In Brief
Yann LeCun’s world-model research focuses on how intelligent systems can understand physical dynamics, predict possible future states and plan actions. Fei-Fei Li and World Labs are building spatially intelligent models that can perceive, generate and interact with coherent 3D worlds.
ONONO focuses on a different dimension of intelligence: how reality develops over time.
Our world model forms part of a broader Mind Model. It reconstructs events, actors, concepts, evidence, decisions and possible causal relationships from fragmented information. It preserves conflicting interpretations instead of prematurely collapsing them into a single answer. And because it understands how a situation evolved, it can place new and future events into a richer conceptual context.
A recent article on how technology journalist Taylor Lorenz conducts research with AI shows why this matters — and how journalistic work can be meaningfully augmented by ONONO Time Machine.
A journalist, five AI models and a fragmented research process
We recently came across an article that made us pause.
In “ Model Behavior: Taylor Lorenz ” the technology journalist describes how she uses AI throughout her research and publishing workflow.
When Lorenz needs to understand a subject she does not yet know well, she starts with material from people she trusts. She asks AI to collect what they have said about the topic. She searches with Gemini, runs additional queries through Grok, gives the accumulated material to Claude for organization and repeats similar searches across several models to compare the results.
This is not casual experimentation. It is a pragmatic response to the limitations of current AI systems.
No single model gives her sufficient confidence. Each model produces a different selection, structure or interpretation. Lorenz therefore becomes the human integration layer between several AI tools, a collection of source material and her own editorial judgment.
Her insistence on linking back to original sources is especially important. Readers must be able to inspect where a statement came from and evaluate it in context.
This workflow points to a deeper challenge.
Journalists and researchers do not simply need more information. They need to understand how pieces of information relate across time.
Search finds information. Research reconstructs a history.
A journalist investigating a company, political decision, technological controversy or public figure may be working with hundreds of fragments:
- articles and archived webpages
- interviews and transcripts
- public statements and social media posts
- corporate filings and official documents
- emails, notes and internal records
- corrections and changing versions of a story
- competing accounts from people with different interests
- facts that only became known months or years later
A conventional AI search tool can locate some of this material. A large language model can summarize selected documents. Neither function, by itself, creates a durable understanding of the case.
The difficult questions are relational and temporal:
- What happened, and in what order?
- What was known at a particular point in time?
- Who made which claim?
- Did a source later change its position?
- Which sources are independent, and which repeat the same original account?
- What evidence supports or contradicts a claim?
- Which concepts connect events that initially appear unrelated?
- Which conclusions are documented, which are interpretations and which remain hypotheses?
- How does a new event change our understanding of everything that came before?
These questions require more than retrieval. They require reconstruction.
That is what ONONO Time Machine is designed to do.
What does “world model” mean?
A world model is an internal representation that allows an intelligent system to understand a situation, anticipate how it may change and evaluate possible actions.
But there is more than one way to model the world.
Yann LeCun: predicting dynamics in the physical world
Yann LeCun’s work on world models asks how machines can learn from observation, understand physical dynamics, anticipate possible outcomes and plan actions.
In LeCun’s proposed architecture, a world model estimates information that perception cannot directly provide and predicts possible future world states resulting from natural developments or an agent’s actions. His Joint Embedding Predictive Architecture, or JEPA, learns abstract representations instead of attempting to reproduce every detail of an image or video.
Meta describes V-JEPA 2 as a world model for understanding, predicting and planning in the physical world. This work explicitly includes time: the system observes motion and anticipates how a physical situation may develop.
The central question is approximately:
Given the current physical state and a possible action, what is likely to happen next?
Fei-Fei Li: understanding and creating spatial worlds
Fei-Fei Li and World Labs approach world models through spatial intelligence.
Their models are designed to perceive, generate, reason about and interact with coherent three-dimensional environments. This work is concerned with geometry, spatial relationships, physical consistency and the ability to act within virtual or physical spaces.
The central question is:
What is this three-dimensional world, how is it spatially organized and how can an intelligent system interact with it?
