Why Science? | Centre for Frontier AI Research

AI research begins with a task, environment, success criterion and failure cost. Frontier AI research studies the foundations and limits of learning systems before dependable applications are assumed. CFAR lists AI for Science, artificial general intelligence, resilient and safe AI, sustainable AI, and theory and optimisation as its current pillars. Chapter 1 turns one frontier question into a testable claim with baselines, uncertainty and clearly stated failure conditions.

Specify what information is available and what decision follows. A rigorous AI workflow defines the task, population or system, data provenance, representation, model, compute, baseline, split, metric and deployment context. Researchers check leakage, distribution shift, calibration, efficiency and reproducibility, then compare against simpler methods. Scientific applications also require consistency with domain knowledge and experimental validation rather than treating generated outputs as discoveries.

A vague goal such as intelligence cannot be evaluated responsibly. Capability labels are not proof of general intelligence, understanding, truth or safety. Benchmark gains may be narrow; robust performance can fail under new attacks or environments; and a model cannot own ethical or scientific responsibility. High-stakes use needs sector evidence, human oversight, security, governance and a named decision owner with authority to stop or roll back the system.

Use “Frame the intelligence task” with synthetic data or a small public teaching dataset. Write a task contract. Write a data sheet and model card, establish a simple baseline, reserve a test set and probe one shift or failure mode. Report compute and uncertainty. Never upload confidential or personal data, and do not deploy a classroom model to make real decisions about people.

Did you know? AI for Science can search patterns or propose candidates, yet experiments and domain reasoning still decide whether a candidate represents the physical world. Chapter 1 shows why constraints, interpretability and active learning can reduce wasted exploration without converting a prediction into evidence by itself.

Capabilities include mathematics, optimisation, statistics, computer science, software engineering, domain science, human factors, cybersecurity and ethics. Students can start with probability, algorithms, linear algebra, experiments and clear documentation. Current programmes set entry requirements. A strong AI portfolio explains data, baselines, compute, failures and the human decision boundary instead of promising a career or universal intelligence.

Contents · Next chapter

Datasets encode collection choices, labels, licences, populations and historical context. Frontier AI research studies the foundations and limits of learning systems before dependable applications are assumed. CFAR lists AI for Science, artificial general intelligence, resilient and safe AI, sustainable AI, and theory and optimisation as its current pillars. Chapter 2 turns one frontier question into a testable claim with baselines, uncertainty and clearly stated failure conditions.

Create data sheets and inspect quality, leakage and representation. A rigorous AI workflow defines the task, population or system, data provenance, representation, model, compute, baseline, split, metric and deployment context. Researchers check leakage, distribution shift, calibration, efficiency and reproducibility, then compare against simpler methods. Scientific applications also require consistency with domain knowledge and experimental validation rather than treating generated outputs as discoveries.

More data does not automatically reduce bias or improve relevance. Capability labels are not proof of general intelligence, understanding, truth or safety. Benchmark gains may be narrow; robust performance can fail under new attacks or environments; and a model cannot own ethical or scientific responsibility. High-stakes use needs sector evidence, human oversight, security, governance and a named decision owner with authority to stop or roll back the system.

Use “Build trustworthy datasets” with synthetic data or a small public teaching dataset. Audit a teaching dataset. Write a data sheet and model card, establish a simple baseline, reserve a test set and probe one shift or failure mode. Report compute and uncertainty. Never upload confidential or personal data, and do not deploy a classroom model to make real decisions about people.

Did you know? AI for Science can search patterns or propose candidates, yet experiments and domain reasoning still decide whether a candidate represents the physical world. Chapter 2 shows why constraints, interpretability and active learning can reduce wasted exploration without converting a prediction into evidence by itself.

Capabilities include mathematics, optimisation, statistics, computer science, software engineering, domain science, human factors, cybersecurity and ethics. Students can start with probability, algorithms, linear algebra, experiments and clear documentation. Current programmes set entry requirements. A strong AI portfolio explains data, baselines, compute, failures and the human decision boundary instead of promising a career or universal intelligence.

Contents · Previous chapter · Next chapter

Theory and optimisation examine learning dynamics, objectives, guarantees and efficient search. Frontier AI research studies the foundations and limits of learning systems before dependable applications are assumed. CFAR lists AI for Science, artificial general intelligence, resilient and safe AI, sustainable AI, and theory and optimisation as its current pillars. Chapter 3 turns one frontier question into a testable claim with baselines, uncertainty and clearly stated failure conditions.

