[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7AK4ZCRDPFDFGDA1RGSC4YA":3,"skill-same:01M7AK4ZCRDPFDFGDA1RGSC4YA":48},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":29,"ownerHtmlUrl":30,"ownerBlogUrl":31,"ownerDescription":32,"updatedAt":33,"install":34,"resources":47},"01M7AK4ZCRDPFDFGDA1RGSC4YA","ai-ml-landscape",null,"Comprehensive practitioner reference for the 2025–2026 AI\u002FML landscape covering frontier model selection and routing, open-weight vs hosted API trade-offs, reasoning models and test-time compute, RAG and agent production patterns, classical ML for tabular data, deep learning foundations (transformers, attention, MoE, SSMs), LLM training and post-training (DPO\u002FGRPO\u002FLoRA\u002FQLoRA), inference serving (vLLM\u002FSGLang), evaluation benchmark skepticism, safety\u002Falignment, EU AI Act governance, and hardware selection. Use when advising on model selection, AI architecture decisions, LLM deployment, fine-tuning strategy, benchmark interpretation, or AI governance compliance.","the-vibey-project\u002Fvibey","the-vibey-project","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F329298758?v=4","01M7AJHGZ9722CTJ4JNQZVRH0Q","the vibey project","01M7AJHGZVMG2K7TGXQ6XFHVG3",5,"MIT","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fai-and-data\u002Fskills\u002Fai-ml-landscape\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: ai-ml-landscape\ndescription: \"Comprehensive practitioner reference for the 2025–2026 AI\u002FML landscape covering frontier model selection and routing, open-weight vs hosted API trade-offs, reasoning models and test-time compute, RAG and agent production patterns, classical ML for tabular data, deep learning foundations (transformers, attention, MoE, SSMs), LLM training and post-training (DPO\u002FGRPO\u002FLoRA\u002FQLoRA), inference serving (vLLM\u002FSGLang), evaluation benchmark skepticism, safety\u002Falignment, EU AI Act governance, and hardware selection. Use when advising on model selection, AI architecture decisions, LLM deployment, fine-tuning strategy, benchmark interpretation, or AI governance compliance.\"\n---\n\n# AI & Machine Learning Landscape — Practitioner Reference (2026)\n\n## The Three-Layer Structure of AI in 2026\n\n1. **Frontier model race** — fast-moving, narrow quality gaps between top models (~2.7% lead at the top as of March 2026)\n2. **Production engineering stack** — maturing: vLLM\u002FSGLang serving, MCP-standardized agents, contextual-retrieval RAG\n3. **Classical ML core** — stable: gradient boosting still wins most tabular problems\n\n**Inference cost collapse**: >280-fold reduction from $20.00 to $0.07 per million tokens between Nov 2022 and Oct 2024 (GPT-3.5-equivalent quality, per Stanford HAI). This makes capability cheap and **engineering discipline** the binding constraint.\n\n---\n\n## Frontier Model Landscape (Mid-2026)\n\nNo single model dominates. The top model leads by only ~2.7% (Stanford HAI AI Index, March 2026). The leading closed-weight model's edge over the top open-weight model on Chatbot Arena narrowed from 8.04% (Jan 2024) to 1.70% (Feb 2025).\n\n**Key practitioner principle:** Build behind an abstraction layer (LiteLLM\u002FOpenRouter) to avoid lock-in. Model names and rankings shift weekly.\n\n---\n\n## Hosted APIs vs Open Weights — Decision Rule\n\n| Factor | Hosted API | Open Weights |\n|---|---|---|\n| Peak capability, zero ops | ✓ | |\n| Privacy, air-gap, data residency | | ✓ |\n| Fine-tuning control | | ✓ |\n| Predictable cost at sustained high QPS | | ✓ |\n| Fastest to start | ✓ | |\n\n**Open-weight options**: DeepSeek-V4 (MIT-licensed, 1M context, 1.6T\u002F49B active MoE), Qwen3, Meta Llama 4, Mistral, IBM Granite 4.0.