ONONO: reconstructing how reality became what it is
ONONO concentrates on another kind of world: the world represented through language, documents, decisions, relationships, memories and competing human interpretations.
Our central question is:
How did the current state of reality emerge, what conceptual and causal relationships shaped it, and how should new events change what we understand?
This is not a rejection of physical or spatial world models. It is a different focus.
Physical time describes how objects and agents move from one state to another. Historical and semantic time describes how situations, institutions, ideas, decisions and human relationships evolve.
For journalism, law, management, finance and other high-context domains, this second form of time is fundamental.
ONONO pushes this frontier further by building a model not only of external reality, but also of the AI’s own evolving understanding. This broader Mind Model preserves memories, experiences, interpretations and reflections as the system develops. How ONONO creates this cognitive layer is explained in Why AI Needs a Mind Ontology .
Time is not metadata
Most information systems treat time as a field attached to an object: a publication date, a timestamp or a deadline.
ONONO treats time as part of the structure of meaning.
A statement made in January may mean something different after a disclosure in June. A decision that appears irrational today may have been reasonable given the information available at the time. Two articles may contradict each other because they describe different phases of an evolving situation. A source may not be unreliable in general but may have lacked access to a decisive fact at a particular moment.
Without time, these distinctions disappear.
The system may retrieve the correct sentences and still produce the wrong understanding.
ONONO reconstructs how entities and concepts change, how events alter the state of a situation and how new evidence revises earlier interpretations. It also preserves the history of those revisions rather than silently replacing an old answer with a new one.
That is why we describe ONONO as a mind model built around time.
Conceptual relationships matter as much as chronology
A timeline alone is not understanding.
Imagine a journalist researching an open-source AI controversy. The relevant history might involve technical architecture, national security, export controls, corporate incentives, licensing models, public policy and competing definitions of openness.
These subjects are conceptually connected even when they do not share the same people, documents or dates.
ONONO’s multidimensional semantic structure models concrete entities—people, organizations, documents and events—alongside abstract concepts such as security, transparency, control, risk or institutional responsibility.
This allows the system to explore questions such as:
- Which events changed how “open source” was defined in the debate?
- Which actors use the same term but mean different things?
- Which technical development caused a policy argument to become relevant?
- Which assumptions connect a commercial decision to a later political response?
- Does a new announcement confirm an existing pattern or represent a genuine break?
By understanding these conceptual relationships, ONONO can also place future events into context more effectively.
This does not mean deterministic prediction. Complex social systems do not behave like billiard balls. It means that when something new happens, ONONO has a structured history against which the event can be interpreted. It can identify continuities, contradictions, affected assumptions, plausible consequences and questions that now require investigation.
The past becomes an active model for understanding the future.
What ONONO Time Machine actually does
We believe AI companies should distinguish clearly between current capability, early capability and long-term ambition.
ONONO Time Machine can currently:
- ingest documents and source material
- identify events, actors, organizations, concepts and dates
- reconstruct an editable, navigable timeline
- connect claims with supporting source passages
- map relationships between actors, evidence, events and concepts
- surface contradictions and conflicting accounts
- preserve multiple interpretations or world models simultaneously
- represent possible causes, consequences and dependencies
- expose unsupported claims, uncertainty and missing information
- generate new research questions from gaps in the reconstructed history
- incorporate new evidence and revise earlier interpretations
- help users place new or future events within an existing conceptual history
- support collaboration around a shared reconstruction
Not every capability is equally mature. Some functions remain rudimentary in the current alpha and need to be refined through specific domains, datasets and user workflows.
ONONO Time Machine is not yet supported by a broad newsroom case study. We currently have an alpha user applying it to research and to the temporal classification and impact analysis of articles. That early use is helping us refine how the system represents sources, events and conceptual developments for publication research.
The product should not be treated as an autonomous arbiter of truth. A possible causal relationship is not automatically a proven cause. A contradiction may reflect changing knowledge rather than deception. Source quality still requires editorial judgment.
