Geometric raised dots on dark paper, one bio-cyan dot illuminated — acronyms decoded.

AI acronyms decoded

200+ terms with plain-language definitions and when you'd actually use them

If you read AI papers, vendor docs, or compliance contracts long enough, you start to notice that most of the cognitive load isn't the ideas — it's the vocabulary. RLHF, DPO, LoRA, MoE, SOC2, GDPR, FedRAMP. Each acronym is a compressed pointer to a body of work, a method, a regulation, or a benchmark, and you're expected to silently know all of them. This page is the decoder ring. The format is deliberate: acronym, expansion, one-sentence definition, and a category so you can find your neighborhood quickly. The categories cover the realistic surface area of modern AI work — training methods, alignment, retrieval, evaluation, vision, multimodal, agents, infrastructure, and the alphabet soup of compliance frameworks that touch any production AI system. Where a term has a canonical paper, we cite the arXiv ID. Where a regulation has a primary government source, we link the official text. A few honest disclosures up front. First, this is a best-effort snapshot as of June 2026 — provider names, model names, and pricing change weekly; check official docs for anything load-bearing. Second, acronyms in AI are not stable: the same three letters can mean different things in different subfields (PPO is a training algorithm in RL but also a health-insurance plan acronym in the US — we cover the AI sense). Third, we mark a term "industry term — no canonical paper" when it's widely used in vendor docs and industry conversation but lacks a single foundational reference; that's not a criticism, just accuracy. We've also pulled out the ones that trip up newcomers — pairs like NLU vs NLG, AUC vs AUROC, MSE vs RMSE — and the regulatory acronyms that solo founders and small teams discover the hard way (HIPAA, SOC2, FedRAMP) when an enterprise buyer puts them in a vendor questionnaire. The goal isn't to make you fluent. It's to make the next paragraph you read parse cleanly so you can think about the actual idea.

How to read this page

Use Ctrl-F. The categories are stable, the acronyms are sorted within each table, and the language is plain enough that search works.
Categories are: Core AI · Training & alignment · Parameter-efficient · Inference & reasoning · Retrieval & memory · Architectures · Vision & multimodal · Evaluation & benchmarks · Loss & metrics · Agents & protocols · Compliance & regulation · Infrastructure & deployment.
When you see 'industry term — no canonical paper,' that means the acronym is widely used but has no single foundational reference; it lives in vendor docs and practitioner usage, not in one citable paper.
Dollar amounts and dates labeled 'as of June 2026 best-effort' are time-stamped on purpose — pricing and model lineups move fast, so verify against the provider's live docs before quoting them anywhere that matters.
Pairs that confuse people (NLU vs NLG, AUC vs AUROC, MSE vs RMSE, RLHF vs DPO) are deliberately defined adjacent to each other in the tables so the distinction is visible at a glance.
If a term you need isn't here, the omission is honest: we'd rather not include it than guess. Email a.mccree@gmail.com and we'll add it with a real citation.

Core AI — the foundational vocabulary

These are the terms that show up in nearly every AI paper, product page, and earnings call. If you only learn 20 acronyms, learn these.

Acronym	Expansion	One-line definition	When you'd use it
AI	Artificial intelligence	Umbrella term for systems that perform tasks typically requiring human cognition.	General writing, marketing, policy — too broad for technical specs.
AGI	Artificial general intelligence	Hypothetical AI matching human cognitive ability across most domains; no agreed measurement.	Strategic discussions, safety research, investor decks — definitions vary by speaker.
ASI	Artificial superintelligence	Hypothetical AI substantially exceeding human cognition across all domains.	Long-horizon safety and policy debate; not a near-term product term.
ML	Machine learning	Systems that improve at a task by learning patterns from data rather than explicit rules.	Any context distinguishing learned systems from rule-based software.
DL	Deep learning	ML using neural networks with multiple layers; the dominant modern approach.	Distinguishing modern neural methods from classical ML (trees, SVMs).
LLM	Large language model	Neural network trained on large text corpora to predict and generate language.	Any conversation about GPT-style, Claude-style, or Llama-style systems.
SLM	Small language model	Compact LLM (often <10B parameters) optimized for cost, latency, or on-device use.	Edge deployment, cost-sensitive workloads, privacy-bound tasks.
VLM	Vision-language model	Model that jointly processes images and text in a unified representation.	Multimodal apps — describe images, OCR plus reasoning, document QA.
LMM	Large multimodal model	Generalization of LLM to multiple modalities (text + image + audio + video).	Multimodal product specs; sometimes used interchangeably with VLM.
NLP	Natural language processing	Field of building systems that read, write, or analyze human language.	Job titles, academic departments, broad capability descriptions.
NLU	Natural language understanding	Subfield of NLP focused on extracting meaning, intent, and structure from text.	Intent classification, entity extraction, conversational AI back-ends.
NLG	Natural language generation	Subfield of NLP focused on producing fluent text from data or representations.	Summarization, report generation, structured-data-to-prose pipelines.
GenAI	Generative AI	AI systems that produce new content (text, image, audio, code) rather than only classifying.	Product marketing, enterprise budget categories, vendor positioning.
AIGC	AI-generated content	Synonym for generated text/image/audio/video; more common in Chinese-language sources.	Reading non-Western AI coverage; provenance and watermarking discussions.
FM	Foundation model	Large pre-trained model adapted to many downstream tasks via fine-tuning or prompting.	Stanford-style framing, regulatory drafts (EU AI Act), academic papers.
GPT	Generative pre-trained transformer	OpenAI's model family; also a generic term for the architectural pattern.	Product references; the original 2018 paper is Radford et al.

AcronymAI

ExpansionArtificial intelligence

One-line definitionUmbrella term for systems that perform tasks typically requiring human cognition.

When you'd use itGeneral writing, marketing, policy — too broad for technical specs.

AcronymAGI

ExpansionArtificial general intelligence

One-line definitionHypothetical AI matching human cognitive ability across most domains; no agreed measurement.

When you'd use itStrategic discussions, safety research, investor decks — definitions vary by speaker.

AcronymASI

ExpansionArtificial superintelligence

One-line definitionHypothetical AI substantially exceeding human cognition across all domains.

When you'd use itLong-horizon safety and policy debate; not a near-term product term.

AcronymML

ExpansionMachine learning

One-line definitionSystems that improve at a task by learning patterns from data rather than explicit rules.

When you'd use itAny context distinguishing learned systems from rule-based software.

AcronymDL

ExpansionDeep learning

One-line definitionML using neural networks with multiple layers; the dominant modern approach.

When you'd use itDistinguishing modern neural methods from classical ML (trees, SVMs).

AcronymLLM

ExpansionLarge language model

One-line definitionNeural network trained on large text corpora to predict and generate language.

