2020s
844 words first used in this period, under the year each was born.
2020
- AI copyright litigation
- mechanistic interpretability
- physical AI
- retrieval-augmented generation
- scaling laws
- AlphaFold
- best-of-n sampling
- circuit
- completion
- data contamination
- data mixture
- dense retrieval
- diffusion sampler
- digital replica
- EleutherAI
- factuality
- faithfulness
- feature (interpretability)
- few-shot prompting
- GPT-3
- in-context learning
- MMLU
- model editing
- neural radiance field
- polysemanticity
- risk-based approach
- role prompting
- scaling hypothesis
- small language model
- sparse retrieval
- superposition
- training data extraction
- usage policy
- vision transformer
- win rate
- 3D parallelism
- AI Incident Database
- biological anchors
- Books3
- ColBERT
- concept bottleneck model
- Conformer
- decentralized training
- DeepSpeed
- denoising diffusion implicit model
- denoising diffusion probabilistic model
- DETR
- Global Partnership on Artificial Intelligence
- goodput
- hardware lottery
- iterative nullspace projection
- late interaction
- linear attention
- logit lens
- maximal update parametrization
- MLCommons
- National AI Research Resource
- neural collapse
- N:M sparsity
- octopus test
- open-vocabulary detection
- parametric knowledge
- patch embedding
- prompt optimization
- QK-norm
- router
- sharpness-aware minimization
- shortcut learning
- sliding window attention
- sponge example
- SwiGLU
- The Pile
- universality hypothesis
- VQGAN
- windfall clause
2021
- AI watermarking
- Anthropic
- copilot
- GitHub Copilot
- high-risk AI system
- prompt engineering
- stochastic parrot
- sycophancy
- AI assistant
- AI coding assistant
- AI regulatory sandbox
- Character.AI
- citation
- classifier-free guidance
- CLIP
- code generation
- Codex
- conformity assessment
- Content Credentials
- DALL·E
- data-centric AI
- dead internet theory
- defense in depth
- evaluation harness
- foundation model
- general-purpose AI
- GSM8K
- helpful, honest and harmless
- homogenization
- honesty
- HumanEval
- induction head
- instruction dataset
- instruction tuning
- instruction-tuned model
- LAION
- latent diffusion model
- low-rank adaptation
- MATH
- pass@k
- prohibited AI practices
- prompt chaining
- prompt sensitivity
- prompt template
- provider and deployer
- remote biometric identification
- residual stream
- scheming
- scratchpad
- serious incident
- social scoring
- test-time compute
- transparency obligations
- truthfulness
- TruthfulQA
- UNESCO Recommendation on the Ethics of AI
- verifier
- AI bug bounty
- ALiBi
- Alignment Research Center
- automated employment decision tool
- BEIR
- biometric categorisation
- byte-level model
- C2PA
- classifier guidance
- CLIP score
- context distillation
- cross-embodiment learning
- decision transformer
- diffusion language model
- DINO
- direct logit attribution
- Distributed AI Research Institute
- edge of stability
- eliciting latent knowledge
- emotion recognition system
- explosive growth
- fully sharded data parallel
- grokking
- intended purpose
- interpretability illusion
- inverse scaling
- it's over, we're so back
- knowledge neuron
- masked autoencoder
- miniF2F
- MiniMax (company)
- MITRE ATLAS
- multimodal neuron
- natural abstraction hypothesis
- neural audio codec
- post-market monitoring
- prefix tuning
- privileged basis
- prompt tuning
- prover–verifier game
- QK and OV circuits
- Redwood Research
- reward is enough
- rotary positional embedding
- RWKV
- sandwiching
- sequence parallelism
- state space model
- substantial modification
- Torment Nexus
- Transformer Circuits Thread
- unembedding
- WebGPT
2022
- AI agent
- AI slop
- chain of thought
- ChatGPT
- export controls on AI chips
- jailbreak
- knowledge cutoff
- Midjourney
- prompt injection
