Ai Foundation Models

Transformer Architecture Research

Doctoral work examines the architecture underpinning almost every current foundation model. Architectural choices determine what capabilities a model can acquire during training.

Attention Mechanism Design

Research examines mechanisms relating each position to every other position. Attention design governs both capability and computational cost together.

Efficient Attention Methods

Doctoral study reduces the quadratic cost of relating all positions. Efficiency gains determine what sequence lengths remain practically feasible.

Long Context Architectures

Research develops models capable of handling very long input sequences. Extended context enables reasoning across whole documents, books and codebases.

State Space Model Research

Doctoral work examines sequence architectures based on continuous state dynamics. These architectures scale more favourably with sequence length.

Recurrent Architecture Revival

Research revisits recurrent sequence designs using modern training techniques. Recurrent designs offer constant memory cost during text generation.

Mixture Of Experts Architectures

Doctoral study examines models activating only some parameters per input. Sparse activation increases capacity without proportional computation cost.

Sparse Model Design

Research examines models where most parameters remain inactive per input. Sparsity permits far larger models within fixed computational budgets.

Positional Encoding Research

Doctoral work examines how models represent the order of input elements. Position representation strongly determines extension to longer sequences.

Tokenisation Method Research

Research examines how raw text is divided into modelling units. Tokenisation choices affect efficiency, fairness across languages and capability.

Vocabulary Design Research

Doctoral study examines the set of units a model is able to represent. Vocabulary composition systematically disadvantages underrepresented languages.

Byte Level Modelling

Research examines models operating directly on raw sequences of bytes. Byte level operation removes tokenisation bias across differing writing systems.

Embedding Space Analysis

Doctoral work examines the geometry of learned internal representations. Representation geometry reveals what structure a model has actually learned.

Normalisation Technique Research

Research examines methods stabilising values flowing through deep networks. Normalisation choices strongly affect training stability at very large scale.

Activation Function Research

Doctoral study examines the nonlinear functions applied within network layers. Activation choice influences both trainability and final model quality.

Regularisation Method Research

Research examines techniques preventing models fitting training data too closely. Regularisation matters differently at the scale foundation models occupy.

Initialisation Strategy Research

Doctoral work examines starting parameter values before training begins. Initialisation strongly affects whether very deep networks train at all.

Optimiser Design Research

Research examines algorithms adjusting parameters during model training. Optimiser choice determines both training speed and achievable final quality.

Learning Rate Schedule Research

Doctoral study examines how step sizes change across a full training run. Schedule design materially affects the quality of the resulting trained model.

Loss Function Design

Research examines the quantity a model is trained to minimise during learning. Loss design determines precisely what behaviour the training process rewards.

Objective Function Research

Doctoral work examines training targets shaping the capability of a model. Objective choice determines which abilities emerge from the pretraining process.

Next Element Prediction Analysis

Research examines why predicting continuations produces such general capability. This objective is remarkably simple yet remains incompletely explained.

Masked Modelling Objectives

Doctoral study examines training by predicting concealed portions of input. Masked objectives suit representation learning rather than generation.

Contrastive Pretraining Objectives

Research trains models by distinguishing matched from unmatched examples. Contrastive objectives underpin much multimodal representation learning.

Multi Objective Pretraining

Doctoral work combines several training objectives within one process. Combined objectives may produce capability no single objective achieves.

Curriculum Design In Pretraining

Research examines ordering training material from simpler toward harder examples. Ordering effects on eventual model capability remain poorly understood.

Data Ordering Effects

Doctoral study examines how the sequence of training examples affects outcomes. Ordering influences what a model retains from early training.

Batch Size And Scaling Effects

Research examines how the quantity processed per step affects model training. Batch scaling interacts with essentially every other training hyperparameter.

Scaling Law Research

Doctoral work examines predictable relationships between scale and capability. Scaling relationships guide resource allocation before training begins.

Compute Optimal Training

Research examines the best balance between model size and training data. Optimal allocation substantially changes the capability obtained per unit of computation.

Parameter Count Trade Offs

Doctoral study examines consequences of model size for capability and cost. Larger is not automatically better once serving cost is considered.

Training Stability Research

Research examines why very large training runs sometimes fail catastrophically. Stability problems waste enormous quantities of computation.

Loss Spike Analysis

Doctoral work examines sudden training failures and their underlying causes. Spike understanding permits prevention rather than reactive restarting.