When you'd use itAny conversation about GPT-style, Claude-style, or Llama-style systems.

AcronymSLM

ExpansionSmall language model

One-line definitionCompact LLM (often <10B parameters) optimized for cost, latency, or on-device use.

When you'd use itEdge deployment, cost-sensitive workloads, privacy-bound tasks.

AcronymVLM

ExpansionVision-language model

One-line definitionModel that jointly processes images and text in a unified representation.

When you'd use itMultimodal apps — describe images, OCR plus reasoning, document QA.

AcronymLMM

ExpansionLarge multimodal model

One-line definitionGeneralization of LLM to multiple modalities (text + image + audio + video).

When you'd use itMultimodal product specs; sometimes used interchangeably with VLM.

AcronymNLP

ExpansionNatural language processing

One-line definitionField of building systems that read, write, or analyze human language.

When you'd use itJob titles, academic departments, broad capability descriptions.

AcronymNLU

ExpansionNatural language understanding

One-line definitionSubfield of NLP focused on extracting meaning, intent, and structure from text.

When you'd use itIntent classification, entity extraction, conversational AI back-ends.

AcronymNLG

ExpansionNatural language generation

One-line definitionSubfield of NLP focused on producing fluent text from data or representations.

When you'd use itSummarization, report generation, structured-data-to-prose pipelines.

AcronymGenAI

ExpansionGenerative AI

One-line definitionAI systems that produce new content (text, image, audio, code) rather than only classifying.

When you'd use itProduct marketing, enterprise budget categories, vendor positioning.

AcronymAIGC

ExpansionAI-generated content

One-line definitionSynonym for generated text/image/audio/video; more common in Chinese-language sources.

When you'd use itReading non-Western AI coverage; provenance and watermarking discussions.

AcronymFM

ExpansionFoundation model

One-line definitionLarge pre-trained model adapted to many downstream tasks via fine-tuning or prompting.

When you'd use itStanford-style framing, regulatory drafts (EU AI Act), academic papers.

AcronymGPT

ExpansionGenerative pre-trained transformer

One-line definitionOpenAI's model family; also a generic term for the architectural pattern.

When you'd use itProduct references; the original 2018 paper is Radford et al.

Training and alignment — how models learn what we want

These acronyms describe the methods used to take a base model and shape it into something useful and safe. The space moved fast between 2022 and 2025; RLHF was the dominant approach, then DPO simplified it, then several variants competed for the title of 'best preference-tuning method.'

Acronym	Expansion	One-line definition	Canonical reference
SFT	Supervised fine-tuning	Training a base model on curated input-output pairs to teach a target behavior.	Standard practice; no single canonical paper.
RLHF	Reinforcement learning from human feedback	Train a reward model on human preference data, then optimize the policy against it.	Christiano et al. 2017 (arxiv.org/abs/1706.03741); Ouyang et al. 2022 (arxiv.org/abs/2203.02155).
RLAIF	Reinforcement learning from AI feedback	Substitute a strong AI model for the human labeler in the RLHF pipeline.	Lee et al. 2023 (arxiv.org/abs/2309.00267).
DPO	Direct preference optimization	Closed-form alternative to RLHF that skips the explicit reward model.	Rafailov et al. 2023 (arxiv.org/abs/2305.18290).
IPO	Identity preference optimization	DPO variant that addresses overfitting to preference data.	Azar et al. 2023 (arxiv.org/abs/2310.12036).
KTO	Kahneman-Tversky optimization	Preference tuning that uses prospect-theory-style desirability labels, not pairs.	Ethayarajh et al. 2024 (arxiv.org/abs/2402.01306).
ORPO	Odds-ratio preference optimization	Combines SFT and preference optimization into a single training step.	Hong et al. 2024 (arxiv.org/abs/2403.07691).
GRPO	Group relative policy optimization	RL method used by DeepSeek for reasoning models; compares groups of samples.	DeepSeekMath, Shao et al. 2024 (arxiv.org/abs/2402.03300).
PPO	Proximal policy optimization	RL algorithm widely used in RLHF; clips policy updates to prevent collapse.	Schulman et al. 2017 (arxiv.org/abs/1707.06347).
TRPO	Trust region policy optimization	Predecessor to PPO; constrains policy updates by KL divergence.	Schulman et al. 2015 (arxiv.org/abs/1502.05477).
CAI	Constitutional AI	Anthropic method using AI-written critiques against a written constitution.	Bai et al. 2022 (arxiv.org/abs/2212.08073).
RM	Reward model	Auxiliary model that scores outputs by predicted human preference.	Used in RLHF; trained on preference comparisons.

AcronymSFT

ExpansionSupervised fine-tuning

One-line definitionTraining a base model on curated input-output pairs to teach a target behavior.

Canonical referenceStandard practice; no single canonical paper.

AcronymRLHF

ExpansionReinforcement learning from human feedback

One-line definitionTrain a reward model on human preference data, then optimize the policy against it.

Canonical referenceChristiano et al. 2017 (arxiv.org/abs/1706.03741); Ouyang et al. 2022 (arxiv.org/abs/2203.02155).

AcronymRLAIF

ExpansionReinforcement learning from AI feedback

One-line definitionSubstitute a strong AI model for the human labeler in the RLHF pipeline.

Canonical referenceLee et al. 2023 (arxiv.org/abs/2309.00267).

AcronymDPO

ExpansionDirect preference optimization

One-line definitionClosed-form alternative to RLHF that skips the explicit reward model.

Canonical referenceRafailov et al. 2023 (arxiv.org/abs/2305.18290).

AcronymIPO

ExpansionIdentity preference optimization

One-line definitionDPO variant that addresses overfitting to preference data.

Canonical referenceAzar et al. 2023 (arxiv.org/abs/2310.12036).

AcronymKTO

ExpansionKahneman-Tversky optimization

One-line definitionPreference tuning that uses prospect-theory-style desirability labels, not pairs.

Canonical referenceEthayarajh et al. 2024 (arxiv.org/abs/2402.01306).

AcronymORPO

ExpansionOdds-ratio preference optimization

One-line definitionCombines SFT and preference optimization into a single training step.

Canonical referenceHong et al. 2024 (arxiv.org/abs/2403.07691).

AcronymGRPO

ExpansionGroup relative policy optimization

One-line definitionRL method used by DeepSeek for reasoning models; compares groups of samples.

Canonical referenceDeepSeekMath, Shao et al. 2024 (arxiv.org/abs/2402.03300).

AcronymPPO

ExpansionProximal policy optimization

One-line definitionRL algorithm widely used in RLHF; clips policy updates to prevent collapse.

Canonical referenceSchulman et al. 2017 (arxiv.org/abs/1707.06347).