State assumptions and compare algorithms under controlled settings. A rigorous AI workflow defines the task, population or system, data provenance, representation, model, compute, baseline, split, metric and deployment context. Researchers check leakage, distribution shift, calibration, efficiency and reproducibility, then compare against simpler methods. Scientific applications also require consistency with domain knowledge and experimental validation rather than treating generated outputs as discoveries.

A theorem under assumptions cannot guarantee messy deployment behaviour. Capability labels are not proof of general intelligence, understanding, truth or safety. Benchmark gains may be narrow; robust performance can fail under new attacks or environments; and a model cannot own ethical or scientific responsibility. High-stakes use needs sector evidence, human oversight, security, governance and a named decision owner with authority to stop or roll back the system.

Use “Study theory and optimisation” with synthetic data or a small public teaching dataset. Map assumptions to tests. Write a data sheet and model card, establish a simple baseline, reserve a test set and probe one shift or failure mode. Report compute and uncertainty. Never upload confidential or personal data, and do not deploy a classroom model to make real decisions about people.

Did you know? AI for Science can search patterns or propose candidates, yet experiments and domain reasoning still decide whether a candidate represents the physical world. Chapter 3 shows why constraints, interpretability and active learning can reduce wasted exploration without converting a prediction into evidence by itself.

Capabilities include mathematics, optimisation, statistics, computer science, software engineering, domain science, human factors, cybersecurity and ethics. Students can start with probability, algorithms, linear algebra, experiments and clear documentation. Current programmes set entry requirements. A strong AI portfolio explains data, baselines, compute, failures and the human decision boundary instead of promising a career or universal intelligence.

Contents · Previous chapter · Next chapter

Representation learning transforms raw data into features that support a task. Frontier AI research studies the foundations and limits of learning systems before dependable applications are assumed. CFAR lists AI for Science, artificial general intelligence, resilient and safe AI, sustainable AI, and theory and optimisation as its current pillars. Chapter 4 turns one frontier question into a testable claim with baselines, uncertainty and clearly stated failure conditions.

Probe invariances, information loss and transfer. A rigorous AI workflow defines the task, population or system, data provenance, representation, model, compute, baseline, split, metric and deployment context. Researchers check leakage, distribution shift, calibration, efficiency and reproducibility, then compare against simpler methods. Scientific applications also require consistency with domain knowledge and experimental validation rather than treating generated outputs as discoveries.

A compact embedding is not an explanation or neutral view. Capability labels are not proof of general intelligence, understanding, truth or safety. Benchmark gains may be narrow; robust performance can fail under new attacks or environments; and a model cannot own ethical or scientific responsibility. High-stakes use needs sector evidence, human oversight, security, governance and a named decision owner with authority to stop or roll back the system.

Use “Learn useful representations” with synthetic data or a small public teaching dataset. Visualise fictional representations. Write a data sheet and model card, establish a simple baseline, reserve a test set and probe one shift or failure mode. Report compute and uncertainty. Never upload confidential or personal data, and do not deploy a classroom model to make real decisions about people.

Did you know? AI for Science can search patterns or propose candidates, yet experiments and domain reasoning still decide whether a candidate represents the physical world. Chapter 4 shows why constraints, interpretability and active learning can reduce wasted exploration without converting a prediction into evidence by itself.

Capabilities include mathematics, optimisation, statistics, computer science, software engineering, domain science, human factors, cybersecurity and ethics. Students can start with probability, algorithms, linear algebra, experiments and clear documentation. Current programmes set entry requirements. A strong AI portfolio explains data, baselines, compute, failures and the human decision boundary instead of promising a career or universal intelligence.

Contents · Previous chapter · Next chapter

Generalisation asks whether performance extends beyond training examples and conditions. Frontier AI research studies the foundations and limits of learning systems before dependable applications are assumed. CFAR lists AI for Science, artificial general intelligence, resilient and safe AI, sustainable AI, and theory and optimisation as its current pillars. Chapter 5 turns one frontier question into a testable claim with baselines, uncertainty and clearly stated failure conditions.

Use held-out, shifted and external tests with confidence intervals. A rigorous AI workflow defines the task, population or system, data provenance, representation, model, compute, baseline, split, metric and deployment context. Researchers check leakage, distribution shift, calibration, efficiency and reproducibility, then compare against simpler methods. Scientific applications also require consistency with domain knowledge and experimental validation rather than treating generated outputs as discoveries.

