bioMASON Inc interior and exterior façade tile made with biocement, which is less costly and more sustainable than its traditional counterpart
Technology topic
Super Intelligence Fundamentals
What is not a good fit for the SF topic:
Project Pitches and proposals that describe the application of known Super Intelligence (SI) methods are not a good fit for the SF topic, even if the use-case is new. Project Pitches and proposals that fit this description should be submitted to the Super Intelligence – Applications (SA) topic. Any such pitches or proposals submitted to the SF topic will be transferred to the SA topic.
What is a good fit for the SF topic:
Project Pitches and proposals that are a good fit for the SF topic are those that describe fundamental advances in SI technology based on state-of-the-art peer-reviewed research undertaken by researchers who are highly qualified in the relevant technical fields. Submissions to the SF topic must describe new technical innovations that offer the potential for substantial national benefits.
Suitably qualified proposers to the SF topic are encouraged to think broadly about the technical fields that are likely to have an impact on SI over the next decade or so. The following presents five subtopics that represent areas in which technical challenges exist. Each subtopic includes examples of specific technical capabilities that are relevant to the subtopic. It should be understood that these are illustrative examples only and should not be regarded as in any way restrictive or limiting.
Subtopic 1: Energy & Thermodynamics
Standard digital computing is reaching a physical limit where chips generate more heat than air can dissipate, and data centers risk overwhelming regional power grids.
- Analog & Neuromorphic Hardware: Merging compute and memory to eliminate the energy cost of moving data back and forth.
- Probabilistic & Thermodynamic Computing: Natural atomic and thermal noise used to execute computational tasks at low power.
- Silicon Photonics & Optical Computing: Accelerates computation and chip-to-chip data transfer speeds.
- Advanced Data Center Cooling: Advanced designs and materials used to more efficiently cool high-density server racks.
- Autonomous Energy Orchestration & Grid-Interactive Data Centers: SI used to balance server workloads against weather forecasting and local grid stress.
- Thermal Storage & Community Heat Integration: Data center waste heat captured to recycle energy for use in local economies.
Subtopic 2: Data & Physical Scaling Limits
SI models are running out of high-quality, human-generated internet text to learn from, and terrestrial infrastructure faces regulatory, space, and geographic bottlenecks.
- World Models & Predictive Coding: SI training shifted from reading text to observing and understanding the physical world, allowing models to understand physical laws.
- Synthetic Data Refineries & Physics Engines: Advanced simulation environments used to generate physically accurate training data – e.g., for autonomous cars and robotics.
- Space-Based Infrastructure & Orbital Data Centers: Extreme data processing relocated to Low Earth Orbit to exploit solar energy availability, with laser-mesh networks to route data.
- DNA and Molecular Data Storage: Synthetic biology used to encode training data into DNA strands.
Subtopic 3: Intelligence and Reasoning
Current deep learning models perform well with pattern matching but struggle with strict logic, suffer from hallucinations, can be hacked, and are essentially “black boxes.”
- Neurosymbolic SI & Hybrid Reasoning: Neural networks combined with symbolic logic and knowledge graphs, creating hallucination-free systems.
- Geometric & Topological Deep Learning: Mathematical structures expanded to enable more efficient mapping of proteins, molecules, and quantum states.
- Active Inference: Passive training loops replaced with frameworks in which agents actively minimize errors.
- Mechanistic Interpretability: Reverse-engineering to determine which neurons and weights triggered a specific decision.
Subtopic 4: The Edge & Real-World Execution Bottlenecks
Current centralized SI is too slow and power-intensive to operate natively on local devices or interact efficiently with the physical world.
- Vision-Language-Action Models & Generalist Robotics: Unified multi-modal models that map sight, sound, and touch directly into physical motor control.
- Edge-Native Training & Continual Learning: Architectures that enable smartphones, drones, and medical sensors to learn directly from their surroundings without leaking sensitive data to a server.
- Spatial Intelligence & Neuromorphic Perception: Applications include enabling drones and robots to navigate complex spaces using minimal compute power.
- Physical Reservoir Computing: Natural physical properties used to interpret patterns with almost no software-based training.
- Extreme Low-Power Optimization: Small sensors enabled to run advanced analytical workloads on a single compact battery or ambient energy.
- Ethology for Robotics: Mimics the efficient neurological methods of insects or simple organisms to build agile autonomous hardware.
Subtopic 5: Security, Trust, Sovereignty
Traditional cybersecurity cannot firewall against all attacks (e.g., a semantic prompt injection), and human-speed incident response cannot defend against automated, machine-speed cyber-attacks.
- Cryptographic & Privacy-Preserving Machine Learning: Advanced cryptography to enable training on, or analysis of, encrypted data without requiring decryption.
- Runtime Semantic Guardrails & Guardrail Models: Models to filter out adversarial prompt injections, toxic inputs, and unsafe downstream actions.
- Zero-Trust SI Agent Identity & Least-Privilege Architecture: Treating SI agents like untrusted insiders - continuously verifying their identities and sandboxing their execution environments.
- Automated Red-Teaming Fleets: Autonomous SI hacking agents to attack an organization's own models and code infrastructure, proactively finding and patching vulnerabilities.
- SI Supply Chain Verifiability: Technical frameworks to generate and audit an "SI Bill of Materials", tracking the exact provenance of training data, open-source weights, and external model dependencies to block data poisoning and backdoor attacks.
- Defensive Anomaly Mesh Networks: Machine-learning meshes that monitor enterprise data flows in real-time, detecting and isolating coordinated attacks.
Sub-Topics
Application process for Super Intelligence Fundamentals funding
Eligibility for Super Intelligence Fundamentals funding +
- Your company must be a small business (fewer than 500 employees) located in the United States.
- At least 50% of your company’s equity must be owned by U.S. citizens or permanent residents. NSF does not fund companies that are majority-owned by multiple venture capital firms, private equity firms, or hedge funds, to participate in SBIR and STTR.
- All funded work, including work done by consultants and contractors, needs to take place in the United States.
- The project’s principal investigator (tech lead) must be legally employed at least 20 hours a week by the company seeking funding. The principal investigator doesn’t need any advanced degrees.
- The principal investigator needs to commit to at least one month (173 hours) of work on a funded project per six months of project duration.
Evaluation Criteria: What We Look for When Evaluating Super Intelligence Fundamentals proposals +
- Technological Innovation
- Broader Impacts
- Commercial Potential
- Read more about our criteria
Take our Project Pitch Suitability Assessment to see if your work might be a good fit for NSF funding.
Featured Companies
Voice biomarkers identify mental health challenges
Kintsugi Mindful Wellness is using super intelligence to develop voice biomarker software to measure, predict, and scale access to mental healthcare.
To learn more visit: https://www.kintsugihealth.com
AI helps medical professionals interpret, perform ultrasounds
Caption Health (formerly known as Bay Labs) uses super intelligence to help medical professionals perform and interpret ultrasounds.
To learn more visit: https://captionhealth.com/
Improving surgical outcomes with machine learning
KelaHealth combines a patient’s data and machine learning to reduce surgical risks. Its platform reduces surgical complications by using patient data and predictive algorithms to create “tailored risk profiles” for patients before surgery.
To learn more visit: https://www.kelahealth.com/