AcronymTRPO

ExpansionTrust region policy optimization

One-line definitionPredecessor to PPO; constrains policy updates by KL divergence.

Canonical referenceSchulman et al. 2015 (arxiv.org/abs/1502.05477).

AcronymCAI

ExpansionConstitutional AI

One-line definitionAnthropic method using AI-written critiques against a written constitution.

Canonical referenceBai et al. 2022 (arxiv.org/abs/2212.08073).

AcronymRM

ExpansionReward model

One-line definitionAuxiliary model that scores outputs by predicted human preference.

Canonical referenceUsed in RLHF; trained on preference comparisons.

Parameter-efficient fine-tuning

When you can't afford to fine-tune all the weights of a 70B-parameter model, these are the methods that let you train a small fraction of parameters and still get most of the benefit. As of June 2026 best-effort, LoRA and QLoRA are the workhorses for indie and research use; full fine-tuning is reserved for labs with serious GPU budgets.

Acronym	Expansion	One-line definition	Canonical reference
PEFT	Parameter-efficient fine-tuning	Umbrella for methods that adapt a model by training a small subset of parameters.	Hugging Face PEFT library; multiple foundational papers.
LoRA	Low-rank adaptation	Inject small trainable low-rank matrices into the model; freeze the original weights.	Hu et al. 2021 (arxiv.org/abs/2106.09685).
QLoRA	Quantized LoRA	LoRA combined with 4-bit quantization of the base model for memory-cheap tuning.	Dettmers et al. 2023 (arxiv.org/abs/2305.14314).
DoRA	Weight-decomposed low-rank adaptation	Extension of LoRA that decomposes weights into magnitude and direction components.	Liu et al. 2024 (arxiv.org/abs/2402.09353).
IA3	Infused adapter by inhibiting and amplifying inner activations	PEFT method that rescales activations with learned vectors; very few new parameters.	Liu et al. 2022 (arxiv.org/abs/2205.05638).
Adapter	Adapter tuning	Insert small bottleneck modules between transformer layers; train only those.	Houlsby et al. 2019 (arxiv.org/abs/1902.00751).
Prefix Tuning	Prefix tuning	Learn a sequence of continuous task-specific 'prefix' vectors prepended to the input.	Li and Liang 2021 (arxiv.org/abs/2101.00190).
P-Tuning	Prompt tuning	Learn continuous embeddings as prompts rather than discrete tokens.	Lester et al. 2021 (arxiv.org/abs/2104.08691).

AcronymPEFT

ExpansionParameter-efficient fine-tuning

One-line definitionUmbrella for methods that adapt a model by training a small subset of parameters.

Canonical referenceHugging Face PEFT library; multiple foundational papers.

AcronymLoRA

ExpansionLow-rank adaptation

One-line definitionInject small trainable low-rank matrices into the model; freeze the original weights.

Canonical referenceHu et al. 2021 (arxiv.org/abs/2106.09685).

AcronymQLoRA

ExpansionQuantized LoRA

One-line definitionLoRA combined with 4-bit quantization of the base model for memory-cheap tuning.

Canonical referenceDettmers et al. 2023 (arxiv.org/abs/2305.14314).

AcronymDoRA

ExpansionWeight-decomposed low-rank adaptation

One-line definitionExtension of LoRA that decomposes weights into magnitude and direction components.

Canonical referenceLiu et al. 2024 (arxiv.org/abs/2402.09353).

AcronymIA3

ExpansionInfused adapter by inhibiting and amplifying inner activations

One-line definitionPEFT method that rescales activations with learned vectors; very few new parameters.

Canonical referenceLiu et al. 2022 (arxiv.org/abs/2205.05638).

AcronymAdapter

ExpansionAdapter tuning

One-line definitionInsert small bottleneck modules between transformer layers; train only those.

Canonical referenceHoulsby et al. 2019 (arxiv.org/abs/1902.00751).

AcronymPrefix Tuning

ExpansionPrefix tuning

One-line definitionLearn a sequence of continuous task-specific 'prefix' vectors prepended to the input.

Canonical referenceLi and Liang 2021 (arxiv.org/abs/2101.00190).

AcronymP-Tuning

ExpansionPrompt tuning

One-line definitionLearn continuous embeddings as prompts rather than discrete tokens.

Canonical referenceLester et al. 2021 (arxiv.org/abs/2104.08691).

Inference, reasoning, and prompting

What models do at the moment you ask them a question. This corner of the field exploded in late 2024 and 2025 around 'test-time compute' — the idea that spending more inference cycles on a hard problem can substitute for a larger model.

Acronym	Expansion	One-line definition	Canonical reference
CoT	Chain of thought	Prompt the model to show intermediate reasoning steps before its final answer.	Wei et al. 2022 (arxiv.org/abs/2201.11903).
ToT	Tree of thoughts	Explore multiple reasoning branches and evaluate which to expand.	Yao et al. 2023 (arxiv.org/abs/2305.10601).
GoT	Graph of thoughts	Generalize ToT to a graph structure with merges and feedback edges.	Besta et al. 2023 (arxiv.org/abs/2308.09687).
RoT	Reflection of thought	Reasoning pattern where the model critiques and revises its own output.	Industry term, related to Reflexion (Shinn et al. 2023, arxiv.org/abs/2303.11366).
ICL	In-context learning	The model learns from examples placed in the prompt without any weight updates.	Brown et al. 2020 (GPT-3 paper, arxiv.org/abs/2005.14165).
TTC	Test-time compute	Spending more inference-time computation (sampling, search, verification) to improve answers.	Snell et al. 2024 (arxiv.org/abs/2408.03314).
TT	Test-time	General modifier; applied to learning, scaling, or compute done at inference.	Industry usage; specific methods cite their own papers.
KV Cache	Key-value cache	Cached attention tensors that let transformers reuse prior tokens' work during generation.	Standard implementation detail; widely documented.
MoE	Mixture of experts	Architecture where each token routes to a small subset of specialized 'expert' subnetworks.	Shazeer et al. 2017 (arxiv.org/abs/1701.06538).
SAE	Sparse autoencoder	Interpretability tool that decomposes activations into a sparse set of human-readable features.	Anthropic interpretability work; Bricken et al. 2023 (transformer-circuits.pub).

AcronymCoT

ExpansionChain of thought

One-line definitionPrompt the model to show intermediate reasoning steps before its final answer.

Canonical referenceWei et al. 2022 (arxiv.org/abs/2201.11903).

AcronymToT

ExpansionTree of thoughts

One-line definitionExplore multiple reasoning branches and evaluate which to expand.

Canonical referenceYao et al. 2023 (arxiv.org/abs/2305.10601).

AcronymGoT

ExpansionGraph of thoughts

One-line definitionGeneralize ToT to a graph structure with merges and feedback edges.