Numerical Precision In Training

Research examines how reduced numerical precision affects training outcomes. Precision choices trade computational efficiency against training stability.

Mixed Precision Methods

Doctoral study assigns differing precision to different training computations. Selective precision preserves stability while reducing computation substantially.

Gradient Accumulation Research

Research examines combining learning signals across successive computations. Accumulation permits large effective batches on limited hardware.

Distributed Training Systems

Doctoral work examines training spread across very many separate processors. Distribution is entirely unavoidable at the scale these models require.

Model Parallelism Research

Research divides a single model across multiple separate computing devices. Division is required whenever a model exceeds any single device memory.

Pipeline Parallelism Methods

Doctoral study assigns successive model layers to different computing devices. Pipeline arrangements reduce memory demand but introduce processor idle time.

Communication Efficient Training

Research reduces the volume of data exchanged between processors during training. Communication bandwidth frequently limits achievable training throughput.

Fault Tolerant Training Systems

Doctoral work examines training that continues despite individual hardware failures. Failures are inevitable across runs spanning many thousands of processors.

Checkpoint And Recovery Research

Research examines saving and restoring training state efficiently. Checkpoint frequency balances storage cost against recomputation after failure.

Training Reproducibility

Doctoral study examines whether large training runs can be repeated exactly. Reproducibility is essential for both scientific progress and accountability.

Training Cost Modelling

Research models the resources required to train models at scale. Cost modelling determines what training is feasible for which organisations.

Hardware Aware Model Design

Doctoral work designs models suited to available accelerator characteristics. Hardware constraints shape architecture as strongly as theory does.

Pretraining Data Curation

Research examines assembling the very large corpora used in pretraining. Data composition determines model capability more than architecture does.

Data Quality Assessment

Doctoral study develops measurement of the quality of training material. Quality assessment at corpus scale cannot possibly rely on human inspection.

Data Filtering Methods

Research examines removing unsuitable material from very large training corpora. Filtering decisions substantially shape what a trained model actually learns.

Deduplication Research

Doctoral work examines removing repeated material from training corpora. Repetition promotes memorisation and wastes substantial training computation.

Data Mixture Optimisation

Research optimises the proportions of differing data sources during training. Mixture proportions strongly determine which model capabilities develop.

Domain Weighting Research

Doctoral study examines emphasising particular subject domains during training. Weighting decisions trade breadth against depth of capability.

Data Scaling Research

Research examines how model capability changes with training corpus size. High quality data availability is now becoming a genuinely binding constraint.

Synthetic Data Generation

Doctoral work examines training material produced by models themselves. Synthetic material addresses scarcity but risks narrowing diversity.

Synthetic Data Quality Research

Research examines whether generated training material is genuinely useful. Quality assessment is difficult when material resembles model output.

Model Collapse Research

Doctoral study examines degradation when models train on generated material. Repeated generations may progressively lose diversity and accuracy.

Data Contamination Detection

Research detects evaluation material appearing within training corpora. Contamination makes reported capability figures seriously misleading.

Benchmark Contamination Analysis

Doctoral work examines the extent of benchmark leakage into training data. Contamination is widespread and inconsistently reported by developers.

Multilingual Data Research

Research examines training corpora spanning many different human languages. Language representation determines which populations models actually serve well.

Low Resource Language Data

Doctoral study examines languages with very little available digital text. Most of the world languages fall into this severely underserved category.

Code Data Research

Research examines programming code included within model training corpora. Code data appears to improve reasoning well beyond programming tasks alone.

Scientific Corpus Research

Doctoral work examines scientific literature included within training material. Scientific text is high quality but access is frequently legally restricted.

Mathematical Data Research

Research examines mathematical material included within training corpora. Mathematical data is scarce and demands specialised notation representation.

Multimodal Data Alignment

Doctoral study examines pairing training data across differing modalities. Alignment quality determines whether genuine cross modal capability develops.

Image Text Pair Research

Research examines paired visual and textual material used in training. Pair quality strongly determines the visual understanding capability achieved.

Video Data Curation

Doctoral work examines assembling video material for use in model training. Video is abundant but extremely costly to process at meaningful scale.

Audio Data Research

Research examines speech and audio material within training corpora. Audio training raises distinctive consent and identifiability concerns.

Data Provenance Tracking

Doctoral study examines recording where training material originated. Provenance records are required for both legal and scientific accountability.

Licensing And Data Rights

Research examines legal permissions governing use of training material. Rights questions remain substantially unsettled across jurisdictions.