\n\n**Threshold to self-host**: sustained high QPS where API spend exceeds fully-loaded GPU fleet cost, or a hard data-residency requirement.\n\n---\n\n## Model Selection Strategy (Staged)\n\n1. **Start with hosted frontier behind an abstraction layer.** Default to strong mid-tier (Claude Sonnet-class, GPT-4.1-mini-class, Gemini Flash-class); escalate to top-tier only for hard tasks.\n\n2. **Implement model routing** once volume justifies it. A cheap classifier sends easy queries to small\u002Fcheap models with confidence-gated escalation. RouteLLM (UC Berkeley\u002FAnyscale\u002FCanva, ICLR 2025) achieved 95% of GPT-4 quality using only ~26% of GPT-4 calls, with up to 85% cost reduction on MT-Bench.\n\n3. **Move to open weights self-hosted on vLLM\u002FSGLang** when privacy, air-gap, predictable cost at scale, or fine-tuning control becomes the binding constraint.\n\n---\n\n## Build Pattern for LLM Applications\n\n**Prompting → RAG → Fine-tuning (in that order)**\n\n- **Few-shot prompting** for behavioral change\n- **RAG** for dynamic\u002Ffactual\u002Fauditable knowledge — use hybrid retrieval (BM25 + dense) + reranking + contextual retrieval\n- **Fine-tune** (LoRA\u002FQLoRA) only for consistent format\u002Fstyle\u002Fjargon or cost reduction\n- **GRPO\u002FDPO** only when you have verifiable rewards or preference data and a reasoning\u002Fbehavior target\n\n---\n\n## Reasoning Models & Test-Time Compute\n\nThe o1\u002Fo3\u002FR1 paradigm: spending more inference compute on chain-of-thought \"thinking tokens\" is a complementary scaling axis to parameters\u002Fdata.\n\n**Process Reward Models (PRMs)**: score intermediate reasoning steps; outperform outcome-only rewards on math (GenPRM: 1.5B model beats GPT-4o on ProcessBench via test-time scaling).\n\n**GRPO** (Group Relative Policy Optimization): critic-free, group-relative RL with verifiable rewards (RLVR). Dominant post-training method for reasoning. Spawned variants DAPO, Dr. GRPO, GSPO, GMPO.\n\n**When reasoning tokens are worth it**: genuine multi-step reasoning (math, code generation, legal analysis). Wasteful when wired into pipelines expecting short outputs.\n\n---\n\n## Classical ML — Tabular Data\n\nGradient-boosted trees remain state-of-the-art on most tabular problems. Deep learning requires more tuning and rarely wins on tabular without genuine unstructured signal.\n\n**2025 challenger**: TabPFN (transformer tabular foundation model, published in Nature) beats GBDTs on small\u002Fmedium single tables. XGBoost still wins on large tables.\n\n**Practitioner default**: GBDT-first (LightGBM → CatBoost → XGBoost depending on data characteristics). Only reach for deep learning with genuine unstructured signal or very large data.\n\n---\n\n## Mathematical Foundations\n\n**Linear algebra**: tensors are batched multidimensional arrays; matmul\u002Fdot products implement every linear layer and attention score; SVD underpins PCA and low-rank adaptation (LoRA literally learns low-rank ΔW). Norms drive regularization (L1 sparsity\u002FLasso, L2 weight decay\u002FRidge) and gradient clipping.\n\n**Probability**: KL\u002FJS\u002FWasserstein divergences anchor VAEs, GANs, and distribution matching. Cross-entropy is the default classification loss (= MLE under a categorical model). Bias–variance is operationalized by regularization and ensembling.