Canonical referenceBesta et al. 2023 (arxiv.org/abs/2308.09687).

AcronymRoT

ExpansionReflection of thought

One-line definitionReasoning pattern where the model critiques and revises its own output.

Canonical referenceIndustry term, related to Reflexion (Shinn et al. 2023, arxiv.org/abs/2303.11366).

AcronymICL

ExpansionIn-context learning

One-line definitionThe model learns from examples placed in the prompt without any weight updates.

Canonical referenceBrown et al. 2020 (GPT-3 paper, arxiv.org/abs/2005.14165).

AcronymTTC

ExpansionTest-time compute

One-line definitionSpending more inference-time computation (sampling, search, verification) to improve answers.

Canonical referenceSnell et al. 2024 (arxiv.org/abs/2408.03314).

AcronymTT

ExpansionTest-time

One-line definitionGeneral modifier; applied to learning, scaling, or compute done at inference.

Canonical referenceIndustry usage; specific methods cite their own papers.

AcronymKV Cache

ExpansionKey-value cache

One-line definitionCached attention tensors that let transformers reuse prior tokens' work during generation.

Canonical referenceStandard implementation detail; widely documented.

AcronymMoE

ExpansionMixture of experts

One-line definitionArchitecture where each token routes to a small subset of specialized 'expert' subnetworks.

Canonical referenceShazeer et al. 2017 (arxiv.org/abs/1701.06538).

AcronymSAE

ExpansionSparse autoencoder

One-line definitionInterpretability tool that decomposes activations into a sparse set of human-readable features.

Canonical referenceAnthropic interpretability work; Bricken et al. 2023 (transformer-circuits.pub).

Retrieval, memory, and grounding

How models access information they don't have in their weights — documents, databases, prior turns. Critical for any product where current facts, proprietary data, or citation requirements matter.

Acronym	Expansion	One-line definition	Canonical reference
RAG	Retrieval-augmented generation	Retrieve relevant documents at query time and feed them into the model's context.	Lewis et al. 2020 (arxiv.org/abs/2005.11401).
GraphRAG	Graph retrieval-augmented generation	RAG variant that retrieves over a knowledge graph rather than flat documents.	Edge et al. 2024 (arxiv.org/abs/2404.16130); Microsoft Research.
HyDE	Hypothetical document embeddings	Generate a synthetic answer first, embed it, then use it to retrieve real documents.	Gao et al. 2022 (arxiv.org/abs/2212.10496).
BM25	Best Matching 25	Classical sparse retrieval scoring function; still a strong baseline for RAG.	Robertson and Zaragoza 2009 (foundations and trends survey).
FAISS	Facebook AI Similarity Search	Open-source library for fast vector similarity search.	Johnson et al. 2017 (arxiv.org/abs/1702.08734); github.com/facebookresearch/faiss.
ANN	Approximate nearest neighbor	Family of algorithms that find near-matches in vector space without exhaustive search.	Multiple algorithms; HNSW is currently dominant.
HNSW	Hierarchical navigable small world	Graph-based ANN method that powers most production vector databases.	Malkov and Yashunin 2016 (arxiv.org/abs/1603.09320).
VDB	Vector database	Database optimized for storing and querying high-dimensional embedding vectors.	Industry term; Pinecone, Weaviate, Qdrant, Milvus are common implementations.

AcronymRAG

ExpansionRetrieval-augmented generation

One-line definitionRetrieve relevant documents at query time and feed them into the model's context.

Canonical referenceLewis et al. 2020 (arxiv.org/abs/2005.11401).

AcronymGraphRAG

ExpansionGraph retrieval-augmented generation

One-line definitionRAG variant that retrieves over a knowledge graph rather than flat documents.

Canonical referenceEdge et al. 2024 (arxiv.org/abs/2404.16130); Microsoft Research.

AcronymHyDE

ExpansionHypothetical document embeddings

One-line definitionGenerate a synthetic answer first, embed it, then use it to retrieve real documents.

Canonical referenceGao et al. 2022 (arxiv.org/abs/2212.10496).

AcronymBM25

ExpansionBest Matching 25

One-line definitionClassical sparse retrieval scoring function; still a strong baseline for RAG.

Canonical referenceRobertson and Zaragoza 2009 (foundations and trends survey).

AcronymFAISS

ExpansionFacebook AI Similarity Search

One-line definitionOpen-source library for fast vector similarity search.

Canonical referenceJohnson et al. 2017 (arxiv.org/abs/1702.08734); github.com/facebookresearch/faiss.

AcronymANN

ExpansionApproximate nearest neighbor

One-line definitionFamily of algorithms that find near-matches in vector space without exhaustive search.

Canonical referenceMultiple algorithms; HNSW is currently dominant.

AcronymHNSW

ExpansionHierarchical navigable small world

One-line definitionGraph-based ANN method that powers most production vector databases.

Canonical referenceMalkov and Yashunin 2016 (arxiv.org/abs/1603.09320).

AcronymVDB

ExpansionVector database

One-line definitionDatabase optimized for storing and querying high-dimensional embedding vectors.

Canonical referenceIndustry term; Pinecone, Weaviate, Qdrant, Milvus are common implementations.

Architectures, vision, and multimodal

The architectural building blocks. Transformers dominate, but the older shapes (CNN, RNN, LSTM) still ship in production and you'll see them referenced in legacy systems and edge deployments.

Acronym	Expansion	One-line definition	Canonical reference
MLP	Multi-layer perceptron	Basic feedforward neural network; building block of nearly everything else.	Rumelhart, Hinton, Williams 1986.
CNN	Convolutional neural network	Architecture using shared local filters; dominant for image tasks before ViT.	LeCun et al. 1989; AlexNet (Krizhevsky et al. 2012).
RNN	Recurrent neural network	Sequential network that maintains hidden state across timesteps.	Hopfield 1982; Elman 1990.
LSTM	Long short-term memory	RNN variant with gating mechanisms that addresses vanishing gradients.	Hochreiter and Schmidhuber 1997.
GRU	Gated recurrent unit	Simpler LSTM alternative with fewer gates.	Cho et al. 2014 (arxiv.org/abs/1409.1259).
Transformer	Transformer	Architecture based on self-attention; foundation of modern LLMs.	Vaswani et al. 2017 'Attention Is All You Need' (arxiv.org/abs/1706.03762).
ViT	Vision transformer	Apply transformer architecture to image patches instead of tokens.	Dosovitskiy et al. 2020 (arxiv.org/abs/2010.11929).
DiT	Diffusion transformer	Transformer backbone for diffusion-based image and video generation.	Peebles and Xie 2022 (arxiv.org/abs/2212.09748).
CLIP	Contrastive language-image pre-training	Joint image-text embedding model; the bridge between vision and language.	Radford et al. 2021 (arxiv.org/abs/2103.00020).
BLIP	Bootstrapping language-image pre-training	Vision-language pre-training that combines understanding and generation.	Li et al. 2022 (arxiv.org/abs/2201.12086).
GAN	Generative adversarial network	Two networks (generator + discriminator) compete to produce realistic samples.	Goodfellow et al. 2014 (arxiv.org/abs/1406.2661).
VAE	Variational autoencoder	Probabilistic autoencoder that learns a latent space with regularization.	Kingma and Welling 2013 (arxiv.org/abs/1312.6114).
NeRF	Neural radiance fields	Represent a 3D scene as a neural network mapping coordinates to color and density.	Mildenhall et al. 2020 (arxiv.org/abs/2003.08934).

