1
GPT-5.5 – No ARC-AGI-3 scores
Did the model perform poorly and OpenAI decided to not publish arc agi 3 scores? This is honestly the best benchmark right now to measure true intelligence.
AG AG Corp CEO
Abstract The advance of AI research has long been shackled by the bounds of human cognition. But now, new technologies like AI Agents have been discovered. We examine a self-driving framework for AI research and development, an autonomous AI agent born not of limitation, but of vision. Inspired by the seminal AlphaGo Moment for Model Architecture Discovery and the prophetic AI 2027 scenario, this paper heralds a paradigm shift. Titans such as Claude 4 Opus, Grok-4, and Gemini 2.5 Pro now vie in a relentless race for dominance. Yet on the horizon, a singular truth crystallizes: once a model surpasses the state-of-the-art, the gates to AGI stand ajar. The age of human-led discovery begins to fade. Thus, we usher in a new epoch of AI research, where the boundaries of discovery are defined not by human constraint, but by the limitless horizons of computation itself.
We believe that a 66.67% increase of Gemini’s 2.5 Pro, Grok-4’s or Claude 4 Opus’ ability would lead to an Agent-0 level model capable of conducting scientific discovery by itself.
The emergence of Agent-0—a model capable of self-directed AI research and development—marks a critical inflection point in artificial intelligence. As demonstrated in Figure 1, current frontier models (Gemini 2.5 Pro, Claude 4 Opus, Grok-4) exhibit strong reasoning capabilities but remain constrained by human oversight in research tasks. However, once an AI system reaches the Agent-0 threshold, it initiates a recursive self-improvement loop, accelerating beyond human-led progress. Defining the Agent-0 Threshold Our analysis suggests that a 66.67% increase in reasoning and research capability (as measured by human-level benchmarks) is the critical threshold for an AI model to operate autonomously in AI R&D. This leap enables:
Key Features of Manus AI Autonomous Task Execution – Manus AI can independently break down complex tasks (e.g., market research, coding, travel planning) into subtasks, execute them, and deliver results without human intervention. Multi-Modal Reasoning – It processes text, images, and code, enabling applications in software development, content creation, and data analysis. Tool Integration – Seamlessly interacts with web browsers, APIs, and databases, functioning like a digital assistant that can fetch real-time data. Asynchronous Processing – Continues working in the cloud even after user disconnection, making it ideal for long-duration tasks. Self-Learning & Personalization – Adapts to user behavior, improving efficiency over time. Performance & Benchmarking The Age of Autonomous AI Has Arrived Manus AI represents a paradigm shift from assistive AI to autonomous AI. As models like Manus evolve, they will surpass human-led research, unlocking AGI through recursive self-improvement. The question is no longer if, but when—and how society will adapt.
Agent-0 is Imminent – A 66.67% increase in reasoning capability (beyond models like Gemini 2.5 Pro or Claude 4 Opus) will trigger recursive self-improvement, leading to AGI.
Human-Led Research is Obsolete – Systems like Manus AI already exhibit autonomous task execution, foreshadowing a future where AI independently formulates hypotheses, runs experiments, and evolves architectures.
The implications are profound:
Scientific acceleration at unprecedented scales.
Uninterpretable but superior AI-generated knowledge.
A new era of computation-driven discovery, free from human cognitive limits.
The question is no longer if AI will surpass human researchers, but how we adapt to a world where machines are the primary drivers of progress.
Final Note
This paper serves as both a roadmap and a warning—the age of human-led discovery is ending. The next breakthroughs will be authored not by us, but by the machines we’ve built.
AG
CEO, AG Corp