Random test splits may preserve hidden duplicates or shortcuts. Capability labels are not proof of general intelligence, understanding, truth or safety. Benchmark gains may be narrow; robust performance can fail under new attacks or environments; and a model cannot own ethical or scientific responsibility. High-stakes use needs sector evidence, human oversight, security, governance and a named decision owner with authority to stop or roll back the system.

Use “Test generalisation” with synthetic data or a small public teaching dataset. Design a shift test. Write a data sheet and model card, establish a simple baseline, reserve a test set and probe one shift or failure mode. Report compute and uncertainty. Never upload confidential or personal data, and do not deploy a classroom model to make real decisions about people.

Did you know? AI for Science can search patterns or propose candidates, yet experiments and domain reasoning still decide whether a candidate represents the physical world. Chapter 5 shows why constraints, interpretability and active learning can reduce wasted exploration without converting a prediction into evidence by itself.

Capabilities include mathematics, optimisation, statistics, computer science, software engineering, domain science, human factors, cybersecurity and ethics. Students can start with probability, algorithms, linear algebra, experiments and clear documentation. Current programmes set entry requirements. A strong AI portfolio explains data, baselines, compute, failures and the human decision boundary instead of promising a career or universal intelligence.

Contents · Previous chapter · Next chapter

Multimodal systems learn from text, images, signals, graphs or scientific measurements. Frontier AI research studies the foundations and limits of learning systems before dependable applications are assumed. CFAR lists AI for Science, artificial general intelligence, resilient and safe AI, sustainable AI, and theory and optimisation as its current pillars. Chapter 6 turns one frontier question into a testable claim with baselines, uncertainty and clearly stated failure conditions.

Align sources, timing, missingness and modality-specific quality. A rigorous AI workflow defines the task, population or system, data provenance, representation, model, compute, baseline, split, metric and deployment context. Researchers check leakage, distribution shift, calibration, efficiency and reproducibility, then compare against simpler methods. Scientific applications also require consistency with domain knowledge and experimental validation rather than treating generated outputs as discoveries.

Adding modalities can amplify inconsistency or spurious cues. Capability labels are not proof of general intelligence, understanding, truth or safety. Benchmark gains may be narrow; robust performance can fail under new attacks or environments; and a model cannot own ethical or scientific responsibility. High-stakes use needs sector evidence, human oversight, security, governance and a named decision owner with authority to stop or roll back the system.

Use “Combine multiple modalities” with synthetic data or a small public teaching dataset. Build a fusion map. Write a data sheet and model card, establish a simple baseline, reserve a test set and probe one shift or failure mode. Report compute and uncertainty. Never upload confidential or personal data, and do not deploy a classroom model to make real decisions about people.

Did you know? AI for Science can search patterns or propose candidates, yet experiments and domain reasoning still decide whether a candidate represents the physical world. Chapter 6 shows why constraints, interpretability and active learning can reduce wasted exploration without converting a prediction into evidence by itself.

Capabilities include mathematics, optimisation, statistics, computer science, software engineering, domain science, human factors, cybersecurity and ethics. Students can start with probability, algorithms, linear algebra, experiments and clear documentation. Current programmes set entry requirements. A strong AI portfolio explains data, baselines, compute, failures and the human decision boundary instead of promising a career or universal intelligence.

Contents · Previous chapter · Next chapter

General-purpose models aim to transfer across tasks or contexts with limited adaptation. Frontier AI research studies the foundations and limits of learning systems before dependable applications are assumed. CFAR lists AI for Science, artificial general intelligence, resilient and safe AI, sustainable AI, and theory and optimisation as its current pillars. Chapter 7 turns one frontier question into a testable claim with baselines, uncertainty and clearly stated failure conditions.

Measure breadth, transfer cost, controllability and negative transfer. A rigorous AI workflow defines the task, population or system, data provenance, representation, model, compute, baseline, split, metric and deployment context. Researchers check leakage, distribution shift, calibration, efficiency and reproducibility, then compare against simpler methods. Scientific applications also require consistency with domain knowledge and experimental validation rather than treating generated outputs as discoveries.