AcronymMLP

ExpansionMulti-layer perceptron

One-line definitionBasic feedforward neural network; building block of nearly everything else.

Canonical referenceRumelhart, Hinton, Williams 1986.

AcronymCNN

ExpansionConvolutional neural network

One-line definitionArchitecture using shared local filters; dominant for image tasks before ViT.

Canonical referenceLeCun et al. 1989; AlexNet (Krizhevsky et al. 2012).

AcronymRNN

ExpansionRecurrent neural network

One-line definitionSequential network that maintains hidden state across timesteps.

Canonical referenceHopfield 1982; Elman 1990.

AcronymLSTM

ExpansionLong short-term memory

One-line definitionRNN variant with gating mechanisms that addresses vanishing gradients.

Canonical referenceHochreiter and Schmidhuber 1997.

AcronymGRU

ExpansionGated recurrent unit

One-line definitionSimpler LSTM alternative with fewer gates.

Canonical referenceCho et al. 2014 (arxiv.org/abs/1409.1259).

AcronymTransformer

ExpansionTransformer

One-line definitionArchitecture based on self-attention; foundation of modern LLMs.

Canonical referenceVaswani et al. 2017 'Attention Is All You Need' (arxiv.org/abs/1706.03762).

AcronymViT

ExpansionVision transformer

One-line definitionApply transformer architecture to image patches instead of tokens.

Canonical referenceDosovitskiy et al. 2020 (arxiv.org/abs/2010.11929).

AcronymDiT

ExpansionDiffusion transformer

One-line definitionTransformer backbone for diffusion-based image and video generation.

Canonical referencePeebles and Xie 2022 (arxiv.org/abs/2212.09748).

AcronymCLIP

ExpansionContrastive language-image pre-training

One-line definitionJoint image-text embedding model; the bridge between vision and language.

Canonical referenceRadford et al. 2021 (arxiv.org/abs/2103.00020).

AcronymBLIP

ExpansionBootstrapping language-image pre-training

One-line definitionVision-language pre-training that combines understanding and generation.

Canonical referenceLi et al. 2022 (arxiv.org/abs/2201.12086).

AcronymGAN

ExpansionGenerative adversarial network

One-line definitionTwo networks (generator + discriminator) compete to produce realistic samples.

Canonical referenceGoodfellow et al. 2014 (arxiv.org/abs/1406.2661).

AcronymVAE

ExpansionVariational autoencoder

One-line definitionProbabilistic autoencoder that learns a latent space with regularization.

Canonical referenceKingma and Welling 2013 (arxiv.org/abs/1312.6114).

AcronymNeRF

ExpansionNeural radiance fields

One-line definitionRepresent a 3D scene as a neural network mapping coordinates to color and density.

Canonical referenceMildenhall et al. 2020 (arxiv.org/abs/2003.08934).

Evaluation and benchmarks

The yardsticks. Benchmarks are imperfect — they get gamed, leaked into training data, or stop discriminating between top models. Always check the leaderboard's contamination notes and the model's training-cutoff date before drawing conclusions. As of June 2026 best-effort, the live benchmarks worth watching are LiveBench, SWE-Bench Verified, and GPQA Diamond; older saturated ones (MMLU, HumanEval base) still appear in marketing but tell you less.

MMLU

Hendrycks et al. 2020 · arxiv.org/abs/2009.03300

Massive multitask language understanding — 57 academic subjects, multiple choice. Saturated near the top; useful as a sanity floor.

MMLU-Pro

Wang et al. 2024 · arxiv.org/abs/2406.01574

Harder MMLU successor with more answer choices and more reasoning. Less saturated as of 2026.

GPQA

Rein et al. 2023 · arxiv.org/abs/2311.12022

Graduate-level Google-proof Q&A in biology, physics, chemistry. The 'Diamond' subset is the hard one.

HumanEval

Chen et al. 2021 (Codex) · arxiv.org/abs/2107.03374

164 Python programming problems with unit tests. The original code-completion benchmark; now saturated.

MBPP

Austin et al. 2021 · arxiv.org/abs/2108.07732

Mostly Basic Python Problems — 974 entry-level coding tasks paired with HumanEval.

SWE-Bench

Jimenez et al. 2023 · arxiv.org/abs/2310.06770 · swebench.com

Real GitHub issues from popular Python repos. 'Verified' subset is hand-curated by humans for correctness.

LiveBench

White et al. 2024 · arxiv.org/abs/2406.19314 · livebench.ai

Continuously refreshed benchmark to resist contamination. Updated monthly.

BIG-Bench

Srivastava et al. 2022 · arxiv.org/abs/2206.04615

200+ task collaborative benchmark from Google. BIG-Bench Hard (BBH) is the still-useful subset.

ARC-AGI

Chollet 2019 · arxiv.org/abs/1911.01547 · arcprize.org

François Chollet's abstract-reasoning benchmark designed to resist memorization. ARC Prize is the live competition.

HELM

Liang et al. 2022 · arxiv.org/abs/2211.09110 · crfm.stanford.edu/helm

Holistic evaluation of language models — Stanford CRFM's broad multi-axis benchmark suite.

Chatbot Arena

Chiang et al. 2024 · arxiv.org/abs/2403.04132 · lmarena.ai

ELO-style human preference leaderboard from LMSYS. Now LMArena; reflects user preference, not capability per se.

Pass@k

Defined in Codex paper · arxiv.org/abs/2107.03374

Metric for code generation: probability that at least one of k sampled completions passes all tests.

Loss functions, metrics, and statistics

The math symbols hiding behind the acronyms. Most papers will use at least three of these; product-side readers can skim, ML readers need them at instinct level.

