AI Engineer 공간 "사부작 사부작"

Tags

• transformer
• Engineering at Anthropic
• catastrophic forgetting
• RLHF
• attention
• 트랜스포머
• SK AI SUMMIT 2025
• context engineering
• extended thinking
• test-time scaling
• fréchet inception distance
• chain-of-thought
• reinforcement learning from human feedback
• flashattention
• rotary position embedding
• model context protocol
• re-ranking
• gqa
• langgraph
• PEFT
• 토크나이저
• MQA
• Langchain
• CoT
• MHA
• BLEU
• Positional Encoding
• Multi-Head Attention
• Embedding
• self-attention
• tokenizer
• Rag
• reasoning
• prompt
• Lora
• FID
• LLM
• Rope
• context
• MCP
• clip
• benchmark
• Code execution with MCP: Building more efficient agents
• context rot
• attention budget
• Effective context engineering for AI agents
• Writing effective tools for agents
• Writing effective tools for agents — with agents
• 제9회 데보션(DEVOCEAN) 테크 데이
• NAVER Cloud AI DevDay
• 엔지니어링 밋업
• Agentic AI 핸즈온
• 프롬프트 캐시
• Time-to-First-Token
• TTFT
• 프롬프트 캐싱
• Prompt Caching
• Prompt Cache
• Train-time Scaling
• Multi Agents
• Deep Agents
• dynamic prompt
• custom prompt
• reasoning model
• checkpoints
• response metadata
• 스페셜 토큰
• special tokens
• subword tokenizer
• 서브워드 토크나이저
• pre-ln
• post-ln
• 레이어 정규화
• 학습률 예열
• learning rate warm-up
• http+sse
• fastmcp
• streamable http
• best-of-n
• best-of-n distillation
• budget forcing
• misaligned persona
• bad boy persona
• 창발적 오정렬
• 에이전트 오정렬
• 오정렬
• deceptive misalignment
• emergent misalignment
• agentic misalignment
• misalignment
• pad masking
• look-ahead masking
• skip link
• residual link
• swish gated linear unit
• element-wise multiplication
• swiglu
• silu
• implicit gating
• explicit gating
• 게이팅
• gating
• needle-in-a-haystack
• u자형 성능 곡선
• u-shaped performance curve
• reranker
• reranking
• lost in the middle
• mulit head attention
• multi-query attention
• grouped-query attention
• text-to-image retrieval
• recall@k
• coco captions
• visual question answering
• ablation study
• mechanistic interpretability
• inner activation exploration
• feature attribution
• attention visualization
• contrastive language-image pre-training
• humaneval
• hellaswag
• big-bench
• gsm8k
• 연속 배칭
• 정적 배칭
• pagedattention
• open neural network exchange
• io-awareness
• 타겟 모델
• 초안 모델
• 추측 디코딩
• speculative decoding
• halueval
• llm-as-a-judge
• inception score
• human evaluation
• automated metrics
• 에이전트 프레임워크
• ai agent framework
• reason + act
• cot 한계
• tree of thoughts
• 인 컨텍스트 학습
• int8
• bf16
• qat
• ptq
• gptq
• awq
• zero redundancy optimizer
• pytorch fully sharded data parallel
• fsdp
• deepspeed
• roformer
• semantic segmentation map
• human pose
• canny edges
• dpm-solver
• guidance scale
• 클래시파이어-프리 가이던스
• diffusion transformer
• 확산 트랜스포머
• conditional generation
• 잠재 확산 모델
• latent diffusion model
• iterative/recursive retrieval
• query transformation
• modular rag
• ragas
• rag score
• rag ratings
• unanswerability evaluation
• answer correctness
• answer relevancy
• ndcg
• context recall
• context precision
• hybrid search
• 하이브리드 검색
• semantic search
• 의미적 검색
• lexical search
• 어휘적 검색
• 파국적 망각
• 지속적 사전 훈련
• continuous pre-training
• weight averaging
• 가중치 평균
• 구면 선형 보간법
• 태스크 산술
• task arithmetic
• 모델 병합
• model merging
• tiesmerging
• quantized lora
• parameter-efficient fine-tuning
• reinforcement learning from verifiable rewards
• rlvr
• format rewards
• 형식 보상
• group relative policy optimization
• direct preference optimization
• reinforcement learning from ai feedback
• ai 피드백 기반 강화 학습
• 인간 피드백 기반 강화 학습
• rlaif
• 지도 파인튜닝
• 명령어 튜닝
• 거부 샘플링
• byte pair encoding
• 그룹-쿼리 어텐션 (grouped-query attention)
• 멀티-쿼리 어텐션 (multi-query attention)
• kv 캐싱
• kv cashing
• 창발적 능력
• emergent abilities
• 친칠라의 함정
• chinchilla trap
• scaling laws
• 스케일링 법칙
• low-rank approximation
• state space models
• efficient / linearized attention
• decoder-only
• encoder-only
• transformer architectures
• attention with linear biases
• 멀티 헤드 어텐션
• mixture-of-experts
• out-of-vocabulary
• oov
• 연속 학습
• kv cache
• 스토리 포인트
• 전문가 혼합
• mcp client
• grpo
• mcp server
• langchain-mcp-adapters
• claude code
• manus
• story point
• deep research
• textual inversion
• agents
• prompting
• instruction tuning
• deepseek-r1
• agentic ai
• slerp
• ddim
• deepseek
• qlora
• reward hacking
• adamw
• low-rank adaptation
• supervised fine-tuning
• retrieval-augmented generation
• truthfulqa
• mmlu
• vllm
• llm 한계
• batching
• llama3
• 포지셔널 인코딩
• rejection sampling
• 데보션
• mode collapse
• sparse attention
• Retrieval
• LLaVa
• positional embedding
• Multimodal
• HyperClova X
• Benchmarks
• context window
• Chinchilla
• Classifier-Free Guidance
• Diffusion Models
• ddpm
• Layer Normalization
• BPE
• 셀프 어텐션
• claude
• DreamBooth
• controlnet
• Fine-tuning
• Transfomer
• residual
• encoder-decoder
• bge
• DEVOCEAN
• fp16
• contrastive learning
• in-context learning
• Inductive bias
• 동적 프롬프트
• NAVER CLOUD PLATFORM
• challenges
• VQA
• 임베딩
• onnx
• 어텐션
• Observation
• sft
• NCP
• BM25
• Skip Connection
• Residual Connection
• image captioning
• Overfitting
• masking
• continual learning
• dpo
• TensorRT
• QUANTIZATION
• bias
• alibi
• TF-IDF
• depth
• flamingo
• 양자화
• fibonacci
• MOE
• stdio
• Continuous
• react
• T5
• 밋업
• Docker
• Add-on
• 프롬프트
• llama
• Scribble
• 잔차
• KAPLAN
• Dare
• mrr
• LDM
• 마스킹
• thinking
• cfg
• fairness
• Bert
• Cider
• TOT
• 피보나치
• container
• SSE
• Thought
• Training
• bond
• DRAFT
• rouge
• metadata
• Warm-up
• position
• FT
• Fragmentation
• dynamic
• jira
• DROP
• Factory
• Token
• sandbox
• Action
• DIT
• gaming
• PATCH
• padding
• AWS
• math
• ssm
• Spice
• server
• GPT
• bon
• HTTP
• calibration
• target
• FUSION
• ICL
• zero
• hyde
• static
• pattern
• 벤치마크
• retriever
• ARC
• Client
• 패딩
• Think
• IS