- refusal
- Stable Diffusion
- agent framework
- agent loop
- AI constitution
- AI-generated content labeling
- audio language model
- automated red teaming
- BIG-bench
- Blueprint for an AI Bill of Rights
- Center for AI Safety
- chain-of-thought monitoring
- chat model
- Chinchilla scaling laws
- CHIPS and Science Act
- code execution
- compute-optimal training
- Constitutional AI
- continuous batching
- diffusion transformer
- DreamBooth
- effective accelerationism
- emergent abilities
- Epoch AI
- FlashAttention
- flow matching
- FP8
- goal misgeneralization
- Gymnasium
- H100
- HELM
- Imagen
- InstructGPT
- instruction following
- instruction-based image editing
- LangChain
- METR
- model merging
- MTEB
- negative prompt
- opt-out
- Perplexity AI
- prefill
- ReAct
- reasoning trace
- reinforcement learning from AI feedback
- reward overoptimization
- safety training
- scaling wall
- self-consistency
- self-improvement
- shoggoth
- situational awareness
- sparse autoencoder
- speculative decoding
- supervised fine-tuning
- system card
- tokens per parameter
- tool
- tool use
- Whisper
- activation patching
- AI hands
- algorithm filing
- causal scrubbing
- causal tracing
- Cicero
- concept erasure
- contrast-consistent search
- ConvNeXt
- DAN
- deep synthesis
- diffusion distillation
- Digital Services Act
- fill-in-the-middle
- Galactica
- Gato
- GPTQ
- helpful-only model
- hybrid architecture
- HyDE
- indirect object identification
- Inflection AI
- IsoFLOP analysis
- ISO/IEC 22989
- joint-embedding predictive architecture
- LAION-5B
- LaMDA sentience controversy
- least-to-most prompting
- linear representation hypothesis
- LlamaIndex
- LLM.int8()
- Matryoshka embedding
- mechanistic anomaly detection
- MEMIT
- model FLOPs utilization
- model soup
- monosemanticity
- MRKL system
- Othello-GPT
- outcome reward model
- outcome supervision
- PaLM
- path patching
- process reward model
- process supervision
- program-aided language model
- prompt leaking
- PyTorch Foundation
- rectified flow
- ROME
- safetensors
- safetywashing
- score distillation sampling
- Self-Instruct
- self-taught reasoner
- shard theory
- sharp left turn
- simulator theory
- SmoothQuant
- sparse upcycling
- steering vector
- structured access
- tabular foundation model
- task arithmetic
- textual inversion
- The Stack
- training gaming
- TransformerLens
- Turing trap
- z-loss
2023
- agent memory
- agentic AI
- AI Safety Institute
- AI Safety Summit
- Claude
- DeepSeek
- frontier AI
- frontier model
- Gemini
- general-purpose AI model
- Google DeepMind
- guardrails
- Llama
- LMArena
- open-weight model
- post-training
- SWE-bench
- xAI
- activation steering
- AI control
- AI deception
- AI engineer
- AI engineering
- AI factory
- AI pause
- AutoGPT
- Bletchley Declaration
- capability threshold
- CBRN risk
- ChatGPT moment
- compute threshold
- ControlNet
- dangerous capability evaluation
- direct preference optimization
- Executive Order 14110
- frontier lab
- function calling
- G7 Hiroshima AI Process
- GAIA
- Gaussian splatting
- general-purpose AI model with systemic risk
- generative engine optimization
- GPQA
- GPT-4
- Grok
- indirect prompt injection
- Interim Measures for the Management of Generative AI Services
- ISO/IEC 42001
- jagged frontier
- llama.cpp
- LLM-as-a-judge
- LLM-isms
- lost in the middle
- Mistral AI
- MMMU
- moat
- model collapse
- model routing
- model weight security
- Moonshot AI
- multi-agent system
- multimodal large language model
- needle in a haystack
- NIST AI Risk Management Framework
- NO FAKES Act
- Ollama
- OpenAI leadership crisis (2023)
- OpenAI-compatible API
- orchestration
- over-refusal
- overtraining
- OWASP Top 10 for LLM Applications
- Pause Giant AI Experiments
- plan-and-execute
- pre-deployment testing