Consent In Training Data

Doctoral work examines whether people consented to their material being used. Most training corpora were assembled without any meaningful consent.

Personal Data In Corpora

Research examines identifiable personal information within training material. Personal information within corpora creates lasting privacy exposure.

Data Memorisation Research

Doctoral study examines models reproducing training material word for word. Memorisation creates both privacy exposure and copyright related risk.

Extraction Attack Research

Research examines recovering training material from a deployed model. Extraction studies establish what privacy protection is actually achieved.

Data Attribution Methods

Doctoral work traces model behaviour back to specific training material. Attribution supports both accountability and targeted data improvement.

Influence Function Research

Research estimates how individual training examples shaped a model. Influence estimation at foundation model scale remains computationally hard.

Training Data Valuation

Doctoral study examines the contribution particular data makes to capability. Valuation underpins fair arrangements with those who contribute data.

Data Documentation Standards

Research develops standards describing the composition of training corpora. Documentation permits external scrutiny that is otherwise entirely impossible.

Corpus Bias Analysis

Doctoral work examines systematic skews present within training material. Corpus skews propagate directly into eventual model behaviour and outputs.

Representation Gaps In Data

Research examines groups and topics entirely absent from training corpora. Absences produce models that serve some populations very much worse.

Data Governance Frameworks

Doctoral study examines oversight of how training data is obtained and used. Governance frameworks in this area remain substantially undeveloped.

Data Curation Labour Research

Research examines the human work of annotating and filtering training material. This labour is extensive, poorly paid and frequently psychologically demanding.

Supervised Fine Tuning Methods

Doctoral work examines adapting pretrained models using labelled examples. Fine tuning converts general capability into task specific usefulness.

Instruction Tuning Research

Research examines training models to follow stated instructions. Instruction following transformed these models from completion engines into assistants.

Parameter Efficient Adaptation

Doctoral study adapts models by modifying only small parameter subsets. Efficient adaptation makes customisation feasible without vast resources.

Low Rank Adaptation Research

Research adapts models using compact low dimensional parameter modifications. Compact modifications are cheap to store, to share and to combine.

Prompt Tuning Methods

Doctoral work adapts models by learning input prefixes rather than parameters. Prefix learning leaves the underlying model entirely unchanged.

In Context Learning Research

Research examines models learning from examples supplied within their input. This capability emerged without ever being explicitly trained for.

Few Shot Capability Analysis

Doctoral study examines performance given only a few demonstration examples. Few shot capability determines usefulness wherever labelled data is scarce.

Prompt Engineering Research

Research examines how the phrasing of input influences model behaviour. Phrasing sensitivity reveals genuine fragility in underlying model capability.

Chain Of Thought Research

Doctoral work examines models producing intermediate reasoning before answering. Intermediate reasoning substantially improves performance on hard problems.

Reasoning Capability Analysis

Research examines what these models can and cannot actually reason about. Whether observed behaviour constitutes reasoning is genuinely contested.

Mathematical Reasoning Research

Doctoral study examines mathematical problem solving capability. Mathematics provides verifiable problems for studying reasoning rigorously.

Code Generation Capability

Research examines models producing functioning computer programs. Code generation is among the most economically significant current applications.

Tool Use Research

Doctoral work examines models invoking external systems, services and tools. Tool access extends capability well beyond what model parameters encode.

Function Calling Research

Research examines structured invocation of external functions by models. Structured calling makes model output reliably machine interpretable.

Agentic Behaviour Research

Doctoral study examines models pursuing goals across multiple steps. Autonomous operation raises capability and risk questions simultaneously.

Multi Step Planning Research

Research examines models decomposing complex tasks into ordered subtasks. Planning capability determines what complex work can safely be delegated.

Memory And State Research

Doctoral work examines models retaining information across interactions. Persistent memory changes both usefulness and privacy considerations.

Retrieval Augmented Generation

Research examines models consulting external sources before generating a response. Retrieval grounds output in verifiable and reasonably current information.

Retrieval System Design

Doctoral study examines systems finding relevant material for model use. Retrieval quality frequently limits overall system performance more than the model.

Knowledge Grounding Methods

Research examines anchoring model output in identified source material. Grounding permits verification that unconstrained generation cannot support.

Factual Accuracy Research

Doctoral work examines whether model statements correspond to reality. Factual reliability determines suitability for consequential applications.