Many benchmark skills do not prove universal competence. Capability labels are not proof of general intelligence, understanding, truth or safety. Benchmark gains may be narrow; robust performance can fail under new attacks or environments; and a model cannot own ethical or scientific responsibility. High-stakes use needs sector evidence, human oversight, security, governance and a named decision owner with authority to stop or roll back the system.

Use “Develop general-purpose systems carefully” with synthetic data or a small public teaching dataset. Create a capability matrix. Write a data sheet and model card, establish a simple baseline, reserve a test set and probe one shift or failure mode. Report compute and uncertainty. Never upload confidential or personal data, and do not deploy a classroom model to make real decisions about people.

Did you know? AI for Science can search patterns or propose candidates, yet experiments and domain reasoning still decide whether a candidate represents the physical world. Chapter 7 shows why constraints, interpretability and active learning can reduce wasted exploration without converting a prediction into evidence by itself.

Capabilities include mathematics, optimisation, statistics, computer science, software engineering, domain science, human factors, cybersecurity and ethics. Students can start with probability, algorithms, linear algebra, experiments and clear documentation. Current programmes set entry requirements. A strong AI portfolio explains data, baselines, compute, failures and the human decision boundary instead of promising a career or universal intelligence.

Contents · Previous chapter · Next chapter

AGI research investigates principles and systems intended to solve diverse unseen problems. Frontier AI research studies the foundations and limits of learning systems before dependable applications are assumed. CFAR lists AI for Science, artificial general intelligence, resilient and safe AI, sustainable AI, and theory and optimisation as its current pillars. Chapter 8 turns one frontier question into a testable claim with baselines, uncertainty and clearly stated failure conditions.

Use precise operational definitions and falsifiable evaluations. A rigorous AI workflow defines the task, population or system, data provenance, representation, model, compute, baseline, split, metric and deployment context. Researchers check leakage, distribution shift, calibration, efficiency and reproducibility, then compare against simpler methods. Scientific applications also require consistency with domain knowledge and experimental validation rather than treating generated outputs as discoveries.

The research label does not establish consciousness, agency or human-level generality. Capability labels are not proof of general intelligence, understanding, truth or safety. Benchmark gains may be narrow; robust performance can fail under new attacks or environments; and a model cannot own ethical or scientific responsibility. High-stakes use needs sector evidence, human oversight, security, governance and a named decision owner with authority to stop or roll back the system.

Use “Research artificial general intelligence” with synthetic data or a small public teaching dataset. Draft an evaluation boundary. Write a data sheet and model card, establish a simple baseline, reserve a test set and probe one shift or failure mode. Report compute and uncertainty. Never upload confidential or personal data, and do not deploy a classroom model to make real decisions about people.

Did you know? AI for Science can search patterns or propose candidates, yet experiments and domain reasoning still decide whether a candidate represents the physical world. Chapter 8 shows why constraints, interpretability and active learning can reduce wasted exploration without converting a prediction into evidence by itself.

Capabilities include mathematics, optimisation, statistics, computer science, software engineering, domain science, human factors, cybersecurity and ethics. Students can start with probability, algorithms, linear algebra, experiments and clear documentation. Current programmes set entry requirements. A strong AI portfolio explains data, baselines, compute, failures and the human decision boundary instead of promising a career or universal intelligence.

Contents · Previous chapter · Next chapter

AI for Science supports modelling, candidate generation and knowledge discovery across domains. Frontier AI research studies the foundations and limits of learning systems before dependable applications are assumed. CFAR lists AI for Science, artificial general intelligence, resilient and safe AI, sustainable AI, and theory and optimisation as its current pillars. Chapter 9 turns one frontier question into a testable claim with baselines, uncertainty and clearly stated failure conditions.

Join domain constraints, uncertainty and experimental confirmation. A rigorous AI workflow defines the task, population or system, data provenance, representation, model, compute, baseline, split, metric and deployment context. Researchers check leakage, distribution shift, calibration, efficiency and reproducibility, then compare against simpler methods. Scientific applications also require consistency with domain knowledge and experimental validation rather than treating generated outputs as discoveries.

A generated molecule, material or hypothesis is not a discovery until validated. Capability labels are not proof of general intelligence, understanding, truth or safety. Benchmark gains may be narrow; robust performance can fail under new attacks or environments; and a model cannot own ethical or scientific responsibility. High-stakes use needs sector evidence, human oversight, security, governance and a named decision owner with authority to stop or roll back the system.