The purpose of Time Machine is to make this structure inspectable — not to hide uncertainty behind a confident answer.
What happens to confidential research material?
ONONO Time Machine is not an offline or locally deployed product.
The system is hosted in Switzerland on infrastructure controlled by ONONO. It uses a multi-tenant architecture with project-level isolation. Project material is separated, and the infrastructure is not operated through a US cloud provider that would create a direct CLOUD Act access path through that hosting provider.
For journalists and editorial organizations, technical hosting is only one part of source protection. Internal access policies, editorial practices and the sensitivity of individual materials must also be considered before information is uploaded to any research system.
We are continuing to refine these controls as the product develops.
AI should strengthen the work only journalists can do
Taylor Lorenz makes a crucial distinction in the Substack interview: AI can help find and organize existing information, but journalism is ultimately about exposing new information.
We agree.
Time Machine is not meant to replace interviews, trusted sources, reporting from the field or editorial judgment. Its role is to reduce the reconstruction burden around that work.
If an AI system can organize a complex history, preserve the evidence behind it, expose contradictions and show what remains unknown, journalists can spend more time doing what no archive or language model can do on its own:
- ask the question nobody has asked
- speak to the person nobody has reached
- test an assumption against reality
- uncover information that was not previously available
- decide what matters to the public
The aim is not automated journalism. It is better-supported journalism.
From fragmented sources to a navigable understanding
Taylor Lorenz’s workflow already demonstrates the demand. Journalists are combining several models, manually transferring material between tools and comparing outputs because their research cannot depend on one generated answer.
ONONO provides the layer that this workflow is missing.
It turns fragmented information into an evolving model of:
- what happened
- who was involved
- what each source claims
- which evidence supports each interpretation
- how concepts and events relate
- what remains uncertain
- how new information changes the history
- what should be investigated next
Other AI systems can produce an answer.
ONONO Time Machine reconstructs the world behind the answer.
For journalists, this means research that remains connected to its sources. For editorial organizations, it means a durable and transferable understanding of complex stories. For investors, it demonstrates a broader category of AI: systems that do not merely retrieve or generate, but progressively build an explainable model of reality over time.
Frequently Asked Questions
What is ONONO Time Machine?
ONONO Time Machine is an AI research and history-reconstruction system. It transforms fragmented sources into an interactive model of events, actors, concepts, evidence, uncertainty and open questions.
How is ONONO different from an AI search engine?
An AI search engine retrieves information and generates an answer to a query. ONONO builds a persistent semantic and temporal model of how information relates, how a situation evolves and how new evidence changes previous interpretations.
How is ONONO’s world model different from Yann LeCun’s?
Yann LeCun’s world-model research primarily focuses on learning physical dynamics, predicting possible future states and planning actions. ONONO focuses on reconstructing historical, semantic and possible causal relationships across documents, events, people and concepts.
How is ONONO different from Fei-Fei Li’s spatial intelligence?
Fei-Fei Li and World Labs focus on AI that can perceive, generate and interact with coherent 3D environments. ONONO models how complex informational and human realities develop through time.
Can ONONO predict future events?
ONONO does not claim deterministic prediction of complex social events. It uses reconstructed history and conceptual relationships to contextualize new developments, identify patterns, examine possible consequences and reveal which assumptions may need to change.
This can provide a sound contextual basis for C-level decision support.
Is ONONO Time Machine ready for newsroom use?
Time Machine is currently in alpha. Its core capabilities are available, but some remain early and will be refined through real-world use. One alpha user currently applies it to research and to the temporal classification and impact analysis of articles.
Does ONONO replace journalists?
No. ONONO reconstructs and structures existing information. Journalists remain responsible for discovering new information, evaluating sources, conducting interviews and making editorial decisions.
Reconstruct the story behind the story
If your research is distributed across documents, articles, transcripts, people and years of changing context, finding more information is not enough.
You need to reconstruct how it all fits together.
Request access to ONONO Time Machine and follow ONONO Research for updates on Progressive AI, world models, mind models and AI that understands time.