
çæAIåºç€â¡ LLMçšèªé
ã¯ããã«
ååèšäºã¯ãã¡ãã
ä»åã¯LLMé¢é£ã®çšèªéãå
¬éããŸãã
LLMã«ã€ããŠæŽçããèæ¯
é¢é£æ
å ±ã«å¯Ÿããè§£å床ãäžããã
èšäºã®äœçœ®ã¥ã
åºç€/æŠèŠãæŽçããããã®ã¡ã¢
ä»çµã¿ã«èå³ããã人åã
å©çšããã¢ãã« (ä¿¡é Œæ§ã®åèçš)
ä»åã¯äž»ã«Gemini 3 (Thinking) ã䜿çšããŠæšæ²ããŠããŸãã
(ä¿¡é Œæ§ãäžããããã«çšèªæ¯ã«URLæç€ºããããšãæ€èšããŸããããèšäºã®å°ç·ãèæ
®ããçµæãããŸãã)
1. åºç€æ§é ãšä»çµã¿
ããŒã¯ã³ (Token): ã¢ãã«ãåŠçããæå°åäœãåèªãæåã®æçãæéãåŠçé (Context Window) ã®åºæºã
次ããŒã¯ã³äºæž¬ (Next Token Prediction): LLMã®åºæ¬åäœããæèã«ç¶ãæã確çã®é«ã次ã®åèªããäºæž¬ãç¶ããã
åã蟌㿠(Embedding): ããã¹ãã倿¬¡å ã®æ°å€ïŒãã¯ãã«ïŒã«å€æããæå³ã®è¿ããèšç®å¯èœã«ããæè¡ã
Transformer: çŸåšã®LLMã®åºç€ãšãªããã¥ãŒã©ã«ãããã¯ãŒã¯æ§é ã䞊ååŠçãšé·è·é¢ã®æèçè§£ã«åªããã
Attention (Self-Attention): æç« å ã®ãã©ã®åèªãéèŠãããèšç®ãããæèçè§£ã®æ žå¿ã¡ã«ããºã ã
ãã©ã¡ãŒã¿ / éã¿ (Parameters / Weights): ã¢ãã«å éšã®èª¿æŽå€ãæ°ãå€ãã»ã©è€éãªç¥èãä¿æã§ããããèšç®ã³ã¹ããå¢ãã
ã¹ã±ãŒãªã³ã°å (Scaling Laws): ããŒã¿ã®éãèšç®éãã¢ãã«ãµã€ãºã倧ããããã»ã©ãæ§èœãäºæž¬å¯èœã«åäžãããšããæ³åã
ãã«ãã¢ãŒãã« (Multimodal): ããã¹ãã»ç»åã»é³å£°ãªã©ãçµ±åããŠçè§£ã»çæããèœåãçŸåšã¯åäžã¢ãã«ã§å šæ å ±ãçŽæ¥åŠçããããã€ãã£ããã«ãã¢ãŒãã«ããäž»æµã
2. åŠç¿ã®ããã»ã¹
äºååŠç¿ (Pre-training): èšå€§ãªããŒã¿ããèšèªã®åºç€ãåŠã¶ãåºç€ã¢ãã« (Foundation Model) äœæã®ç¬¬äžæ®µéã
ããŒã¿ã¯ã¬ã³ãžã³ã° / éè€æé€ (Data Cleansing / Deduplication): åŠç¿ããŒã¿ãããã€ãºãéè€ãé€ããã¢ãã«ã®è³ªãšå¹çãé«ããäœæ¥ã
æç€ºèª¿æŽ (SFT: Supervised Fine-Tuning): ã質åã«çããããªã©ã®å¯Ÿè©±åœ¢åŒã«é©å¿ãããããã®è¿œå åŠç¿ã
RLHF (Reinforcement Learning from Human Feedback) /