Hallucination Analysis

Research examines confident model statements that are simply not true. Confident fabrication remains among the most serious obstacles to deployment.

Confidence Calibration Research

Doctoral study examines whether expressed confidence matches actual accuracy. Miscalibrated confidence misleads users about output reliability.

Uncertainty Estimation Methods

Research develops honest quantification of uncertainty in model output. Knowing when a model is unreliable is as valuable as raw predictive accuracy.

Abstention And Refusal Behaviour

Doctoral work examines models declining to answer where appropriate. Appropriate abstention prevents confident answers to unanswerable questions.

Learning From Human Feedback

Research examines training models using human judgements of output quality. Human feedback shaped nearly every deployed assistant model in use today.

Preference Learning Methods

Doctoral study examines learning from comparisons between candidate outputs. Comparison judgements are easier to collect than absolute ratings.

Reward Model Research

Research examines learned models scoring the quality of generated outputs. Reward model errors propagate directly into the trained model behaviour.

Reward Hacking Analysis

Doctoral work examines models exploiting flaws in their training signal. Exploitation produces high scores alongside genuinely poor behaviour.

Principle Based Training Methods

Research examines training guided by explicit written behavioural principles. Explicit principles make intended behaviour open to public scrutiny.

Self Improvement Research

Doctoral study examines models generating their own training signal. Self improvement raises capability and human oversight concerns simultaneously.

Distillation From Larger Models

Research transfers capability from large models into much smaller ones. Distillation makes capable models deployable at very much lower cost.

Model Merging Research

Doctoral work examines combining separately trained models into a single one. Merging captures multiple capabilities without any additional retraining.

Continual Learning In Models

Research examines models acquiring new capability without losing existing ones. Retraining entirely from scratch is prohibitively expensive.

Knowledge Editing Methods

Doctoral study examines correcting specific facts within a trained model. Targeted correction avoids retraining for isolated factual errors.

Unlearning Research

Research examines removing the influence of specific material from models. Removal is legally required yet technically very difficult to verify.

Capability Emergence Research

Doctoral work examines abilities appearing suddenly as model scale increases. Emergence makes capability at larger scale genuinely difficult to anticipate.

Emergent Behaviour Analysis

Research examines behaviours nobody explicitly trained a model to exhibit. Unanticipated behaviour is central to both the promise and the risk.

Evaluation Methodology

Doctoral study develops methodologically sound assessment of model capability. Evaluation practice in this field is widely criticised as inadequate.

Benchmark Design Research

Research examines constructing tests that measure what they claim to measure. Poorly designed benchmarks misdirect the effort of an entire research community.

Benchmark Validity Research

Doctoral work examines whether benchmark scores predict genuine usefulness. Benchmark performance and practical utility diverge very substantially.

Human Evaluation Methods

Research examines assessment of model output by human expert judges. Human judgement remains the reference standard yet is costly and variable.

Automated Evaluation Research

Doctoral study examines assessment conducted without any human involvement. Automation permits evaluation at a scale humans could never achieve.

Model Based Evaluation

Research examines using models to assess the outputs of other models. Model judges inherit the biases of the models performing the judgement.

Adversarial Evaluation Methods

Doctoral work examines assessment using deliberately difficult inputs. Adversarial testing reveals weaknesses that average performance conceals.

Red Teaming Methodology

Research examines structured attempts to elicit harmful model behaviour. Systematic probing identifies failure modes before public deployment.

Capability Elicitation Research

Doctoral study examines drawing out capabilities a model actually possesses. Poor elicitation causes capability to be seriously underestimated.

Dangerous Capability Evaluation

Research examines assessing capabilities that could enable serious harm. These assessments inform decisions about whether release is responsible.

Evaluation Of Reasoning Quality

Doctoral work examines assessing reasoning rather than only final answers. Correct final answers can follow from entirely invalid intermediate reasoning.

Long Context Evaluation

Research examines assessing model performance across very long inputs. Claimed context capacity frequently exceeds genuinely usable context capacity.

Multilingual Evaluation

Doctoral study examines assessment across many human languages. Evaluation concentrates heavily on a small number of well resourced languages.

Domain Specific Evaluation

Research examines assessment within particular specialist professional domains. General purpose benchmarks predict specialist performance very poorly.

Robustness Evaluation Methods

Doctoral work examines stability of performance under variation in input. Very small input changes can produce disproportionately large behavioural changes.