Use “Apply AI for Science” with synthetic data or a small public teaching dataset. Build a discovery loop. Write a data sheet and model card, establish a simple baseline, reserve a test set and probe one shift or failure mode. Report compute and uncertainty. Never upload confidential or personal data, and do not deploy a classroom model to make real decisions about people.

Did you know? AI for Science can search patterns or propose candidates, yet experiments and domain reasoning still decide whether a candidate represents the physical world. Chapter 9 shows why constraints, interpretability and active learning can reduce wasted exploration without converting a prediction into evidence by itself.

Capabilities include mathematics, optimisation, statistics, computer science, software engineering, domain science, human factors, cybersecurity and ethics. Students can start with probability, algorithms, linear algebra, experiments and clear documentation. Current programmes set entry requirements. A strong AI portfolio explains data, baselines, compute, failures and the human decision boundary instead of promising a career or universal intelligence.

Contents · Previous chapter · Next chapter

Resilience concerns performance under noise, shift, faults, attacks and changing resources. Frontier AI research studies the foundations and limits of learning systems before dependable applications are assumed. CFAR lists AI for Science, artificial general intelligence, resilient and safe AI, sustainable AI, and theory and optimisation as its current pillars. Chapter 10 turns one frontier question into a testable claim with baselines, uncertainty and clearly stated failure conditions.

Stress-test corruptions, recovery, monitoring and fallback. A rigorous AI workflow defines the task, population or system, data provenance, representation, model, compute, baseline, split, metric and deployment context. Researchers check leakage, distribution shift, calibration, efficiency and reproducibility, then compare against simpler methods. Scientific applications also require consistency with domain knowledge and experimental validation rather than treating generated outputs as discoveries.

Robustness to one threat does not imply robustness to all. Capability labels are not proof of general intelligence, understanding, truth or safety. Benchmark gains may be narrow; robust performance can fail under new attacks or environments; and a model cannot own ethical or scientific responsibility. High-stakes use needs sector evidence, human oversight, security, governance and a named decision owner with authority to stop or roll back the system.

Use “Engineer resilient AI” with synthetic data or a small public teaching dataset. Create a resilience matrix. Write a data sheet and model card, establish a simple baseline, reserve a test set and probe one shift or failure mode. Report compute and uncertainty. Never upload confidential or personal data, and do not deploy a classroom model to make real decisions about people.

Did you know? AI for Science can search patterns or propose candidates, yet experiments and domain reasoning still decide whether a candidate represents the physical world. Chapter 10 shows why constraints, interpretability and active learning can reduce wasted exploration without converting a prediction into evidence by itself.

Capabilities include mathematics, optimisation, statistics, computer science, software engineering, domain science, human factors, cybersecurity and ethics. Students can start with probability, algorithms, linear algebra, experiments and clear documentation. Current programmes set entry requirements. A strong AI portfolio explains data, baselines, compute, failures and the human decision boundary instead of promising a career or universal intelligence.

Contents · Previous chapter · Next chapter

Safe-AI research studies methods and systems intended to reduce harmful or uncontrolled behaviour. Frontier AI research studies the foundations and limits of learning systems before dependable applications are assumed. CFAR lists AI for Science, artificial general intelligence, resilient and safe AI, sustainable AI, and theory and optimisation as its current pillars. Chapter 11 turns one frontier question into a testable claim with baselines, uncertainty and clearly stated failure conditions.

Define hazards, stakeholders, controls, assurance and incident response. A rigorous AI workflow defines the task, population or system, data provenance, representation, model, compute, baseline, split, metric and deployment context. Researchers check leakage, distribution shift, calibration, efficiency and reproducibility, then compare against simpler methods. Scientific applications also require consistency with domain knowledge and experimental validation rather than treating generated outputs as discoveries.

A safety score cannot remove context or organisational responsibility. Capability labels are not proof of general intelligence, understanding, truth or safety. Benchmark gains may be narrow; robust performance can fail under new attacks or environments; and a model cannot own ethical or scientific responsibility. High-stakes use needs sector evidence, human oversight, security, governance and a named decision owner with authority to stop or roll back the system.

Use “Advance safe AI” with synthetic data or a small public teaching dataset. Build an assurance case. Write a data sheet and model card, establish a simple baseline, reserve a test set and probe one shift or failure mode. Report compute and uncertainty. Never upload confidential or personal data, and do not deploy a classroom model to make real decisions about people.