DPO (Direct Preference Optimization): 人éã®ãã£ãŒãããã¯ã«åºã¥ããåçã®å質ãå®å šæ§ãé«ããåŸ®èª¿æŽææ³ã2026幎çŸåšã¯ãããå¹ççãªDPOãäž»æµãPEFT(Parameter-Efficient Fine-Tuning) /
LoRA (Low-Rank Adaptation): äžéšã®ãã©ã¡ãŒã¿ã®ã¿ãæŽæ°ããäœã³ã¹ãã»çæéã§è¿œå åŠç¿ãè¡ãææ³ãåæããŒã¿ (Synthetic Data): AIèªèº«ãçæããåŠç¿çšããŒã¿ã2026幎çŸåšã¯ãé«å質ãªãæšè«éçšããå«ãåæããŒã¿ãã¢ãã«æ§èœã®é£èºã«è²¢ç®ã
ã¢ãããŒã·ã§ã³ (Annotation): ããŒã¿ã«å¯ŸããŠãæ£è§£ã©ãã«ããªã©ãä»äžããAIãåŠç¿ã§ããããã«ããäœæ¥ã
3. äž»ãªã¢ãã«
ç§æ (Proprietary)
GPT (OpenAI): æ¥çæšæºã2026幎çŸåšã¯ãGPT-5ãäžä»£ãäž»æµãšãªããé«åºŠãªæšè«èœåãšå®ç§ãªãã«ãã¢ãŒãã« (Multimodal) æ©èœãçµ±åã
Claude (Anthropic): é«ãå®å šæ§ãšã人éã«è¿ãèªç¶ãªæç« 衚çŸãç¹åŸŽãå«çæ§èœãé«ããæ°çŸäžããŒã¯ã³ã®å ¥åã«å¯Ÿå¿ã
Gemini (Google): ãã€ãã£ããã«ãã¢ãŒãã«ããããªã»é³å£°ã»ã³ãŒãã®çµ±åçè§£ã«åªãã1,000äžããŒã¯ã³ãè¶ ããè¶ é·æã³ã³ããã¹ã (Long Context) åŠçã匷ã¿ã
ãªãŒãã³ãœãŒã¹ (Open Weights)
Llama (Meta): ãªãŒãã³ãœãŒã¹ã®ä»£è¡šæ Œã2026幎çŸåšã¯Llama 5ãç»å Žããåçšã¢ãã«ã«å¹æµããæ§èœããªã³ãã¬ãã¹ã§å®çŸã
Mistral: 欧å·çºãMoE (æ··åå°éå®¶) æè¡ã«ããé«å¹çãªæšè«ãšãé«ãã«ã¹ã¿ãã€ãºæ§ãç¹åŸŽã
DeepSeek: äžåœçºãå§åçãªã³ã¹ããèªããç¹ã«æ°åŠãã³ãŒãçã®çæ°ç³»æšè«ã«ãããŠäžçãããçŽã®æ§èœãçºæ®ã
Qwen (Alibaba): ã¢ãžã¢åã®èšèªã«åŒ·ããç§åŠæè¡åéã®å°éç¥èŠãšé«åºŠãªãã«ãã¢ãŒãã«æ§èœãå Œåã
Grok (xAI): Xã®ãªã¢ã«ã¿ã€ã æ å ±ãå³åº§ã«åæ ãæ å ±ã®å§åçãªé®®åºŠãšãç¬èªã®ãŠã£ããã«å¯ãã åçã¹ã¿ã€ã«ãæŠåšãäžéšã¢ãã«ã¯ãªãŒãã³ãœãŒã¹åã
SLM (Small Language Models)
(Phi, Gemmaãªã©): ç¹å®ã¿ã¹ã¯ã«ç¹åããã¹ãããPCäžã§é«éåäœããå°èŠæš¡ã¢ãã«ããªã³ããã€ã¹AIã®æ®åã«ããéèŠæ§ãå¢å€§ã
4. æšè«ã®æ°ãããã©ãã€ã