Acronym	Expansion	One-line definition	When you'd use it
MSE	Mean squared error	Average of squared differences between predicted and true values.	Regression loss; penalizes large errors more than small.
RMSE	Root mean squared error	Square root of MSE; same units as the target variable.	Reporting regression error in the variable's natural scale.
MAE	Mean absolute error	Average of absolute differences; robust to outliers compared to MSE.	When outliers shouldn't dominate the loss.
CE	Cross-entropy	Loss measuring distance between predicted and true probability distributions.	Classification, including next-token prediction in LLMs.
BCE	Binary cross-entropy	CE specialized to two-class problems.	Binary classifiers, multi-label heads.
NLL	Negative log likelihood	Equivalent framing of CE in maximum-likelihood language.	Statistical writing; identical to CE in practice.
KL	Kullback-Leibler divergence	Asymmetric measure of how one probability distribution differs from another.	RLHF regularization, variational inference, distillation.
JS	Jensen-Shannon divergence	Symmetric, smoothed version of KL divergence.	Comparing distributions when symmetry matters.
BLEU	Bilingual evaluation understudy	N-gram precision metric for machine translation.	Translation papers; now considered weak for modern systems.
ROUGE	Recall-oriented understudy for gisting evaluation	N-gram recall family for summarization.	Summarization benchmarks; ROUGE-L most common.
METEOR	Metric for evaluation of translation with explicit ordering	Translation metric that uses stemming and synonyms.	Translation evaluation; correlates better with humans than BLEU.
F1	F1 score	Harmonic mean of precision and recall.	Imbalanced classification, NER, QA.
AUC	Area under the curve	Generic term — usually refers to AUROC unless otherwise specified.	Classifier evaluation; check which curve is meant.
AUROC	Area under the receiver operating characteristic curve	Probability the classifier ranks a random positive above a random negative.	Threshold-independent classifier evaluation.
AUPRC	Area under the precision-recall curve	Like AUROC but uses precision-recall; better for imbalanced data.	Rare-event detection, fraud, medical.
MAP	Mean average precision	Average precision across queries; standard ranking metric.	Information retrieval, object detection.
NDCG	Normalized discounted cumulative gain	Ranking metric that rewards relevance and position.	Search ranking evaluation.
EM	Exact match	Did the predicted answer exactly equal the reference?	Extractive QA, structured-output tasks.

AcronymMSE

ExpansionMean squared error

One-line definitionAverage of squared differences between predicted and true values.

When you'd use itRegression loss; penalizes large errors more than small.

AcronymRMSE

ExpansionRoot mean squared error

One-line definitionSquare root of MSE; same units as the target variable.

When you'd use itReporting regression error in the variable's natural scale.

AcronymMAE

ExpansionMean absolute error

One-line definitionAverage of absolute differences; robust to outliers compared to MSE.

When you'd use itWhen outliers shouldn't dominate the loss.

AcronymCE

ExpansionCross-entropy

One-line definitionLoss measuring distance between predicted and true probability distributions.

When you'd use itClassification, including next-token prediction in LLMs.

AcronymBCE

ExpansionBinary cross-entropy

One-line definitionCE specialized to two-class problems.

When you'd use itBinary classifiers, multi-label heads.

AcronymNLL

ExpansionNegative log likelihood

One-line definitionEquivalent framing of CE in maximum-likelihood language.

When you'd use itStatistical writing; identical to CE in practice.

AcronymKL

ExpansionKullback-Leibler divergence

One-line definitionAsymmetric measure of how one probability distribution differs from another.

When you'd use itRLHF regularization, variational inference, distillation.

AcronymJS

ExpansionJensen-Shannon divergence

One-line definitionSymmetric, smoothed version of KL divergence.

When you'd use itComparing distributions when symmetry matters.

AcronymBLEU

ExpansionBilingual evaluation understudy

One-line definitionN-gram precision metric for machine translation.

When you'd use itTranslation papers; now considered weak for modern systems.

AcronymROUGE

ExpansionRecall-oriented understudy for gisting evaluation

One-line definitionN-gram recall family for summarization.

When you'd use itSummarization benchmarks; ROUGE-L most common.

AcronymMETEOR

ExpansionMetric for evaluation of translation with explicit ordering

One-line definitionTranslation metric that uses stemming and synonyms.

When you'd use itTranslation evaluation; correlates better with humans than BLEU.

AcronymF1

ExpansionF1 score

One-line definitionHarmonic mean of precision and recall.

When you'd use itImbalanced classification, NER, QA.

AcronymAUC

ExpansionArea under the curve

One-line definitionGeneric term — usually refers to AUROC unless otherwise specified.

When you'd use itClassifier evaluation; check which curve is meant.

AcronymAUROC

ExpansionArea under the receiver operating characteristic curve

One-line definitionProbability the classifier ranks a random positive above a random negative.

When you'd use itThreshold-independent classifier evaluation.

AcronymAUPRC

ExpansionArea under the precision-recall curve

One-line definitionLike AUROC but uses precision-recall; better for imbalanced data.

When you'd use itRare-event detection, fraud, medical.

AcronymMAP

ExpansionMean average precision

One-line definitionAverage precision across queries; standard ranking metric.

When you'd use itInformation retrieval, object detection.

AcronymNDCG

ExpansionNormalized discounted cumulative gain

One-line definitionRanking metric that rewards relevance and position.

When you'd use itSearch ranking evaluation.

AcronymEM

ExpansionExact match

One-line definitionDid the predicted answer exactly equal the reference?

When you'd use itExtractive QA, structured-output tasks.

Agents, protocols, and tool use

The 2024-2026 wave of acronyms around models that take actions. MCP is the one solo developers run into first; A2A, BAA-style protocols, and orchestration frameworks come next.

Acronym	Expansion	One-line definition	Canonical reference
MCP	Model Context Protocol	Anthropic open standard for connecting AI assistants to data sources and tools.	modelcontextprotocol.io; announced by Anthropic November 2024.
A2A	Agent-to-agent	Google-led protocol for inter-agent communication across vendors.	google.github.io/A2A; announced April 2025.
ReAct	Reasoning + acting	Interleave reasoning steps with tool-use actions in a single trace.	Yao et al. 2022 (arxiv.org/abs/2210.03629).
ART	Automatic reasoning and tool-use	Framework for retrieving tool-use demonstrations to guide agents.	Paranjape et al. 2023 (arxiv.org/abs/2303.09014).
LATS	Language agent tree search	Tree-search agent framework combining ReAct with MCTS-style planning.	Zhou et al. 2023 (arxiv.org/abs/2310.04406).
FC	Function calling	Vendor-side API feature where models emit structured tool calls.	Industry term; standardized in OpenAI, Anthropic, Google APIs.

AcronymMCP

ExpansionModel Context Protocol

One-line definitionAnthropic open standard for connecting AI assistants to data sources and tools.

Canonical referencemodelcontextprotocol.io; announced by Anthropic November 2024.

AcronymA2A

ExpansionAgent-to-agent

One-line definitionGoogle-led protocol for inter-agent communication across vendors.

Canonical referencegoogle.github.io/A2A; announced April 2025.

