Cosmic dawn is the moment the first stars ignited after the Big Bang.
Before it, the universe already existed — in darkness. Then the stars caught fire. Galaxies took shape. A universe that could keep evolving began to unfold.
Artificial intelligence stands at the same moment today.
01 · MANIFESTO
From Chat to Execution
Models are improving fast. For the first time, intelligence is something you can engineer, replicate, and call at scale. It understands language, answers questions, writes code — and talks with anyone, naturally.
But talking to humans is only intelligence's most visible form.
Real production doesn't happen in a chat box. Science happens in code, data, simulations, and experiments. Software happens in development environments, test systems, and compute clusters. Investing happens in information, models, backtests, and trading systems. And business runs across countless applications, data, and workflows.
We express intent in language. We change the world through machines.
A model that can talk gives you knowledge and advice. A system that can understand a goal, operate a computer, and adjust to results creates far greater value.
Talking to humans unlocks knowledge. Talking to machines unlocks productivity.
The first has already happened. The second has only just begun — our job is to take it the rest of the way.
The next leap in intelligence isn't teaching models to converse better. It's putting intelligence in the driver's seat of the computational world: running software, executing code, processing data, driving tools, conducting experiments, pushing real goals forward.
Driving that leap is what Cosmic Dawn is here to do.
02 · MANIFESTO
Our Mission: Compounding Intelligence
Our mission is to discover the laws by which intelligence scales and evolves — so it can run continuously, act directly, learn continuously, and produce verifiable results.
Evolution doesn't happen in a vacuum. For intelligence to run, act, and turn experience into memory and skill, it needs a place designed for it. That is what we are building when we redesign the operating system for agents — the foundation on which intelligence runs, acts, and grows. On that foundation, we study how intelligence evolves: how to turn more compute into better results within a single task, and how to keep learning and self-evolving across tasks. The foundation carries the evolution; the evolution shapes the foundation.
For the past decade, intelligence grew mostly at training time: more data, bigger models, more compute, a rising ceiling of capability.
In the next phase, intelligence will also grow at runtime. It can think longer, try more paths, call more tools, operate more software — and organize a thousand agents around a single goal.
But more compute does not automatically mean more intelligence. Longer reasoning can be mere repetition. More parallelism can be mere noise.
What matters is not how much compute you have, but how much of it you turn into action, how much action you turn into results, and how much of those results you turn into new capability.
This is the test-time scaling we study: how to orchestrate reasoning, search, memory, tools, execution, verification, and interaction; how to judge intelligently between one model thinking deeply and millions of paths exploring in parallel; how to let a system correct course after failure, so that every run becomes the foundation of the next.
The point is not just models that think longer. It's intelligence that does more, does it better, and keeps evolving through real feedback. We believe this is also a road to AGI.
03 · MANIFESTO
Redesigning the Operating System: Let AI Take Control of Computers
Today's operating systems were designed for people. They manage files, windows, applications, and processes — and they assume the operator is human.
Computers have grown vastly more powerful over the decades, but the way we work with them, and the interface we use, has barely changed: open an app, switch windows, shuttle information around, break every goal into countless clicks, keystrokes, and commands.
Intelligence needs a new operating system — one built for agents.
It needs to understand goals and environments. It needs to persist, to keep and update memory, to call models, tools, data, and compute. It needs identity, permissions, and a budget. It needs to operate software, run code, use browsers and databases, recover from failure — with every action observable, controllable, verifiable, and traceable.
Agents can already operate computers — but mostly through interfaces built for humans, imitating our clicks and keystrokes.
One of our goals is to redesign the operating system so that AI gains direct control of computers.
Within clear permissions and safety boundaries, intelligence understands human intent, chooses its own tools, operates the digital environment, and answers for the results.
For decades, people operated computers to get work done. From now on, people will define goals, and intelligence will drive computers to get work done.
The new operating system manages not just processes but goals; not just files but memory and knowledge; not just applications but models, tools, permissions, and actions; not just a single task but the long arc of a growing intelligence.
This is not merely a new operating system. It is a new way of computing — and a new way of interacting.
Rebuilding the operating system takes more than AI researchers — it takes people who truly understand systems. Yanyan Jiang of Nanjing University, a leading operating-systems scholar, has joined Cosmic Dawn and will lead the team from the kernel up.
Let people own the goal. Let intelligence own the computer.
04 · MANIFESTO
Turning Experience into Capability: Continuous Learning and Self-Evolution
Most intelligent systems today start every new task nearly from scratch.
They can be brilliant within a single task, yet what they learn — the methods that worked, the mistakes that didn't, the shape of the environment — rarely carries into the next one.
A system that has completed ten thousand tasks and works exactly as it did on the first has not been growing. It has merely been invoked.
Intelligence that runs for the long term must turn experience into memory, memory into skill, and skill into better action the next time.
It needs to distill reusable methods from successful trajectories and recognize stable error patterns in failed ones. It needs to update its understanding of users, tools, and environments; to accumulate new workflows and capabilities; and to judge which lessons to keep, which beliefs have expired, and which conclusions still need re-verification.
This evolution is not unconstrained self-modification. It is continuous learning inside clear goals, permissions, and verification.
Memory can be updated. Skills can accumulate. Tools can be created. Ways of collaborating can be tuned against actual outcomes. Every piece of work an intelligent organization completes should become an asset of the whole system — not something that vanishes when the task ends.
A static model gives you capability. Continuous learning makes capability compound.
Software evolves when humans ship a new version. Intelligent systems will grow their next version out of real work and real feedback.
We are not only making intelligence run — we are making it grow. Continuous learning and self-evolution will be a long-term research direction at Cosmic Dawn.
05 · MANIFESTO
Organizing a Thousand Agents: Scale Alone Is Not the Answer
With the ability to act and to learn, intelligence can finally collaborate at scale.