- prompt caching
- Qwen
- responsible scaling policy
- rogue AI
- RT-2
- sandbagging
- scaffolding
- Segment Anything Model
- self-correction
- sovereign AI
- Statement on AI Risk
- Suno
- superalignment
- Sydney
- SynthID
- system prompt
- third-party AI evaluation
- time to first token
- tool call
- vision-language-action model
- vLLM
- WebArena
- woke AI
- wrapper
- action chunking
- activation addition
- agent swarm
- AGI has been achieved internally
- AI gateway
- AI lie detection
- AI Pact
- AI Safety Level
- AlgoPerf
- Apollo Research
- attention sink
- attribution patching
- augmented language model
- AutoGen
- automated circuit discovery
- automated interpretability
- autonomous replication and adaptation
- AWQ
- BabyAGI
- Balenciaga Pope
- BitNet
- blurry JPEG of the web
- capability elicitation
- chain-of-thought faithfulness
- chat template
- ChatGPT plugins
- chunked prefill
- consistency model
- context extension
- context parallelism
- contrastive activation addition
- control evaluation
- CrewAI
- Cursor
- data-constrained scaling
- developmental interpretability
- diffusion policy
- DiLoCo
- disaggregated serving
- distributed alignment search
- DSPy
- dual-use foundation model
- effective compute
- English is the hottest new programming language
- enumerative safety
- excessive agency
- ExecuTorch
- exploration hacking
- feature splitting
- feel the AGI
- fine-tuning attack
- FP4
- Frontier Model Forum
- function vector
- fundamental rights impact assessment
- FunSearch
- generative agent
- generative recommendation
- GGUF
- Glaze
- glitch token
- GPU poor
- grandma exploit
- grouped-query attention
- GUI agent
- Heart on My Sleeve
- High-level Advisory Body on Artificial Intelligence
- IAEA for AI
- IFEval
- in-context vector
- indicator properties of consciousness
- inference-time intervention
- ISO/IEC 23894
- JSON mode
- know-your-customer for compute
- KV cache compression
- LEACE
- levels of AGI
- Lion
- Llama Guard
- LLM observability
- LLMOps
- lobotomized
- local learning coefficient
- machine psychology
- Mamba
- MemGPT
- microscaling formats
- min-p sampling
- MLX
- model deprecation
- model organism of misalignment
- MT-Bench
- nerfed
- Neuronpedia
- New York Times v. OpenAI
- Nightshade
- notkilleveryoneism
- on-chip governance mechanisms
- on-policy distillation
- Open X-Embodiment
- PagedAttention
- PAIR
- PauseAI
- Preparedness Framework
- prompt management
- Q*
- QLoRA
- Reflexion
- representation engineering
- reversal curse
- self-exfiltration
- semantic entropy
- SGLang
- Sparks of AGI
- stop condition
- superficial alignment hypothesis
- TensorRT-LLM
- TESCREAL
- time per output token
- time series foundation model
- Toolformer
- tree of thoughts
- trusted editing
- trusted monitoring
- tuned lens
- universal adversarial suffix
- untrusted monitoring
- Waluigi effect
- weak-to-strong generalization
- Will Smith eating spaghetti
2024
- coding agent
- computer use
- deep research
- EU AI Act
- Model Context Protocol
- reasoning model
- agent instruction file
- agentic RAG
- agentic workflow
- AI Office
- AI red lines
- AI scientist
- AI whistleblower protections
- AIME
- alignment faking
- annealing
- Colorado AI Act
- country of geniuses in a datacenter
- delve
- FLUX
- Framework Convention on Artificial Intelligence
- Frontier AI Safety Commitments
- FrontierMath
- Gemma
- Global Digital Compact
- glue on pizza
- Golden Gate Claude
- GPT-4o
- GraphRAG
- group relative policy optimization
- International AI Safety Report
- jagged intelligence
- MCP server
- mid-training
- model specification
- Muon
- neocloud
- Nobel Prizes for AI
- one-shot
- OpenAI o1
- orchestrator–worker pattern
- OSWorld
- reinforcement fine-tuning
- reinforcement learning with verifiable rewards
- RL scaling
- Safe Superintelligence Inc.