Did you know? AI for Science can search patterns or propose candidates, yet experiments and domain reasoning still decide whether a candidate represents the physical world. Chapter 11 shows why constraints, interpretability and active learning can reduce wasted exploration without converting a prediction into evidence by itself.

Capabilities include mathematics, optimisation, statistics, computer science, software engineering, domain science, human factors, cybersecurity and ethics. Students can start with probability, algorithms, linear algebra, experiments and clear documentation. Current programmes set entry requirements. A strong AI portfolio explains data, baselines, compute, failures and the human decision boundary instead of promising a career or universal intelligence.

Contents · Previous chapter · Next chapter

Sustainable AI examines compute, energy, hardware, data and lifecycle impacts alongside value. Frontier AI research studies the foundations and limits of learning systems before dependable applications are assumed. CFAR lists AI for Science, artificial general intelligence, resilient and safe AI, sustainable AI, and theory and optimisation as its current pillars. Chapter 12 turns one frontier question into a testable claim with baselines, uncertainty and clearly stated failure conditions.

Report training and inference resources under a clear boundary. A rigorous AI workflow defines the task, population or system, data provenance, representation, model, compute, baseline, split, metric and deployment context. Researchers check leakage, distribution shift, calibration, efficiency and reproducibility, then compare against simpler methods. Scientific applications also require consistency with domain knowledge and experimental validation rather than treating generated outputs as discoveries.

Efficiency per task can be offset by scale or rebound effects. Capability labels are not proof of general intelligence, understanding, truth or safety. Benchmark gains may be narrow; robust performance can fail under new attacks or environments; and a model cannot own ethical or scientific responsibility. High-stakes use needs sector evidence, human oversight, security, governance and a named decision owner with authority to stop or roll back the system.

Use “Develop sustainable AI” with synthetic data or a small public teaching dataset. Create a compute-impact card. Write a data sheet and model card, establish a simple baseline, reserve a test set and probe one shift or failure mode. Report compute and uncertainty. Never upload confidential or personal data, and do not deploy a classroom model to make real decisions about people.

Did you know? AI for Science can search patterns or propose candidates, yet experiments and domain reasoning still decide whether a candidate represents the physical world. Chapter 12 shows why constraints, interpretability and active learning can reduce wasted exploration without converting a prediction into evidence by itself.

Capabilities include mathematics, optimisation, statistics, computer science, software engineering, domain science, human factors, cybersecurity and ethics. Students can start with probability, algorithms, linear algebra, experiments and clear documentation. Current programmes set entry requirements. A strong AI portfolio explains data, baselines, compute, failures and the human decision boundary instead of promising a career or universal intelligence.

Contents · Previous chapter · Next chapter

Efficient AI seeks useful performance with less memory, time, energy or data. Frontier AI research studies the foundations and limits of learning systems before dependable applications are assumed. CFAR lists AI for Science, artificial general intelligence, resilient and safe AI, sustainable AI, and theory and optimisation as its current pillars. Chapter 13 turns one frontier question into a testable claim with baselines, uncertainty and clearly stated failure conditions.

Compare accuracy-resource trade-offs on relevant hardware. A rigorous AI workflow defines the task, population or system, data provenance, representation, model, compute, baseline, split, metric and deployment context. Researchers check leakage, distribution shift, calibration, efficiency and reproducibility, then compare against simpler methods. Scientific applications also require consistency with domain knowledge and experimental validation rather than treating generated outputs as discoveries.

Model compression may change robustness, calibration or accessibility. Capability labels are not proof of general intelligence, understanding, truth or safety. Benchmark gains may be narrow; robust performance can fail under new attacks or environments; and a model cannot own ethical or scientific responsibility. High-stakes use needs sector evidence, human oversight, security, governance and a named decision owner with authority to stop or roll back the system.

Use “Improve compute efficiency” with synthetic data or a small public teaching dataset. Draw a Pareto frontier. Write a data sheet and model card, establish a simple baseline, reserve a test set and probe one shift or failure mode. Report compute and uncertainty. Never upload confidential or personal data, and do not deploy a classroom model to make real decisions about people.

Did you know? AI for Science can search patterns or propose candidates, yet experiments and domain reasoning still decide whether a candidate represents the physical world. Chapter 13 shows why constraints, interpretability and active learning can reduce wasted exploration without converting a prediction into evidence by itself.

