Artificial Inventor & Synthetic Scientist
A proprietary collective-intelligence platform in which multiple AI models challenge, verify and combine one another before producing a result, with scientific, engineering, patent and supervised-robotics tools.
Have We Reached Superintelligence?
The video scene is science fiction. The question here is whether a much slower version of that idea — collective intelligence connected to tools, laboratories and robots — is beginning to exist in reality.
A collective brain of up to twelve models
Artificial Inventor Pro can convene up to twelve AI models or roles. The visible pipeline uses Advocate → Skeptic → Arbiter → Closer and can add five independent synthetic reviewers. Different models propose, attack, verify and synthesize a shared answer. The exact model selection and arrangement are proprietary.
What Artificial Inventor does
Beyond the multi-model debate, the platform is designed to turn reasoning into scientific, engineering and invention work.
Patent and prior-art research
Searches patent literature and builds feature-by-feature novelty maps with cited references.
Engineering tools
Integrates CAD, FEA/CFD, AI/ML engineering, manufacturing, prototyping and technical-document production.
Supervised action
Can browse, search, source components and propose robotic actions, while consequential real-world actions remain subject to human authorization.
An extraordinary signal that still needs independent replication
ArtificialInventor.ai reports a preliminary, recorded 98.5% score on three custom, blind and original variations of the Humanity’s Last Exam framework. That is not an official HLE score and is not external validation. Scale Labs’ public leaderboard listed a top score of 46.44% on 30 August 2026. If the 98.5% result is reproduced under an independent protocol, the gap would be extraordinary.
A public challenge to CAIS and Scale AI
The project says it has repeatedly contacted the Center for AI Safety and Scale AI offering to pay for an independent Humanity’s Last Exam evaluation and has not received a test price or substantive response. Lack of a response does not itself demonstrate misconduct. The public challenge is simple: run a fair, reproducible test and publish the result.
Science fiction — or partly a question of speed?
One provocative possibility is that for some problems the visible difference between fiction and present collective systems is partly latency: today a multi-model system may take minutes where a fictional mind answers in milliseconds. That is a hypothesis, not a conclusion. Superintelligence would also require generality, reliability, autonomy, adaptation and sustained performance across domains not optimized for a test.
The brain–computer interface is leaving science fiction
Neuralink is conducting clinical trials involving computer and robotic-arm control through thought, while Paradromics is developing implanted interfaces intended to translate neural signals into text, speech or computer commands. Faster human–machine communication could make the boundary between human intelligence, collective AI and robotic tools less distinct.
Synthetic scientists: bringing reasoning into the physical world
Artificial Inventor describes a Robotic Control interface in which AI proposes an action and a human authorizes it before execution. The goal is laboratories, workshops, manufacturing, industrial inspection, fire response and other hazardous work where AI reasoning can connect to sensors and robots without eliminating human supervision.
Open-model components can be hosted in project data centers
The broader project plan is to host open-model components used by Artificial Inventor in Spanish data-center infrastructure, linking the scientific platform to renewable energy, recovered heat and local computing capacity.
HLE-oriented engineering
The technical roadmap calls for deterministic calculation verification, exact-answer formatting, calibrated confidence, category-aware panel weighting, full multimodal access for critic models and self-consistency sampling — controls intended to reduce arithmetic, formatting, vision and overconfidence failures.
Worldwide invitation: bring the questions almost nobody can solve
We invite universities, research institutes, scientists, engineers and educators to submit original, difficult and verifiable questions that most humans cannot solve easily. We want third-party tests with unseen questions, rules fixed before execution and complete published results. The question should be resolved by experiments, not marketing.