æšè«ã¢ãã« (Reasoning Models): OpenAIã®o1/o3ã·ãªãŒãºã®ããã«ãå éšçã«ãæèã®é£éããè¡ããæ°åŠãç§åŠã®é£åãè«ççã«è§£ãã¢ãã«ã
System 1 vs System 2 Thinking: çŽæçãªå³ç (System 1: åŸæ¥ã®LLM) ãšãæéããããŠçèããåç (System 2: æšè«ã¢ãã«) ã®äœ¿ãåãã
äžçã¢ãã« (World Models): ç©çæ³åãç©ºéæŠå¿µãçè§£ããåç»çæããããã£ã¯ã¹ãã·ãã¥ã¬ãŒã·ã§ã³ã«å¿çšãããæŠå¿µã
åµçºçèœå (Emergent Abilities): ã¢ãã«ãããäžå®ã®èŠæš¡ãè¶ ããéãæç€ºçã«åŠç¿ããŠããªãè€éãªèœåãçªç¶çŸããçŸè±¡ã
5. ããã³ãããšçæå¶åŸ¡
Context Window: ã¢ãã«ãäžåºŠã«ä¿æã§ããæ å ±éã2026幎ã«ã¯ãç¡éãã«è¿ãé·å€§ãªå ¥åãæ±ãæè¡ãäžè¬åã
n-shot Prompting: äŸç€ºãªã (Zero-shot) ããè€æ°ã®äŸç€º (Few-shot) ãäžããŠåçãå®å®ãããææ³ã
CoT (Chain of Thought): ãstep by stepã§èããŠããšä¿ããè€éãªæšè«ãè§£ãããææ³ã
æ§é ååºå (Structured Output): JSONçã®ç¹å®åœ¢åŒã§åºåãããã·ã¹ãã 飿ºã確å®ã«ããæè¡ã
çæãã©ã¡ãŒã¿ã®æ¯èŒ
Temperatureã
圹å²:ã©ã³ãã æ§ã®å¶åŸ¡
èšå®ã®ãã³ã: 0ã«è¿ã¥ãããšå
å®ã1ã«è¿ã¥ãããšåµé çã«ãªãã
Top-p / Top-kã
圹å²: åè£ã®çµã蟌ã¿
èšå®ã®ãã³ã: 確çã®é«ãåèªã ãã«éžæè¢ãå¶éããåçã®ç Žç¶»ãé²ãã
6. RAGãšå€éšé£æº
RAG (Retrieval-Augmented Generation): å€éšç¥èãæ€çŽ¢ãããã®å 容ãå ã«åçããä»çµã¿ããã«ã·ããŒã·ã§ã³æå¶ã«äžå¯æ¬ ã
ãã¯ãã«ããŒã¿ããŒã¹ (Vector Database): æå³ãè¿ãããŒã¿ãé«éæ€çŽ¢ã§ãããããåã蟌㿠(Embedding) ãä¿åããããšã«ç¹åããDBã
ãã£ã³ãã³ã° (Chunking): é·ãææžãRAGã§æ±ããããé©åãªãµã€ãº (Chunk) ã«åå²ããåŠçã
AIãšãŒãžã§ã³ã (AI Agents): ããŒã«åŒã³åºã (Function Calling) ã䜿ããèªåŸçã«ã¿ã¹ã¯ (æ€çŽ¢ãå®è¡ãäºçŽç) ãéè¡ããã·ã¹ãã ã
MCP (Model Context Protocol): ã¢ãã«ãšå€éšããŒã« (DBããã¡ã€ã«ãã¢ããªç) ãå®å šãã€æšæºçã«æ¥ç¶ããããã®2026å¹Žã®æšæºèŠæ Œã
RAG vs ãã¡ã€ã³ãã¥ãŒãã³ã°
RAG
äž»ãªç®ç:ææ°ã»å€éšæ
å ±ã®åç
§ã
