Computational algorithms trained on data to recognise patterns, make predictions, and perform tasks, including neural networks for content generation, behaviour simulation, computer vision, and natural language processing; the core artefact produced by a machine learning training pipeline.
Semantic Classification
Content
Core Neural Network Types
Convolutional Neural Networks (CNN)
-
Pattern recognition
-
Image processing
-
Object detection
-
Visual analysis
-
Feature extraction
Recurrent Neural Networks (RNN)
-
Sequential data processing
-
Time series analysis
-
Language modelling
-
Memory retention
-
State preservation
Long Short-Term Memory (LSTM)
-
Extended memory
-
Gradient problem solution
-
Complex sequences
-
Speech recognition
-
Text generation
Generative Models
GANs (Generative Adversarial Networks)
-
Generator-discriminator architecture
-
Realistic content synthesis
-
Image generation
-
Style transfer
-
Data augmentation
VAEs (Variational Autoencoders)
-
Latent space learning
-
Content reconstruction
-
Feature interpolation
-
Anomaly detection
-
Data compression
NeRF Technology
-
Neural Radiance Fields
-
3D scene generation
-
2D to 3D conversion
-
Hours to minutes modelling
-
Implicit representations
Metaverse Applications
Content Generation
-
3D asset creation
-
Environment synthesis
-
Avatar design
-
Texture generation
-
World building
Natural Language Processing
-
Chatbot intelligence
-
Virtual assistants
-
NPC dialogue
-
Language translation
-
Voice interaction
Computer Vision
-
AR object detection
-
Pose estimation
-
Scene understanding
-
Gesture recognition
-
Spatial mapping
Digital Twins
-
Physical object replication
-
City planning
-
Assembly line simulation
-
Virtual surgery
-
Process modelling
Training Approaches
Supervised Learning
-
Labelled data training
-
Classification tasks
-
Regression problems
-
Error minimisation
-
Ground truth alignment
Unsupervised Learning
-
Pattern discovery
-
Clustering
-
Dimensionality reduction
-
Anomaly detection
-
Feature learning
Reinforcement Learning
-
Reward-based training
-
Agent behaviour
-
Game AI
-
Decision optimisation
-
Environment interaction
Industry Applications
Healthcare
-
Surgical simulation
-
Diagnostic assistance
-
Treatment planning
-
Medical training
-
Patient interaction
Gaming
-
NPC intelligence
-
Procedural generation
-
Player behaviour prediction
-
Adaptive difficulty
-
Content personalisation
Security
-
Fraud detection
-
Anomaly identification
-
Cyberattack prevention
-
Transaction monitoring
-
Darktrace, Microsoft Defender
Development Tools
AI Engines
-
Claude2
-
Midjourney
-
Runway
-
Stable Diffusion
-
Llama2
Frameworks
-
TensorFlow
-
PyTorch
-
Keras
-
JAX
-
ONNX
Model Capabilities
Prediction
-
Future state estimation
-
Behaviour forecasting
-
Trend analysis
-
Risk assessment
-
Demand prediction
Classification
-
Category assignment
-
Object identification
-
Sentiment analysis
-
Content moderation
-
Intent recognition
Generation
-
Content creation
-
Image synthesis
-
Text generation
-
Music composition
-
Video production
2024 Advancements
GPT-4 Era
-
Unprecedented capabilities
-
Multimodal learning
-
Complex reasoning
-
Creative assistance
-
Task automation
Multimodal Integration
-
Visual understanding
-
Audio processing
-
Text analysis
-
Combined inputs
-
Comprehensive AI
Future Directions
Scalable AI
-
Larger models
-
Efficient training
-
Edge deployment
-
Real-time inference
-
Cost reduction
Intelligent Environments
-
Hyper-personalisation
-
Adaptive content
-
Dynamic experiences
-
Decentralised AI
-
Autonomous systems
Current Landscape (2026)
- The dominant model class shifted from plain instruction-tuned LLMs to explicit reasoning (“thinking”) models: OpenAI shipped o3/o4-mini and then GPT-5 (August 2025), iterating to GPT-5.2 (December 2025) and GPT-5.4 (March 2026), while Google released Gemini 2.5 Pro (March 2025) and Gemini 3 Pro with Deep Think (November 2025), the first model to break 1500 Elo on LMArena.
- Anthropic’s Claude 4.5 family (Sonnet, Haiku, Opus, rolled out September–November 2025) pushed real-world coding state-of-the-art, with Sonnet reaching 77.2% on SWE-bench Verified, reflecting a broader pivot toward long-horizon agentic tool-calling as the headline capability rather than raw next-token quality.
- Open-weight models became genuinely frontier-competitive and disrupted pricing: DeepSeek V3 and R1 (January 2025) matched GPT-4o/o1-class performance at roughly a tenth of the cost, followed by Meta’s Llama 4 Scout and Maverick (April 2025, the firm’s first natively multimodal mixture-of-experts models) and Alibaba’s Qwen3-235B.
- Mixture-of-experts (MoE) architectures became the mainstream design for large models (Llama 4, DeepSeek V3 at 671B total / 37B active, Qwen3, gpt-oss), decoupling total parameter count from per-token inference cost and enabling very long context windows (Gemini 2.5’s 1M tokens; Llama 4 Scout advertised up to 10M).
- OpenAI re-entered the open-weight space with gpt-oss-120b and gpt-oss-20b under an Apache 2.0 licence (August 2025), its first openly downloadable models since GPT-2, signalling that even closed-first labs now treat open weights as strategically necessary.