\n\n**Optimization**: backprop = reverse-mode autodiff applying chain rule over the computational graph. Deep loss landscapes are non-convex but navigable — saddle points (not bad local minima) dominate, and SGD noise helps escape them. **AdamW** is the safe default; cosine annealing with linear warmup is the standard LR schedule. First-order methods dominate because Hessian-based methods don't scale to billions of parameters.\n\n---\n\n## Classical ML Reference\n\n### Regression & Classification\n- OLS with Ridge\u002FLasso\u002FElasticNet for regularized selection\n- GLMs (logistic, Poisson, negative binomial) for non-Gaussian targets\n- Gaussian Process Regression for small-data + calibrated uncertainty\n- Logistic regression: strong, interpretable baseline — always try it first\n- SVMs (kernel trick, RBF): still win on small high-dimensional data\n\n### Calibration\n**Calibration matters in production.** Temperature scaling (post-hoc, single-parameter) for neural nets; isotonic\u002FPlatt for others. Brier score and reliability diagrams for diagnosis.\n\n### Unsupervised & Anomaly Detection\n- Clustering: k-means(++)\u002FDBSCAN\u002FHDBSCAN\u002FGMM\u002Fspectral\n- Dimensionality reduction: PCA\u002FUMAP\u002Ft-SNE (UMAP preferred for preserving global structure and speed)\n- Anomaly detection: isolation forest\u002FLOF\u002Fautoencoders\n\n### Model Selection\n- Stratified\u002Fgroup\u002Ftime-series CV to prevent leakage\n- Metric choice: AUC-PR for imbalanced, MCC as balanced single number, log-loss for probabilistic\n- Optuna (TPE\u002FBayesian) + Hyperband\u002FASHA for hyperparameter tuning\n- **Leakage is the silent killer**: target leakage, train\u002Ftest contamination via preprocessing, and temporal leakage cause the most \"great offline, broken online\" failures\n\n---\n\n## Deep Learning Foundations\n\n### Activations & Normalization\n- **ReLU**: default for CNNs\n- **GELU\u002FSiLU(Swish)**: dominate transformers (smooth, work well with normalization)\n- **BatchNorm**: CNNs\n- **LayerNorm\u002FRMSNorm**: transformers (RMSNorm is cheaper, now standard in LLMs)\n- **Pre-norm** (LayerNorm before sublayer): dominates modern transformers for stability at depth\n\n### Initialization & Precision\n- He\u002FKaiming for ReLU nets; Xavier for tanh; orthogonal for RNNs\n- **BF16**: LLM-training default (wider dynamic range than FP16, no loss scaling)\n- **FP8**: production on Hopper\u002FBlackwell (DeepSeek-V3 trained in FP8)\n- Mixed precision memory tools: gradient checkpointing, gradient accumulation, activation recomputation\n\n### Residual Connections\nEnable very deep networks by giving gradients an identity path.\n\n---\n\n## The Transformer & Attention\n\n**Architecture types:**\n- Encoder-only (BERT): classification\u002Fembedding\n- Encoder-decoder (T5): seq2seq tasks\n- Decoder-only (GPT): dominates generative LLMs\n\n**Attention components:**\n- Scaled dot-product attention over Q,K,V\n- Multi-head attention\n- **RoPE** (rotary positional encoding): now standard; YaRN\u002Fposition interpolation extends context\n- **FlashAttention (v1→v3)**: IO-aware, exact (not approximate), tiling attention in SRAM — mitigates O(n²) attention cost\n- **KV cache**: makes autoregressive decoding tractable\n- FFN uses **SwiGLU**\n\n**Mixture of Experts (MoE)**: dominant frontier-scaling lever. Sparse gating + routing + load balancing. DeepSeek-V3 activates 37B of 671B params per token. Mixtral, DeepSeek-MoE, Qwen3-MoE all use this architecture.