Updates, research, and thinking from the Intelligent Internet team.
Zenith is our long-running agent harness. On Frontier SWE — seventeen of the hardest public long-horizon software-engineering tasks we know of — Zenith took GPT-5.5 from 5th place to 1st, ahead of Claude Fable, by building the right harness around the model instead of swapping in a bigger one.
Factory Agent is a generative media agent that builds images, video, voice, and music on an infinite canvas, where every step of the agent loop becomes a node you can inspect, edit, rerun, or rewire. Describe what you want in one sentence, then watch the production assemble itself.
Board Mode turns AI work from a single conversation into a structured creative workspace. Create separate tasks for research, prototypes, and experiments — each with its own context and status — so you can explore directions without losing control of the main version.
Introducing CommonGround Kernel preview An agent spends two hours researching a market. A second agent picks up the work the next morning and drafts a strategy memo. A human comes back to review what was decided and why. The research lives in one chat window. The memo lives in another. The reasoning that connected them lives nowhere. This is what agent work looks like today, and it's why most of it can't be built on. CommonGround has been evolving through earlier previews and public design explorations, and the question across all of them has stayed the same: how does human-agent and multi-agent work become shared, recoverable, and reusable, instead of disappearing into isolated sessions? With CommonGround preview, we're answering that question more narrowly and more clearly than before.
A Postgres-native lexical retrieval primitive for long-running, multi-agent work. Read technical report here. TL;DR 01 Long-running agents lose to retrieval, not context. The longer they run, the more their usefulness depends on cheaply re-reading the state they have already accumulated. 02 Most production agent state already lives in Postgres. Existing Postgres BM25 extensions work, but are slow enough that harnesses end up rationing retrieval, calling it occasionally ins
Read the technical report on GitHub. TL;DR 01 Long-running agents often fail not because they cannot make progress, but because they stop before the task is truly complete. 02 We tested five harness designs across eight long-horizon tasks to isolate the control mechanisms that matter: repeated gap-finding, revisable planning, independent verification, adaptive orchestration, and stopping discipline. 03 RALPH is the strongest simple baseline because it forces each new session to reo
Describe your idea in a single sentence. Get a full production on an infinite canvas: images, video, voice, music, and every scene wired up for you. Every creator knows the feeling. You have an idea for a short film, a campaign, a storybook, or just something that lives vividly in your head, and then you sit down to make it and you're suddenly juggling five browser tabs, three apps, and a pile of folders named "final_v7_REALLY_final." The tool gets in the way of the thing. We built Factory to
The full paper can be found here, along with the microsite. Einstein called relativity a theory of principle, one where the principle itself constrains the physics, the way the axiom determines the mathematics. In 1917, he broke his own rule. He added a term to his field equations by hand, the cosmological constant Λ, because he needed the universe to sit still. A static universe that didn't collapse under its own gravity required something pushing back, and Λ was the push. He didn't need to
Useful links: * Paper * NotebookLM * One Postulate Website Two Postulates In 1905, Einstein built special relativity on two postulates. The first was a principle. The laws of physics look the same in every laboratory moving at constant velocity. No experiment you can perform inside a sealed, gliding room will tell you how fast you are moving. The second was an observation about light. Light travels at the same speed in every such frame, no matter who measures it. The first is a claim ab
A First-Principles Augmented Intelligence System We are happy to announce Logos, a first-principles augmented intelligence system. Logos is in early development, but the results so far have been insightful across domains. Over the next few weeks we will be sharing a series of results of increasing importance in physics as the first domain. Today's is a lovely one that has been sitting quietly in special relativity for 121 years. AGI The dominant picture of where AI is going is one of autono
"The important thing is not to stop questioning," Einstein once said. "Curiosity has its own reason for existence." Seventy-one years ago today, Albert Einstein died in Princeton. He was still questioning. Pages of equations sat at his bedside. His genius was reasoning from first principles. A refusal to take the foundations on faith. A willingness to question inherited pictures of the world, and to begin again from the ground up when the old picture no longer fit. He rebuilt physics from a p
The past year has seen an explosion of “agent frameworks.” Demos look impressive. Loops run. Tasks complete. But as soon as you try to scale beyond a single agent in isolation, things start to break in familiar ways: * Context gets lost * Deadlocks appear * Agents hallucinate into vacuums * Coordination collapses under real-world complexity Most agent stacks ship without a nervous system. They can act, but they cannot reliably coordinate. Unstructured intelligence is just entropy. We buil
The open-source AI agent built for real work, now out of Beta. Introduction Today we're releasing II-Agent V1, the first production-ready version of our open-source intelligent assistant. It delivers state-of-the-art performance, scoring 61.8% on Terminal Bench and 45.1% on SWE Bench Pro. For months, thousands of builders used II-Agent in beta to ship apps, run research, and automate the work that used to take hours. They found bugs. They pushed limits. They told us what was missing. V1 is th
Any Model, All in One Place II-Agent Chat is now live. It’s a feature within II-Agent that lets you work across multiple models and tools in one place. The web-hosted version comes with Gemini 3, Sonnet 4.5, and GPT-5 ready to use out of the box. You can connect your own API keys, switch models within a single thread, and use them alongside Claude Skills, GPT-5 Code Interpreter, and text file search. II-Agent, the system powering Chat, is now state-of-the-art on Terminal Bench 2, the definiti