- Regulation moved from drafting to enforcement: EU AI Act obligations for general-purpose AI models (defined at >10^23 FLOP, with “systemic risk” models above 10^25 FLOP facing red-teaming, incident reporting and evaluation duties) applied from 2 August 2025, and the AI Office’s enforcement powers plus broad applicability landed on 2 August 2026, backed by fines up to €35m or 7% of global turnover.
- Transparency and provenance became mandated: the AI Act’s Article 50 labelling rules apply from August 2026, with machine-readable watermarking of AI-generated content required from 2 December 2026, forcing model providers to bake output-marking into their systems.
- Open challenges as of 2026 centre on benchmark saturation and contamination (MMLU/GSM8K near-ceiling, driving harder evals like Humanity’s Last Exam and SWE-bench Verified), the compute/energy cost and reliability of test-time “thinking”, copyright and training-data disclosure obligations, and agentic safety as models run autonomous multi-hour tool-calling workflows.
References
-
- European Commission — Digital Strategy (2025). EU rules on general-purpose AI models start to apply, bringing more transparency, safety and accountability. https://digital-strategy.ec.europa.eu/en/news/eu-rules-general-purpose-ai-models-start-apply-bringing-more-transparency-safety-and-accountability
-
- Euronews (2026). EU rules on AI models become enforceable. What’s going to change? https://www.euronews.com/my-europe/2026/08/02/eu-rules-on-ai-models-become-enforceable-whats-going-to-change
-
- Apolo (2026). The Year in AI — Best of 2025, Part I: Reasoning Models, LLM Agents and More. https://www.apolo.us/blog-posts/the-year-in-ai-best-of-2025-part-i-reasoning-models-llm-agents-and-more
-
- Deepest (2025). AI Model Releases in 2025: Every Major Launch, Ranked. https://www.deepest.app/blog/ai-model-releases-2025
-
- IBM (2026). A list of large language models (LLMs). https://www.ibm.com/think/topics/large-language-models-list
Provenance
Merged from Machine Learning Model (Artefact) (needs reconciliation)
The trained artefact produced by a machine learning pipeline: a parameterised function — such as a decision tree ensemble, support vector machine or neural network — whose weights have been fitted to data by an optimisation procedure, and which maps new inputs to predictions, classifications, rankings or generated content; the unit that is evaluated, selected, versioned, deployed, monitored and eventually retrained or retired in production systems.
Semantic Classification
Content
Definition
A machine learning model is the concrete output of training: a function with learned parameters that generalises from observed data to unseen inputs. The term covers everything from a logistic regression with a handful of coefficients to a transformer with hundreds of billions of weights. What unites them is the workflow around the artefact — a hypothesis class is chosen, parameters are fitted by minimising a loss over training data, generalisation is estimated on held-out data, and the frozen result is packaged for inference. The model is distinct from the algorithm that produced it (gradient descent, boosting) and from the system that serves it, a distinction that matters for versioning, reproducibility and governance.
The lifecycle of a model is the organising spine of applied machine learning. Upstream, data preparation, ground-truth labelling and Feature Selection determine what signal is available to fit; model selection and hyperparameter search compare candidate artefacts; threshold optimisation tunes the mapping from scores to decisions. Downstream, the model is serialised (ONNX, SavedModel, safetensors), deployed to servers or edge devices — often after quantisation or distillation to meet latency and memory budgets — and monitored for data drift, performance decay and fairness regressions. Model cards and registries document provenance, intended use and known limitations, requirements that regulation increasingly makes mandatory for high-risk applications.
In this graph the class acts as the hub that evaluation, debugging, selection and deployment concepts attach to: pages on generalisation, cross-validation, principal component analysis and edge deployment all describe operations performed on or for a machine learning model.
Technical Details
Models are usefully classified along three axes. By learning paradigm: supervised (classification, regression), unsupervised (clustering, dimensionality reduction), self-supervised and reinforcement-learned. By hypothesis class: linear models, kernel machines, tree ensembles (random forests, gradient-boosted trees), probabilistic graphical models and neural networks, with deep architectures dominating perception and language tasks while gradient boosting remains highly competitive on tabular data. By parametric character: parametric models with a fixed weight count versus non-parametric methods such as k-nearest neighbours whose effective capacity grows with the data. Key evaluation quantities — bias-variance trade-off, calibration, out-of-distribution robustness — are properties of the trained artefact rather than of the algorithm, which is why rigorous practice always distinguishes the two.
Current Landscape
-
The model artefact is now a regulated object: under the EU AI Act, obligations for providers of general-purpose AI models (Article 53) applied from 2 August 2025 — maintained technical documentation covering architecture, training and evaluation, compute and energy; documentation for downstream integrators; a copyright-compliance policy; and a public training-content summary using the Commission’s mandatory template published on 24 July 2025.
-
Transitional deadlines: GPAI models placed on the EU market before 2 August 2025 must be brought into compliance by 2 August 2027; models above the 10²⁵-FLOP systemic-risk threshold face additional adversarial-testing, incident-reporting, and cybersecurity obligations.
-
Voluntary scaffolding: the Commission’s GPAI Code of Practice and accompanying guidelines (July 2025) operationalise these duties; providers submit systemic-risk notifications and model reports through the EU’s SEND platform.
-
Serialisation and lineage practice: safetensors has displaced pickle-based weight formats for safe distribution on model hubs; model registries and lineage-tracking platforms (MLflow 3’s model-centric redesign, June 2025) increasingly bind the artefact to the exact code, data, prompts and evaluations that produced it — turning the model card from documentation courtesy into compliance evidence.
Sources:
-
https://digital-strategy.ec.europa.eu/en/policies/guidelines-gpai-providers
-
https://www.skadden.com/insights/publications/2025/08/eus-general-purpose-ai-obligations