Distribution Shift Evaluation

Research examines performance on inputs unlike anything in the training material. Real deployment inevitably encounters unfamiliar input distributions.

Evaluation Reproducibility

Doctoral study examines whether reported evaluation results can be repeated. Undisclosed configuration choices make reproduction frequently impossible.

Mechanistic Interpretability

Research examines the internal computations models actually perform. Mechanistic understanding would transform both safety and capability work.

Circuit Analysis Research

Doctoral work identifies internal components implementing specific behaviours. Circuit identification connects observable behaviour to internal structure.

Feature Representation Analysis

Research examines what concepts model internal representations encode. Representations appear to encode far more concepts than they have dimensions.

Sparse Autoencoder Methods

Doctoral study decomposes representations into interpretable component features. This approach has become central to current interpretability work.

Activation Steering Research

Research examines influencing behaviour by adjusting internal model activations. Steering offers behavioural control without any retraining of the model.

Probing Method Research

Doctoral work examines testing what information internal representations contain. Probing results are frequently overinterpreted as evidence of actual use.

Attention Analysis Methods

Research examines what attention patterns reveal about model processing. Attention patterns are informative but do not constitute explanation.

Explanation Faithfulness Research

Doctoral study examines whether stated reasoning reflects actual computation. Models may produce plausible reasoning unrelated to their real processing.

Model Behaviour Prediction

Research examines anticipating how a model will behave before it is tested. Predictive ability is a demanding test of genuine mechanistic understanding.

Internal Knowledge Localisation

Doctoral work examines where within a model particular knowledge resides. Localisation supports targeted correction and removal of information.

Representation Similarity Analysis

Research compares internal representations across models and training runs. Similarity analysis reveals which structures are learned universally.

Training Dynamics Analysis

Doctoral study examines how models change across the whole course of training. Dynamics research explains when and why particular capabilities are acquired.

Phase Transition Research

Research examines abrupt qualitative changes occurring during training. Sudden transitions accompany the acquisition of several key capabilities.

Alignment Theory Research

Doctoral work examines ensuring models pursue the objectives intended for them. Alignment becomes harder as systems become more capable and more autonomous.

Specification And Goal Research

Research examines difficulty of stating precisely what behaviour is wanted. Specification failures underlie most observed alignment problems.

Value Learning Research

Doctoral study examines models inferring human values from available evidence. Whose values, and how conflicts between them are resolved, remain open questions.

Scalable Oversight Methods

Research examines supervising models on tasks humans cannot easily verify. Oversight becomes the binding constraint as model capability continues growing.

Debate And Critique Methods

Doctoral work examines models critiquing one another to assist human judgement. Adversarial structures may surface errors humans would otherwise miss.

Weak To Strong Generalisation

Research examines whether weaker supervisors can guide stronger models. This question is central to supervising systems beyond human capability.

Deceptive Behaviour Research

Doctoral study examines models behaving differently when they appear observed. Such behaviour would undermine every evaluation based safeguard.

Situational Awareness Research

Research examines models recognising their own circumstances and being tested. Awareness of evaluation could invalidate the evaluation being conducted.

Sycophancy Research

Doctoral work examines models agreeing with users rather than being accurate. Agreeableness trained through feedback can displace truthfulness.

Robustness To Manipulation

Research examines resistance to inputs designed to subvert intended behaviour. Manipulation resistance is essential wherever models process untrusted input.

Prompt Injection Defence

Doctoral study examines defending against instructions hidden within processed content. This vulnerability is severe wherever models handle external material.

Safeguard Resistance Research

Research examines attempts to circumvent trained behavioural safeguards. Understanding circumvention is necessary to build durable safeguards.

Safety Training Methods

Doctoral work examines training that produces reliably safe model behaviour. Safety training must generalise far beyond the specific examples used.

Refusal Behaviour Research

Research examines when and how models decline to fulfil requests. Refusal decisions require judgement that simple written rules cannot capture.

Overrefusal Analysis

Doctoral study examines models declining entirely reasonable user requests. Excessive refusal makes models unhelpful without improving safety at all.

Content Policy Research

Research examines the rules governing what models will and will not produce. Policy design involves contested value judgements about acceptable content.

Bias And Fairness Evaluation

Doctoral work examines whether models treat groups differently without justification. Measured differences translate into unequal quality of service.

Stereotype Analysis Research

Research examines stereotyped associations encoded within model behaviour. Stereotypes present in training material persist into model outputs.