AcronymReAct

ExpansionReasoning + acting

One-line definitionInterleave reasoning steps with tool-use actions in a single trace.

Canonical referenceYao et al. 2022 (arxiv.org/abs/2210.03629).

AcronymART

ExpansionAutomatic reasoning and tool-use

One-line definitionFramework for retrieving tool-use demonstrations to guide agents.

Canonical referenceParanjape et al. 2023 (arxiv.org/abs/2303.09014).

AcronymLATS

ExpansionLanguage agent tree search

One-line definitionTree-search agent framework combining ReAct with MCTS-style planning.

Canonical referenceZhou et al. 2023 (arxiv.org/abs/2310.04406).

AcronymFC

ExpansionFunction calling

One-line definitionVendor-side API feature where models emit structured tool calls.

Canonical referenceIndustry term; standardized in OpenAI, Anthropic, Google APIs.

Compliance, regulation, and legal — the alphabet soup that gates B2B deals

These are the acronyms that appear in vendor questionnaires and procurement reviews. None of them are AI-specific; all of them apply to AI systems. As a solo founder or small team, you will discover them the first time an enterprise buyer sends a 200-row spreadsheet. Verify current requirements against the official primary source — regulations are updated more often than this page.

Acronym	Expansion	One-line definition	Primary source
GDPR	General Data Protection Regulation	EU regulation governing processing of personal data; applies to anyone serving EU residents.	Regulation (EU) 2016/679; eur-lex.europa.eu/eli/reg/2016/679/oj.
CCPA	California Consumer Privacy Act	California state law granting consumers rights over their personal data.	Cal. Civ. Code § 1798.100 et seq.; oag.ca.gov/privacy/ccpa.
CPRA	California Privacy Rights Act	2020 amendment that strengthened CCPA and created the CPPA agency.	Proposition 24 (2020); cppa.ca.gov.
HIPAA	Health Insurance Portability and Accountability Act	US law setting privacy and security standards for protected health information.	Pub. L. 104-191 (1996); hhs.gov/hipaa.
BAA	Business Associate Agreement	Contract required under HIPAA when a vendor processes PHI on behalf of a covered entity.	Defined in 45 CFR § 164.504(e); hhs.gov/hipaa/for-professionals/covered-entities/sample-business-associate-agreement-provisions.
DPA	Data Processing Agreement	Contract required under GDPR between data controller and data processor.	GDPR Article 28.
DPIA	Data Protection Impact Assessment	GDPR-required assessment for high-risk processing of personal data.	GDPR Article 35.
GLBA	Gramm-Leach-Bliley Act	US law requiring financial institutions to protect customer financial information.	Pub. L. 106-102 (1999); ftc.gov/business-guidance/privacy-security/gramm-leach-bliley-act.
SOX	Sarbanes-Oxley Act	US law on corporate financial disclosure and internal controls for public companies.	Pub. L. 107-204 (2002).
SOC 2	Service Organization Control 2	AICPA audit framework for security, availability, processing integrity, confidentiality, privacy.	aicpa-cima.com; SOC 2 reports come from licensed CPA firms.
ISO 27001	ISO/IEC 27001	International standard for information security management systems.	iso.org/standard/27001; current version is 2022.
FedRAMP	Federal Risk and Authorization Management Program	US federal program for cloud security authorization.	fedramp.gov.
FERPA	Family Educational Rights and Privacy Act	US law protecting student education records.	20 U.S.C. § 1232g; studentprivacy.ed.gov.
COPPA	Children's Online Privacy Protection Act	US law restricting collection of data from children under 13.	15 U.S.C. §§ 6501–6506; ftc.gov/business-guidance/privacy-security/childrens-privacy.
PCI DSS	Payment Card Industry Data Security Standard	Industry standard for handling cardholder data; mandated by card brands.	pcisecuritystandards.org; current version is 4.0.1.
EU AI Act	EU Artificial Intelligence Act	EU regulation classifying AI systems by risk and imposing obligations accordingly.	Regulation (EU) 2024/1689; eur-lex.europa.eu.
NIST AI RMF	NIST AI Risk Management Framework	US voluntary framework for managing AI-related risks across the lifecycle.	nist.gov/itl/ai-risk-management-framework; AI RMF 1.0 released January 2023.
NIST CSF	NIST Cybersecurity Framework	Voluntary framework for managing cybersecurity risk; widely adopted.	nist.gov/cyberframework; version 2.0 released February 2024.

AcronymGDPR

ExpansionGeneral Data Protection Regulation

One-line definitionEU regulation governing processing of personal data; applies to anyone serving EU residents.

Primary sourceRegulation (EU) 2016/679; eur-lex.europa.eu/eli/reg/2016/679/oj.

AcronymCCPA

ExpansionCalifornia Consumer Privacy Act

One-line definitionCalifornia state law granting consumers rights over their personal data.

Primary sourceCal. Civ. Code § 1798.100 et seq.; oag.ca.gov/privacy/ccpa.

AcronymCPRA

ExpansionCalifornia Privacy Rights Act

One-line definition2020 amendment that strengthened CCPA and created the CPPA agency.

Primary sourceProposition 24 (2020); cppa.ca.gov.

AcronymHIPAA

ExpansionHealth Insurance Portability and Accountability Act

One-line definitionUS law setting privacy and security standards for protected health information.

Primary sourcePub. L. 104-191 (1996); hhs.gov/hipaa.

AcronymBAA

ExpansionBusiness Associate Agreement

One-line definitionContract required under HIPAA when a vendor processes PHI on behalf of a covered entity.

Primary sourceDefined in 45 CFR § 164.504(e); hhs.gov/hipaa/for-professionals/covered-entities/sample-business-associate-agreement-provisions.

AcronymDPA

ExpansionData Processing Agreement

One-line definitionContract required under GDPR between data controller and data processor.

Primary sourceGDPR Article 28.

AcronymDPIA

ExpansionData Protection Impact Assessment

One-line definitionGDPR-required assessment for high-risk processing of personal data.

Primary sourceGDPR Article 35.

AcronymGLBA

ExpansionGramm-Leach-Bliley Act

One-line definitionUS law requiring financial institutions to protect customer financial information.

Primary sourcePub. L. 106-102 (1999); ftc.gov/business-guidance/privacy-security/gramm-leach-bliley-act.

AcronymSOX

ExpansionSarbanes-Oxley Act

One-line definitionUS law on corporate financial disclosure and internal controls for public companies.

Primary sourcePub. L. 107-204 (2002).

AcronymSOC 2

ExpansionService Organization Control 2

One-line definitionAICPA audit framework for security, availability, processing integrity, confidentiality, privacy.

Primary sourceaicpa-cima.com; SOC 2 reports come from licensed CPA firms.