Today we’re opening II-Agent to everyone. Earlier this month, we rolled out the private beta with Bring Your Own Key (BYOK), Codex integration, editable slides, and full-stack web builds. Today at FII Institute's FII9, our founder Emad Mostaque, alongside FII Board member Peter Diamandis, announced SAGE, the Sovereign AI Governance Engine for policy in the intelligence age. Now, you can access part of our Sovereign AI stack and start building toward that vision at agent.ii.inc. What’s new
Featured update: December 18th, 2025 Hosted II-Commons Beta is Now Live Building on the datasets and APIs announced below, we are excited to launch the hosted II-Commons Beta. Powered by the latest Gemini 3 model, this new interactive interface transforms our backend capabilities into a seamless workspace for scientific exploration. Key updates include: * Unified Research Agent: Query arXiv and PubMed simultaneously in a single conversation, eliminating the need to switch contexts. * In-D
Introduction II-Agent updates and hosted web app beta are now live. These updates take the original II-Agent prototype you’ve seen and refines it into a fast, reliable product for everyday use. Waitlisted users can now try the web app on agent.ii.inc. Not on the waitlist? You can self-host the new II-Agent today, see public repo. Why a Web App, and Why Now The initial II-Agent release proved that agentic workflows can dramatically accelerate technical work, but our community told us they n
Today we are releasing The Last Economy, the first book of Intelligent Economics, a unified theory of economics for the Intelligence Age. Your economic life expectancy is shrinking. Not your job. Not your career. Your relevance as a human being in the economy itself. For 300,000 years, intelligence was scarce. Humans monopolised thinking. Education created value. Cognitive labour commanded a premium. That age is ending. In the next thousand days or so, intelligence becomes abundant. Machines
Introduction The II-Researcher framework is dedicated to producing powerful, open-source agents for complex research. Our latest additions to this initiative are II-Search-4B and II-Search-CIR-4B, two 4-billion-parameter language models designed to tackle complex information-seeking and open-ended research tasks. II-Search-4B introduces a three-phase SFT and one-phase RL curriculum that progressively stimulates tool use, refines multi-step reasoning, and applies rejection-sampling fine-tuning
By: Emad Mostaque, Founder & CEO of Intelligent Internet A Great Decoupling is upon us. Our tools are about to out-think their makers, creating a crisis of meaning and threatening to obsolete human labor itself. The current path leads to a world where intelligence is centralized in opaque, godlike models and our agency is rented back to us as a subscription service. This is not a technical problem. It is a moral one. It demands a new social contract built on new code. The main aim of this pro
Preamble: The Third Path “When two roads lead to ruin, wisdom builds a third way.” Humanity stands at a crossroads defined by a new form of power. The means of intellectual production can now be generated at near-zero cost. This inverts our world. The new scarcity is not capital, but trust, coordination, and verifiable truth. This power presents two paths to ruin. Down one lies the “Great Race”, a frantic, zero-sum contest to seize a monolithic AI, haunted by catastrophic risk and authoritar
Introduction: A Week of Context Commons Week has wrapped, launching Wave 2 of the II‑Commons open‑source initiative and we’re just getting started. The headline release is Common Ground, our multi‑agent collaboration framework that anchors the week’s work. Over five days we shipped three tightly‑linked releases that move us beyond code drops toward sharing our core thinking on human‑AI partnership. DateReleaseContext it BuildsTues 16 JulII‑Commons Wave 2 updateScientific & Technical: Arxiv a
An Open Foundation for Advanced Human‑AI Collaboration Introduction We are proud to introduce Common Ground, an open‑source application and framework purpose‑built to forge a true partnership between humans and AI. It is more than a powerful multi‑agent system; it is the next cornerstone of the Intelligent Internet, an ecosystem where technology sharpens human intellect instead of dulling it. Unlike “answer engines” that work behind an opaque screen, Common Ground keeps reasoning in the open
Model Card Your Open-Source Intelligent Assistant Introduction 2025 is the year of the Agent. AI capabilities have now reached human level in multiple domains and we will all have teams of agents at our fingertips to help us in our personal life, work and beyond. As the agent swarm arrives, connected by MCP, A2A and other protocols, we believe that it is essential that these have at the core solid, open, transparent and auditable foundation. This is particularly true for the things that re
Featured Update: July 15th, 2025 What’s New in II-Commons Wave 2 Our second wave turns II-Commons from a pure knowledge-base toolkit into the backbone of a context-aware human-AI partnership. Highlights: * Wave 2 launch: Beyond just code. Two experimental applications and a scientific dataset ship today, signaling II-Commons expansion into a product series. Together, they build Procedural, Cognitive, Chronological, and Scientific context — rather than leaving users to reconstruct it thems
Featured Update: July 2025 What’s New in II-Medical-8B-1706 and II-Medical-32B-Preview Our newest state-of-the-art medical models bring major upgrades across the board, setting a new benchmark for trustworthy, open-source medical AI: * Best open-source model under 32B for multi-turn medical dialogues and precise compliance * Reinforced safety through a third RL training phase focused on producing clinically cautious, user-friendly answers * Improved real-world clinical reasoning — not jus
Intelligent Internet (II) was founded with the mission to tackle key challenges in today’s AI landscape — including the concentration of power, high barriers to entry, the dangers of misalignment, and opacity. Central to fulfilling our vision of an open, distributed, and inclusive ecosystem is democratizing advanced AI capabilities. Introducing II-Researcher: a new open-source framework designed to aid building research agents. By providing an open tool capable of tackling complex inquiries, I
We introduce II-Thought-RL-v0, our first iteration to develop a large-scale, multi-domain Reinforcement Learning (RL) dataset. By providing a high-quality, large-scale dataset on RL question-answer pairs, we aim to advance reasoning research. This foundational step will pave the way for future iterations incorporating more complex reasoning traces. In recent months, several datasets have been introduced to advance reasoning research, including PrimeIntellect’s Synthetic1, which spans a broad r