Capabilities include mathematics, optimisation, statistics, computer science, software engineering, domain science, human factors, cybersecurity and ethics. Students can start with probability, algorithms, linear algebra, experiments and clear documentation. Current programmes set entry requirements. A strong AI portfolio explains data, baselines, compute, failures and the human decision boundary instead of promising a career or universal intelligence.

Contents · Previous chapter · Next chapter

Evaluation should cover baselines, validity, subgroups, shift, calibration, efficiency and task value. Frontier AI research studies the foundations and limits of learning systems before dependable applications are assumed. CFAR lists AI for Science, artificial general intelligence, resilient and safe AI, sustainable AI, and theory and optimisation as its current pillars. Chapter 14 turns one frontier question into a testable claim with baselines, uncertainty and clearly stated failure conditions.

Predefine metrics and protect the test set. A rigorous AI workflow defines the task, population or system, data provenance, representation, model, compute, baseline, split, metric and deployment context. Researchers check leakage, distribution shift, calibration, efficiency and reproducibility, then compare against simpler methods. Scientific applications also require consistency with domain knowledge and experimental validation rather than treating generated outputs as discoveries.

Repeated tuning on a benchmark turns it into training evidence. Capability labels are not proof of general intelligence, understanding, truth or safety. Benchmark gains may be narrow; robust performance can fail under new attacks or environments; and a model cannot own ethical or scientific responsibility. High-stakes use needs sector evidence, human oversight, security, governance and a named decision owner with authority to stop or roll back the system.

Use “Evaluate beyond one benchmark” with synthetic data or a small public teaching dataset. Build an evaluation suite. Write a data sheet and model card, establish a simple baseline, reserve a test set and probe one shift or failure mode. Report compute and uncertainty. Never upload confidential or personal data, and do not deploy a classroom model to make real decisions about people.

Did you know? AI for Science can search patterns or propose candidates, yet experiments and domain reasoning still decide whether a candidate represents the physical world. Chapter 14 shows why constraints, interpretability and active learning can reduce wasted exploration without converting a prediction into evidence by itself.

Capabilities include mathematics, optimisation, statistics, computer science, software engineering, domain science, human factors, cybersecurity and ethics. Students can start with probability, algorithms, linear algebra, experiments and clear documentation. Current programmes set entry requirements. A strong AI portfolio explains data, baselines, compute, failures and the human decision boundary instead of promising a career or universal intelligence.

Contents · Previous chapter · Next chapter

Uncertainty and human oversight help decide when a system should defer, ask or stop. Frontier AI research studies the foundations and limits of learning systems before dependable applications are assumed. CFAR lists AI for Science, artificial general intelligence, resilient and safe AI, sustainable AI, and theory and optimisation as its current pillars. Chapter 15 turns one frontier question into a testable claim with baselines, uncertainty and clearly stated failure conditions.

Test calibration, selective prediction and reviewer workload. A rigorous AI workflow defines the task, population or system, data provenance, representation, model, compute, baseline, split, metric and deployment context. Researchers check leakage, distribution shift, calibration, efficiency and reproducibility, then compare against simpler methods. Scientific applications also require consistency with domain knowledge and experimental validation rather than treating generated outputs as discoveries.

A confidence number is not guaranteed knowledge or safe delegation. Capability labels are not proof of general intelligence, understanding, truth or safety. Benchmark gains may be narrow; robust performance can fail under new attacks or environments; and a model cannot own ethical or scientific responsibility. High-stakes use needs sector evidence, human oversight, security, governance and a named decision owner with authority to stop or roll back the system.

Use “Quantify uncertainty and oversight” with synthetic data or a small public teaching dataset. Design a deferral rule. Write a data sheet and model card, establish a simple baseline, reserve a test set and probe one shift or failure mode. Report compute and uncertainty. Never upload confidential or personal data, and do not deploy a classroom model to make real decisions about people.

Did you know? AI for Science can search patterns or propose candidates, yet experiments and domain reasoning still decide whether a candidate represents the physical world. Chapter 15 shows why constraints, interpretability and active learning can reduce wasted exploration without converting a prediction into evidence by itself.

Capabilities include mathematics, optimisation, statistics, computer science, software engineering, domain science, human factors, cybersecurity and ethics. Students can start with probability, algorithms, linear algebra, experiments and clear documentation. Current programmes set entry requirements. A strong AI portfolio explains data, baselines, compute, failures and the human decision boundary instead of promising a career or universal intelligence.