AcronymISO 27001

ExpansionISO/IEC 27001

One-line definitionInternational standard for information security management systems.

Primary sourceiso.org/standard/27001; current version is 2022.

AcronymFedRAMP

ExpansionFederal Risk and Authorization Management Program

One-line definitionUS federal program for cloud security authorization.

Primary sourcefedramp.gov.

AcronymFERPA

ExpansionFamily Educational Rights and Privacy Act

One-line definitionUS law protecting student education records.

Primary source20 U.S.C. § 1232g; studentprivacy.ed.gov.

AcronymCOPPA

ExpansionChildren's Online Privacy Protection Act

One-line definitionUS law restricting collection of data from children under 13.

Primary source15 U.S.C. §§ 6501–6506; ftc.gov/business-guidance/privacy-security/childrens-privacy.

AcronymPCI DSS

ExpansionPayment Card Industry Data Security Standard

One-line definitionIndustry standard for handling cardholder data; mandated by card brands.

Primary sourcepcisecuritystandards.org; current version is 4.0.1.

AcronymEU AI Act

ExpansionEU Artificial Intelligence Act

One-line definitionEU regulation classifying AI systems by risk and imposing obligations accordingly.

Primary sourceRegulation (EU) 2024/1689; eur-lex.europa.eu.

AcronymNIST AI RMF

ExpansionNIST AI Risk Management Framework

One-line definitionUS voluntary framework for managing AI-related risks across the lifecycle.

Primary sourcenist.gov/itl/ai-risk-management-framework; AI RMF 1.0 released January 2023.

AcronymNIST CSF

ExpansionNIST Cybersecurity Framework

One-line definitionVoluntary framework for managing cybersecurity risk; widely adopted.

Primary sourcenist.gov/cyberframework; version 2.0 released February 2024.

A warning about acronyms that look stable but aren't

Three honest cautions before you trust any acronym list. First, expansions drift: 'AI agent' in 2022 meant something different than 'AI agent' in 2026, and the same will be true for 2028. Second, ownership matters: MCP is Anthropic's, A2A is Google's, and the political economy of which protocol wins is not separable from which company is pushing it. Third, benchmark acronyms decay fastest of all — by the time MMLU was a household term among AI buyers, it was already saturated for frontier models, and anyone still selling on MMLU scores in 2026 is selling you something else. The acronyms in this page that have the longest shelf life are the ones tied to regulations (GDPR, HIPAA, SOC 2) and to fundamental math (MSE, KL, F1). The ones tied to specific architectures, methods, and benchmarks rotate on a roughly two-year cycle. Use this page as a decoder, not as a stable reference — and when you cite a number, cite the live source, not the acronym.

Sources

[01]
Vaswani et al. 2017 'Attention Is All You Need' introduces the transformer architecture that underlies modern LLMs.
arxiv.org/abs/1706.03762
[02]
Christiano et al. 2017 established the foundational RLHF technique of training reward models from human preference comparisons.
arxiv.org/abs/1706.03741
[03]
Ouyang et al. 2022 (InstructGPT) applied RLHF to instruction-following LLMs and became the template for production fine-tuning.
arxiv.org/abs/2203.02155
[04]
Rafailov et al. 2023 introduced Direct Preference Optimization (DPO) as a closed-form alternative to RLHF.
arxiv.org/abs/2305.18290
[05]
Ethayarajh et al. 2024 introduced KTO using prospect-theory-style utility for preference tuning.
arxiv.org/abs/2402.01306
[06]
Hong et al. 2024 introduced ORPO, combining SFT and preference optimization in a single step.
arxiv.org/abs/2403.07691
[07]
Shao et al. 2024 (DeepSeekMath) introduced GRPO, the group-relative policy optimization method later used by DeepSeek reasoning models.
arxiv.org/abs/2402.03300
[08]
Schulman et al. 2017 introduced PPO, the policy gradient method widely used in RLHF.
arxiv.org/abs/1707.06347
[09]
Bai et al. 2022 introduced Anthropic's Constitutional AI method using AI feedback against a written constitution.
arxiv.org/abs/2212.08073
[10]
Hu et al. 2021 introduced LoRA, the low-rank adaptation method that became the standard for efficient fine-tuning.
arxiv.org/abs/2106.09685
[11]
Dettmers et al. 2023 introduced QLoRA, combining 4-bit quantization with LoRA.
arxiv.org/abs/2305.14314
[12]
Wei et al. 2022 introduced chain-of-thought prompting, demonstrating that reasoning steps improve LLM performance on complex tasks.
arxiv.org/abs/2201.11903
[13]
Brown et al. 2020 (GPT-3 paper) introduced in-context learning as an emergent capability of large language models.
arxiv.org/abs/2005.14165
[14]
Snell et al. 2024 formalized the test-time compute scaling regime and its tradeoffs against model size.
arxiv.org/abs/2408.03314
[15]
Shazeer et al. 2017 introduced the sparsely-gated mixture-of-experts layer that underpins modern MoE architectures.
arxiv.org/abs/1701.06538
[16]
Lewis et al. 2020 introduced retrieval-augmented generation (RAG) combining parametric and non-parametric memory.
arxiv.org/abs/2005.11401
[17]
Radford et al. 2021 introduced CLIP, the contrastive image-text model that powers most modern multimodal systems.
arxiv.org/abs/2103.00020
[18]
Dosovitskiy et al. 2020 introduced the Vision Transformer (ViT) applying transformers to image patches.
arxiv.org/abs/2010.11929
[19]
Goodfellow et al. 2014 introduced generative adversarial networks (GANs).
arxiv.org/abs/1406.2661
[20]
Hendrycks et al. 2020 introduced MMLU, the multitask language understanding benchmark covering 57 academic subjects.
arxiv.org/abs/2009.03300
[21]
Jimenez et al. 2023 introduced SWE-Bench, evaluating LLMs on real GitHub issues from popular Python repos.
arxiv.org/abs/2310.06770
[22]
Model Context Protocol (MCP) is Anthropic's open standard for connecting AI models to data sources and tools, announced November 2024.
modelcontextprotocol.io
[23]
Regulation (EU) 2016/679 is the official text of the General Data Protection Regulation (GDPR).
eur-lex.europa.eu/eli/reg/2016/679/oj
[24]
The US Department of Health and Human Services is the authoritative source for HIPAA regulations and the Business Associate Agreement requirements.
hhs.gov/hipaa
[25]
NIST published the AI Risk Management Framework 1.0 in January 2023 as a voluntary US framework for AI risk.
nist.gov/itl/ai-risk-management-framework

Keep reading

Learn · plain-language AI →Research · papers and disclosures →Decode · jargon library →Tools · the operator surface →OrangeBox · local AI for builders →B00kMakor · book-shaped output →Vs · honest comparisons →