- safety case
- SB 1047
- SimpleQA
- sleeper agent
- Sora
- strawberry problem
- structured outputs
- τ-bench
- uplift study
- Veo
- voice agent
- abliteration
- agent identity
- agent–computer interface
- AgentDojo
- AI Futures Project
- AI worm
- AI-assisted vulnerability discovery
- AlphaGeometry
- AlphaProof
- Bartz v. Anthropic
- Black Forest Labs
- botshit
- cache-augmented generation
- character training
- circuit breakers
- co-intelligence
- COLM
- compound AI system
- contamination-resistant benchmark
- contextual retrieval
- control vector
- crosscoder
- Cybench
- deliberative alignment
- Devin
- direct alignment algorithm
- directional ablation
- European Artificial Intelligence Board
- evals-driven development
- evaluator–optimizer
- feature absorption
- feature steering
- FineWeb
- Frontier Safety Framework
- Gemma Scope
- generalist robot policy
- generative reward model
- Goodfire
- guaranteed safe AI
- handoff
- HarmBench
- if-then commitment
- instruction hierarchy
- JailbreakBench
- Kahneman–Tversky optimization
- Kolmogorov–Arnold network
- LangGraph
- latent reasoning
- llms.txt
- Machines of Loving Grace
- many-shot jailbreaking
- marginal risk
- MLE-bench
- model diffing
- multi-head latent attention
- multi-token prediction
- multi-turn jailbreak
- OpenAI o3
- Patchscopes
- pelican on a bicycle
- personhood credential
- Physical Intelligence
- π0
- platonic representation hypothesis
- prefill attack
- Product Liability Directive
- PutnamBench
- rack-scale system
- RE-Bench
- refusal direction
- representation fine-tuning
- reverse acquihire
- RewardBench
- sabotage evaluation
- schedule-free optimization
- Situational Awareness (essay)
- sparse feature circuit
- spatial intelligence
- StrongREJECT
- SWE-agent
- transcoder
- unhobbling
- warmup-stable-decay schedule
- WMDP
- YOLO mode
- YOLO run
2025
- America's AI Action Plan
- GPT-5
- Humanity's Last Exam
- METR time horizon
- Stargate
- vibe coding
- agent skill
- Agent2Agent protocol
- agentic browser
- agentic coding
- agentic commerce
- agentic misalignment
- AI 2027
- AI as normal technology
- AI preemption
- AI psychosis
- background agent
- BrowseComp
- Center for AI Standards and Innovation
- Claude Code
- compaction
- context engineering
- context poisoning
- context rot
- DeepSeek moment
- DeepSeek-R1
- Digital Omnibus
- emergent misalignment
- evaluation awareness
- GDPval
- General-Purpose AI Code of Practice
- gpt-oss
- gradual disempowerment
- harness
- Independent International Scientific Panel on AI
- lethal trifecta
- Meta Superintelligence Labs
- Nano Banana
- OpenClaw
- permission system
- RAISE Act
- RL environment
- SB 53
- Software 3.0
- subagent
- TAKE IT DOWN Act
- Terminal-Bench
- thinking budget
- Thinking Machines Lab
- workslop
- Agentic AI Foundation
- agentic reinforcement learning
- AI Basic Act
- AI co-scientist
- AI Diffusion Rule
- AI Promotion Act
- AI-enabled coup
- alignment audit
- AlphaEvolve
- attribution graph
- benchmaxxing
- circuit tracing
- circular deal
- Claude Plays Pokémon
- code mode
- cognitive debt
- concept injection
- constitutional classifiers
- Darwin Gödel Machine
- entropy collapse
- Epoch Capabilities Index
- era of experience
- extended thinking
- Gemini Robotics
- Genesis Mission
- Ghiblification
- glazing
- Global Dialogue on AI Governance
- HealthBench
- hierarchical reasoning model
- hook
- humanist superintelligence
- hybrid reasoning model
- inoculation prompting
- intelligence curse
- interleaved thinking
- internal deployment
- Italian brainrot
- just-in-time retrieval
- LLM grooming
- MechaHitler
- model biology
- monitorability
- Mutual Assured AI Malfunction
- nested learning
- orbital data center
- parameter decomposition
- persona vector
- personal superintelligence
- Potemkin understanding
- progressive disclosure
- Ralph Wiggum loop
- recursive language model
- Remote Labor Index
- replacement model
- rubric reward
- safe completion
- SB 243
- seahorse emoji
- seemingly conscious AI
- sleep-time compute
- slopsquatting
- spec-driven development
- spiritual bliss attractor state
- Statement on Superintelligence
- subliminal learning
- superhuman coder
- Texas Responsible Artificial Intelligence Governance Act
- tiny recursive model
- tool poisoning
- Vending-Bench
- vibe engineering
- vibe hacking
- vibe porting
- Virology Capabilities Test
- world foundation model
- you're absolutely right
2026
- accidental cyberattack
- Executive Order 14434
- Hugging Face incident
- The Last Dictionary
- assistant axis
- claw
- dark factory
- decision model
- Deep Blue (malaise)
- ExploitGym
- export control directive
- externalization
- Magnifica Humanitas
- Moltbook
- Project Glasswing
- small AI
- swarm scaling
- tokenmaxxing
- World AI Cooperation Organization