preferred remedy.
• Acting, not only chatting. AI agents as co-workers that use tools and carry out tasks, provided they work inside a harness: everything around the model that shapes what an agent does, from the agent program and its instructions to its skills, hooks, tools and checks.

Where the risks are

Speed without direction. Our essay series on the acceleration paradox argues that the feeling of going faster with AI is not a reliable measure, that faster delivery carries hidden costs in quality and learning, and that some work, learning and judgement above all, should not be fast. It argues for working slower and more purposefully while still using AI for what it does well, and says plainly that this middle path has not yet been found. It ends on a warning: used without awareness, and allowed to do everything for us, AI leads toward harm; used consciously, it might help solve problems that would otherwise threaten humanity.

Attention as a resource. We treat human attention as something AI can damage. The same essay asks people to limit what AI knows and does in their lives, as they limit endless scrolling. Our interfaces present long answers a piece at a time so they do not overload working memory.

Ungoverned agents. Running our own work on agents has shown us that agents fail in specific, repeatable ways: reasoning from nearby context instead of the authoritative source, deciding above their delegated authority, or producing a well-argued wrong answer whose coherence hides a missing check. A rule that only asks an agent to behave is not a control. Each recurring failure should become a check in the environment that blocks the action and records that it fired. The harness paradox states the related belief that tightly constrained agents do more useful work than loosely configured ones.

Technology's wider effects. One of our long-term aims is reducing the damaging effects of technology on society. We take from Gregory Bateson the warning that technology amplifies narrow conscious purpose without widening its view, and design with that warning in mind.

The role of the human

We think the person keeps hold of four things: the intent behind the work, the framing that tells an agent what matters, the judgement about whether results are acceptable, and the check that consequences fit the wider system. Our systems put that into practice. Agents follow an explicit rule about when to act or ask; trust is extended one agent at a time, and a stage must succeed before the next begins; and a person, not an agent, merges agent-written code into the main branch.

flowchart LR
  P["Person: intent, framing, judgement"] --> A["Agents: tools, memory, execution"]
  A --> G{"Governance: authority, evidence, limits"}
  G -->|"within bounds"| R["Result shown to the person"]
  G -->|"beyond bounds"| P
  R --> P

Where we put our effort

We expect much general agent tooling to become commodity, and put our effort into what we expect to stay hard: governance, meaning rules that say what an agent may decide, gates that stop it when it goes further, and evidence that a control fired; a person's view and control of agent work; and interfaces that fit how people think. The second of these is the layer VOX OS is designed around: the view through which a person orders and steers many agents. We write our governance rules in terms of authority, evidence and bounded action rather than any one tool, so that they can outlast the tools they were first written for.

Where it shows up: harness engineering, human-in-the-loop gates and evidence of completion.

Further reading

The vision states what these beliefs aim at, and the future of work with AI agents applies them to how people and agents may work together.