\n\n---\n\n## LLM Training & Post-Training\n\n### Tokenization\n- BPE\u002FWordPiece\u002FSentencePiece\u002Ftiktoken; larger vocabulary = shorter sequences but larger embedding tables\n- **Chinchilla scaling** (~20:1 tokens:params) guided compute-optimal training; inference-cost economics now push toward \"overtrain a smaller model\"\n\n### Pre-training Infrastructure\n- Curated\u002Fdeduplicated\u002Fquality-filtered corpora + heavy synthetic data (Phi \"textbooks\" thesis)\n- Tensor\u002Fpipeline\u002Fdata parallelism; ZeRO\u002FDeepSpeed; FSDP; Megatron-LM\n\n### Alignment Pipeline\n1. **SFT** (Supervised Fine-Tuning): high-quality instruction examples; \"LIMA \u002F Less is More\" — a few thousand high-quality examples often beat massive noisy sets\n2. **DPO** (Direct Preference Optimization): reference-model based, no explicit reward model; simple default\n3. **GRPO**: critic-free, group-relative, RLVR; dominates reasoning post-training\n\n### PEFT (Parameter-Efficient Fine-Tuning)\n- **LoRA**: low-rank adapters; workhorses of production fine-tuning\n- **QLoRA**: 4-bit NF4 base + LoRA; enables fine-tuning large models on consumer hardware\n- DoRA\u002FVeRA: refinements of LoRA\n\n---\n\n## LLM Inference Optimization\n\n### Quantization Formats\n| Format | Use case | Notes |\n|---|---|---|\n| **AWQ** | Best throughput in vLLM; INT4 weight-only | \u003C2% degradation on most tasks; noticeable on math\u002Fcode\u002Freasoning |\n| **GPTQ** | Best throughput in vLLM | Similar trade-offs to AWQ |\n| **GGUF** | llama.cpp\u002FOllama; CPU+GPU offload | Q4_K_M–Q6_K near-BF16 quality |\n| **FP8 W8A8** | Near-lossless on Hopper\u002FBlackwell | Production preferred when hardware supports it |\n\n**Rule**: avoid INT4 for math\u002Fcode\u002Freasoning-critical paths.\n\n### Serving Engines\n| Engine | Strength |\n|---|---|\n| **vLLM** | PagedAttention; broadest model\u002Fhardware support; V1 architecture |\n| **SGLang** | RadixAttention; ~29% higher H100 throughput; up to 6.4× on prefix-heavy RAG\u002Fchat |\n| **TensorRT-LLM** | Maximum NVIDIA throughput |\n| **Ollama\u002Fllama.cpp** | Local deployment; CPU offload |\n\n**Continuous batching** + **disaggregated prefill\u002Fdecode** (separate compute-bound prefill and bandwidth-bound decode pools) are the 2026 production patterns.\n\n**Speculative decoding** (draft+verify, Medusa\u002FEAGLE): 2–4× latency wins losslessly.\n\n---\n\n## Evaluation & Benchmark Skepticism\n\n**The contamination problem is severe.** A single model can swing 17–35 points between SWE-bench Verified and the standardized-scaffold Pro leaderboard — scaffold and harness move scores more than model swaps.\n\n**Key benchmarks:**\n- MMLU(-Pro), GPQA Diamond: knowledge\u002Freasoning\n- HumanEval\u002FMBPP, SWE-bench Verified\u002FPro: coding\n- GSM8K\u002FMATH: math reasoning\n- MMMU: multimodal\n- LiveBench\u002FLiveCodeBench: contamination-resistant\n- Chatbot Arena Elo: human preference\n- RULER\u002FHELMET: long-context\n- TruthfulQA\u002FFActScore: factuality\n- RAGAS: RAG evaluation\n\n**Decision rule for benchmarks:**\n- Never select on a single leaderboard number\n- Build a 50–200 case internal eval of your real workflows\n- Weight SWE-bench Pro (standardized scaffold) over Verified for coding\n- When a benchmark and your production results diverge sharply, trust contamination\u002Fscaffold effects over the leaderboard\n- LLM-as-judge is widespread but carries position\u002Fverbosity\u002Fself-preference biases\n\n---\n\n## RAG Fundamentals (Quick Reference)\n\n**Production default** (2026): hybrid search (BM25 + dense via RRF) + cross-encoder reranking + parent-child chunking.\n\n**Anthropic Contextual Retrieval** (Sept 2024): prepend LLM-generated chunk summary before embedding\u002Findexing.