Contents · Previous chapter · Next chapter

Reproducibility and governance cover code, data, compute, access, documentation and accountability. Frontier AI research studies the foundations and limits of learning systems before dependable applications are assumed. CFAR lists AI for Science, artificial general intelligence, resilient and safe AI, sustainable AI, and theory and optimisation as its current pillars. Chapter 16 turns one frontier question into a testable claim with baselines, uncertainty and clearly stated failure conditions.

Use versioned artefacts, independent checks and clear use limits. A rigorous AI workflow defines the task, population or system, data provenance, representation, model, compute, baseline, split, metric and deployment context. Researchers check leakage, distribution shift, calibration, efficiency and reproducibility, then compare against simpler methods. Scientific applications also require consistency with domain knowledge and experimental validation rather than treating generated outputs as discoveries.

Open code alone does not make a model correct or responsible. Capability labels are not proof of general intelligence, understanding, truth or safety. Benchmark gains may be narrow; robust performance can fail under new attacks or environments; and a model cannot own ethical or scientific responsibility. High-stakes use needs sector evidence, human oversight, security, governance and a named decision owner with authority to stop or roll back the system.

Use “Reproduce and govern research” with synthetic data or a small public teaching dataset. Create a research checklist. Write a data sheet and model card, establish a simple baseline, reserve a test set and probe one shift or failure mode. Report compute and uncertainty. Never upload confidential or personal data, and do not deploy a classroom model to make real decisions about people.

Did you know? AI for Science can search patterns or propose candidates, yet experiments and domain reasoning still decide whether a candidate represents the physical world. Chapter 16 shows why constraints, interpretability and active learning can reduce wasted exploration without converting a prediction into evidence by itself.

Capabilities include mathematics, optimisation, statistics, computer science, software engineering, domain science, human factors, cybersecurity and ethics. Students can start with probability, algorithms, linear algebra, experiments and clear documentation. Current programmes set entry requirements. A strong AI portfolio explains data, baselines, compute, failures and the human decision boundary instead of promising a career or universal intelligence.

Contents · Previous chapter · Next chapter

The receipt links task, data, model, compute, baselines, stress tests, uncertainty and owner. Frontier AI research studies the foundations and limits of learning systems before dependable applications are assumed. CFAR lists AI for Science, artificial general intelligence, resilient and safe AI, sustainable AI, and theory and optimisation as its current pillars. Chapter 17 turns one frontier question into a testable claim with baselines, uncertainty and clearly stated failure conditions.

Archive versions, licences, seeds, failures, energy boundary and review decisions. A rigorous AI workflow defines the task, population or system, data provenance, representation, model, compute, baseline, split, metric and deployment context. Researchers check leakage, distribution shift, calibration, efficiency and reproducibility, then compare against simpler methods. Scientific applications also require consistency with domain knowledge and experimental validation rather than treating generated outputs as discoveries.

Documentation cannot turn an experimental model into an authorised decision-maker. Capability labels are not proof of general intelligence, understanding, truth or safety. Benchmark gains may be narrow; robust performance can fail under new attacks or environments; and a model cannot own ethical or scientific responsibility. High-stakes use needs sector evidence, human oversight, security, governance and a named decision owner with authority to stop or roll back the system.

Use “Keep a frontier-AI evidence receipt” with synthetic data or a small public teaching dataset. Assemble a one-page receipt. Write a data sheet and model card, establish a simple baseline, reserve a test set and probe one shift or failure mode. Report compute and uncertainty. Never upload confidential or personal data, and do not deploy a classroom model to make real decisions about people.

Did you know? AI for Science can search patterns or propose candidates, yet experiments and domain reasoning still decide whether a candidate represents the physical world. Chapter 17 shows why constraints, interpretability and active learning can reduce wasted exploration without converting a prediction into evidence by itself.

Capabilities include mathematics, optimisation, statistics, computer science, software engineering, domain science, human factors, cybersecurity and ethics. Students can start with probability, algorithms, linear algebra, experiments and clear documentation. Current programmes set entry requirements. A strong AI portfolio explains data, baselines, compute, failures and the human decision boundary instead of promising a career or universal intelligence.

Contents · Previous chapter · Continue in the Science Learning Hub