ã³ã¹ã: äœ Â (æ¢åã¢ãã«+DB)
ç¥èã®æŽæ°: ããŒã¿å
¥ãæ¿ãã®ã¿
ãã¡ã€ã³ãã¥ãŒãã³ã°
äž»ãªç®ç:å°éçãªå£èª¿ã»åœ¢åŒã®ç¿åŸ
ã³ã¹ã: é« Â (远å ã®èšç®è³æºãå¿
èŠ)
ç¥èã®æŽæ°: ã¢ãã«ã®ååŠç¿ãå¿
èŠ
7. æé©åãšéçš
ãã«ã·ããŒã·ã§ã³ (Hallucination): ã¢ãã«ãäºå®ã«åºã¥ããªãããã£ãšããããåãã€ãçŸè±¡ã
ã°ã©ãŠã³ãã£ã³ã° (Grounding): åçãæ ¹æ (å€éšããŒã¿) ã«åºã¥ãããä¿¡é Œæ§ãé«ããããšã
éåå (Quantization): æ°å€ã®ç²ŸåºŠãèœãšããŠããŒã¿å®¹éãåæžããå®äŸ¡ãªGPUãã¹ããã§åãããããã«ããæè¡ã
ç¥èèžç (Knowledge Distillation): å·šå€§ãªæåž«ã¢ãã«ã®ç¥èãã軜éãªçåŸã¢ãã«ã«ç¶æ¿ãããæè¡ã
ææ©çãã³ãŒã (Speculative Decoding): å°ããªã¢ãã«ã§ãäžæžãããé«éçæãã倧ããªã¢ãã«ã§æ€èšŒããããšã§çæé床ãåçã«äžããææ³ã
LLMOps: LLMã·ã¹ãã ãç¶ç¶çã«ç£èŠã»è©äŸ¡ã»æ¹åã»éçšããããã®ã©ã€ããµã€ã¯ã«ç®¡çã
8. å®å šæ§ãšãªã¹ã¯ç®¡ç
ã¬ãŒãã¬ãŒã« (Guardrails): æå®³ãªåºåãäžæ£æäœãé²ãããã®ãªã¢ã«ã¿ã€ã ã»ãã£ã«ã¿ãŒã
ããã³ããã€ã³ãžã§ã¯ã·ã§ã³ (Prompt Injection): æªæã®ããå ¥åã§ã¢ãã«ã®å¶çŽãåé¿ããæ©å¯æ å ±ãçãæ»æã
è±ç (Jailbreak): å«çãã£ã«ã¿ãŒãå·§åŠã«çªç ŽããçŠæ¢ãããã³ã³ãã³ããçæããã詊ã¿ã
PII (Personally Identifiable Information): å人è奿 å ±ãæŒæŽ©ãé²ãããã®èªåãã¹ãã³ã°å¯Ÿçãå¿ é ã
責任ããAI (Responsible AI): å ¬å¹³æ§ãä¿¡é Œæ§ãéææ§ãæ ä¿ããããã®éçºã»éçšæéã
9. çæã®åŠçããã»ã¹
å ¥å (Input): ãŠãŒã¶ãŒãããã³ãããå ¥åããã
ååŠç (Pre-processing): ããŒã¯ãã€ã¶ãŒãåå²ããåã蟌㿠(Embedding) ã§æ°å€åããã
æ¡åŒµ (Augmentation): (RAGã®å Žå) ãã¯ãã«DBããé¢é£æ å ±ãæ€çŽ¢ãã³ã³ããã¹ãã«è¿œå ããã
æšè« (Inference): Transformerå ã®Attentionãæèãèªã¿ã次ããŒã¯ã³äºæž¬ (Next Token Prediction) ãè¡ãã
å¶åŸ¡ (Decoding Control): Temperatureçã®èšå®ã«åºã¥ãã確çååžããæé©ãªåèªãéžæããã
åºå (Output): ãŠãŒã¶ãŒã«åçã衚瀺ãããã