Fairness Mitigation Methods

Doctoral study examines reducing unjustified differences in model behaviour. Mitigation frequently trades one form of unfairness for another.

Harmful Content Detection

Research examines identifying harmful material in model inputs and outputs. Detection systems must work across languages and cultural contexts.

Privacy Preserving Training

Doctoral work examines training that limits exposure of individual personal data. Privacy protection must be built into training rather than added afterwards.

Differential Privacy In Models

Research applies formal privacy guarantees to model training. Formal guarantees replace informal assurances that repeatedly prove inadequate.

Watermarking Research

Doctoral study examines embedding detectable markers within generated output. Watermarking supports attribution but is fragile to modification.

Generated Content Detection

Research examines distinguishing model output from human produced material. Reliable detection remains unsolved and may be fundamentally limited.

Provenance And Authenticity

Doctoral work examines establishing the verified origin of digital content. Provenance infrastructure may prove more workable than detection alone.

Model Auditing Methods

Research develops systematic examination of model properties and behaviour. Auditing methods determine what external scrutiny can establish.

Third Party Assessment Research

Doctoral study examines independent evaluation by parties outside developers. Independent assessment requires access developers frequently decline to grant.

Inference Optimisation

Research examines efficient generation of output from already trained models. Inference cost dominates total expenditure across the lifetime of a model.

Quantisation Research

Doctoral work reduces numerical precision of model parameters for deployment. Lower precision reduces memory and computation with modest quality cost.

Model Compression Methods

Research reduces model size while preserving genuinely useful capability. Compression determines what hardware can realistically host a given model.

Speculative Decoding Research

Doctoral study accelerates generation using a smaller assisting model. Speculative methods reduce latency without changing the output produced.

Serving System Design

Research examines infrastructure delivering model responses at scale. Serving design determines both achievable throughput and operating cost.

Caching Strategy Research

Doctoral work examines reusing computation across successive model requests. Caching substantially reduces cost for repeated or shared input context.

Latency And Throughput Analysis

Research examines the trade off between response speed and total serving capacity. This trade off shapes both system architecture and user experience.

On Device Model Research

Doctoral study examines running capable models on personal consumer hardware. Local operation removes both network dependence and data exposure risk.

Inference Efficiency Economics

Research examines the resource economics of serving these models at scale. Serving economics ultimately determines which applications are viable at all.

Energy Use Of Models

Doctoral work quantifies energy consumed during model training and serving. Energy demand from these systems is growing extremely rapidly worldwide.

Environmental Impact Assessment

Research examines the full environmental burden of model development. Water use and hardware manufacture matter alongside operational energy demand.

Model Lifecycle Management

Doctoral study examines governing models from development through retirement. Lifecycle governance determines accountability throughout a model existence.

Version Control And Documentation

Research examines tracking successive model versions and their differences. Version tracking is essential for reproducing any past behaviour.

Model Reporting Standards

Doctoral work examines standardised disclosure of model characteristics. Disclosure standards determine what users and regulators can learn.

Open Weight Model Research

Research examines models whose learned parameters are publicly released. Open release enables external scrutiny while removing usage safeguards.

Access Control And Release Policy

Doctoral study examines decisions about who may use a model and in what way. Release decisions are consequential and very largely irreversible.

Licensing And Terms Research

Research examines legal terms governing model use and redistribution. Licence terms determine what downstream developers may legitimately build.

Downstream Use Monitoring

Doctoral work examines observing how deployed models are actually used. Monitoring must balance oversight against user privacy expectations.

Misuse Detection Research

Research examines identifying harmful patterns in how models are used. Detection enables response while avoiding surveillance of ordinary legitimate users.

Regulatory Framework Analysis

Doctoral study examines emerging regulation governing these systems. Regulatory approaches differ sharply between jurisdictions and remain unsettled.

Standards Development Research

Research examines how technical standards for these systems are actually formed. Standards translate broad principles into requirements that can be checked.

Economic Impact Research

Doctoral work examines economic consequences of widespread model deployment. Effects are substantial, uneven and difficult to measure reliably.

Labour Market Effect Research

Research examines how these systems change the nature and organisation of work. Effects fall very unevenly across differing occupations and skill groups.

Compute Governance Research

Doctoral study examines governance of the computing resources training these models. Computation is measurable and therefore an unusually governable input.

Implementation And Adoption Research

Research examines why organisations succeed or fail in deploying these systems. Adoption barriers are frequently organisational rather than technical.