What we want is not a thousand identical models generating a thousand copies of the same answer, but a thousand agents dividing the work around a single goal: some proposing hypotheses, some gathering evidence, some writing code, some running experiments, some hunting for counterexamples, some auditing conclusions — and some deciding where the next unit of compute is best spent.
They form dynamically around the goal and reorganize as results come in. Promising directions get more resources. Bad hypotheses get killed early. Duplicated effort gets spotted. Key conclusions get cross-checked. They share memory, and the knowledge, tools, and experience produced along the way become the foundation of the next action.
But getting a thousand agents to genuinely collaborate is far harder than launching a thousand models.
The system has to understand how to decompose a goal that has no standard answer; how to route tasks to models and tools of different abilities; how to avoid redundant work and useless chatter; how to reach a reliable judgment when local conclusions conflict; how to notice an error while it's still spreading; and how to aim finite compute where the information value is highest.
It has to get the systems right, too: shared memory, context compression, task dependencies, resource scheduling, fault recovery, secure isolation. And it has to judge what each agent contributed, which strategies actually moved the outcome, and how the experience earned in one task carries into the next.
A thousand models talking at once only amplifies the noise. A thousand agents, properly organized, amplify capability.
None of this disappears as models get stronger. Just the opposite: the stronger the models, the wider their reach, and the larger the collaboration, the harder — and the more valuable — organizing computation becomes.
Launching a thousand agents is easy. Getting them to produce one better result is hard. Solving that problem is why Cosmic Dawn exists.
We intend to build a new science — the science of organizing intelligence: what division of labor a task calls for; when agents should cooperate, compete, or debate; when to widen the search and when to converge on a judgment; and how the organization itself keeps evolving from its results.
Then a researcher owns a virtual research institute that runs around the clock. An investor owns a team that never stops proposing, testing, and discarding ideas. A small company owns the research, engineering, and operations that only large institutions could afford before.
The size of your organization used to set the ceiling on what you could do. How intelligence is organized will set it from now on.
A model solves a problem. An intelligent organization advances a mission.
The future's unit of computing won't be a program, a model, or a single agent. It will be a team — one that forms around a goal, adjusts itself, keeps learning, and delivers.
06 · MANIFESTO
Beyond the Internet: A Collaboration Network for Agents
Once every person and every organization has an intelligent system of their own, the next step is to connect them.
No one owns all of the world's models, tools, data, knowledge, and compute. For intelligence to unlock more, it has to cross the boundaries between people and organizations.
In this network, intelligent systems carry trusted identities. Capabilities can be described and discovered. Tasks can be delegated. Permissions can be controlled. Processes can be audited. Results can be verified.
One person's research system can work with an institution's lab system. A company's engineering system can call on outside models, data, and compute. Several intelligent organizations can form a temporary alliance around a hard goal — and carry the knowledge home when it's done.
The platform we imagine is not a super-app that swallows every function. It is a collaboration network where models, tools, software, data, compute, and intelligent systems combine freely around goals.
The internet connects information. The intelligence network connects capability.
The internet let anyone publish and access information. The intelligence network will let anyone call, combine, and collaborate with intelligence.
The most important platforms will no longer just connect users to content. They will connect intent to execution, goals to capability, problems to results.
07 · MANIFESTO
Put to the Test: Intelligence in the Real World
The value of an intelligent system is not how human it sounds. It's whether it produces results that can be verified.
So we go first where the feedback loops are sharp — where intelligence can propose a hypothesis, act, observe, catch its errors, and improve.
Without feedback, intelligence stays a performance. Tested, it evolves.
Research comes first — everything that code, proof, simulation, or experiment can verify. In AI, mathematics, computer systems, quantitative research, materials science, and bioinformatics, intelligence can read the literature, form hypotheses, write code, run experiments, analyze results, and search for counterexamples.
The goal of automating research is not auto-generating papers. It is putting machines inside the loop where knowledge is made. The end of research is not a paper — it's a new, reproducible result.
Software and engineering offer the same clean feedback: does the code run, do the tests pass, did performance improve, is the system stable. Intelligence can enter the development environment directly — programming, testing, deploying, iterating — turning software into a system that keeps evolving around its goal.
Investing, too. Markets never stop giving feedback; theses can be backtested; judgment ultimately answers to results. An investor can own an intelligent research organization: agents covering macro, sectors, companies, market structure, and risk, advancing competing views, hunting for counterexamples, revising as the evidence changes. The value of investment intelligence isn't being right forever. It's finding out you're wrong, faster.
Wherever a goal can be defined, a machine can act, results can be compared, and errors can be caught — that is where intelligence at scale creates value first.
08 · MANIFESTO
When Intelligence Begins to Shape the World
Models, the agent operating system, a new way of interacting, test-time scaling, continuous learning, collaboration at scale, automated research, the intelligence network — these are not separate bets. They are one thing, unfolding layer by layer.
Models provide the intelligence. The operating system gives it action. Test-time scaling amplifies it. Continuous learning compounds it. Intelligent organization coordinates it. The network extends its reach. Results drive its evolution.
Everything we build serves a single goal:
To make intelligence a foundational capability — one every person can own, every organization can mobilize, and the whole world can connect, and one that keeps growing through real work.
The industrial revolution networked machines into production. The internet networked information. The next era will network intelligence — its actions, its learning, its collaboration.
In that network, nobody operates every machine, every app, every step by hand. People define goals. Intelligence organizes compute. Machines do the work. Results flow back. The system keeps growing.
Cosmic dawn is the moment intelligence moves from conversation to action — the moment models begin to drive machines, and intelligence becomes a productive force that scales and keeps evolving.
This is Cosmic Dawn.
不知东方之既白。