\n- Contextual Embeddings alone: 35% reduction in retrieval failure (5.7%→3.7%)\n- + Contextual BM25: 49% reduction (→2.9%)\n- + Reranking: 67% reduction (→1.9%)\n\n**RAG vs fine-tuning vs long-context:**\n- RAG: dynamic\u002Fproprietary knowledge with citations\n- Fine-tuning: changing behavior\u002Fformat\u002Ftone\u002Fdomain style\n- Long-context: single-document deep reasoning where doc fits\n- They combine\n\n---\n\n## Agent Stack & MCP\n\n**Model Context Protocol (MCP)**: Anthropic, Nov 2024; donated to Linux Foundation Dec 2025; adopted by OpenAI, Google, Microsoft, Amazon. \"USB-C for AI tools.\" Over 97M monthly SDK downloads and 10,000 active servers at donation time.\n\n**Production failure modes**: runaway loops, unbounded retries, silent context truncation. Iteration limits, cost ceilings, and termination conditions are mandatory, not polish.\n\n**Single-agent suffices for ~80% of cases.** Multi-agent adds cost and non-determinism.\n\n**Framework selection:**\n- **LangGraph**: production default for regulated\u002Fauditable workflows; graph state machines with checkpointing\n- **CrewAI**: fastest role-based multi-agent prototyping\n- **Pydantic AI**: typed, minimal boilerplate\n\n---\n\n## Generative AI Across Modalities\n\n### Image\n- GANs and VAEs gave way to **diffusion** (DDPM→DDIM→Latent Diffusion = Stable Diffusion)\n- Now **flow matching \u002F rectified flow** (backbone of SD3, FLUX)\n- Classifier-free guidance, ControlNet (pose\u002Fdepth\u002Fedge conditioning), LoRA\u002FDreamBooth personalization are standard\n- FLUX (Black Forest Labs): open-weight image generation leader in 2026\n\n### Video\n- Standardized on **diffusion transformers (DiT) over spacetime patches** (Sora, Veo, Kling, Seedance, WAN, Hunyuan)\n- By early 2026: 8–25s clips at native resolution with synchronized audio and plausible physics\n- Weaknesses: long-range temporal coherence and quadratic attention cost\n\n### Audio\n- TTS: Tacotron→FastSpeech→VITS→XTTS\u002FF5-TTS\u002FKokoro; ElevenLabs leads commercial quality\n- ASR: Whisper dominates\n- Discrete audio tokens: EnCodec\u002FDAC\n\n### Multimodal\u002FVLMs\n- Vision encoders trend SigLIP over CLIP, with tile-based high-res\n- Open options: LLaVA, Qwen-VL, InternVL, Molmo, Pixtral\n\n---\n\n## MLOps & Infrastructure\n\n### Hardware\n| Chip | Key spec | Notes |\n|---|---|---|\n| NVIDIA H100\u002FH200 (Hopper) | 80GB\u002F141GB HBM3 | Current production standard |\n| B200\u002FGB200 (Blackwell) | 192GB HBM3e at 8TB\u002Fs | ~2.5–3× H100 training; GB200 NVL72 trained Llama 3.1 405B in ~10 min on 5,120 GPUs (MLPerf) |\n| AMD MI300X\u002FMI325X\u002FMI355X | 192–288GB HBM3 | More memory; CUDA\u002FROCm moat keeps NVIDIA at ~80%+ share |\n| Google TPU v5\u002Fv6e | — | JAX\u002FXLA affinity |\n| Groq LPU | — | Deterministic high-throughput inference |\n\n### Frameworks\n- **PyTorch**: dominant (eager + torch.compile\u002FInductor)\n- **JAX**: TPU\u002Fresearch\n- **Hugging Face stack**: Transformers, Diffusers, PEFT, TRL, Datasets, Accelerate — the default\n- **Experiment tracking**: W&B (research favorite), MLflow (open, ubiquitous)\n\n### Observability & Monitoring\n- **LLM serving\u002Fobservability**: LangSmith, Langfuse, Arize Phoenix, Helicone, Braintrust, W&B Weave\n- **Drift\u002Fmonitoring**: Evidently, NannyML, Arize, WhyLabs, Fiddler; PSI\u002FKS tests; distinguish covariate\u002Flabel\u002Fconcept shift\n- **Explainability**: SHAP\u002FLIME\u002FIntegrated Gradients (caution: attention weights ≠ explanations)\n\n---\n\n## Advanced Architectures\n\n### State-Space Models (SSMs)\n- **Mamba** (selective state spaces, Mamba-2): linear complexity, ~5× throughput\n- **Critical weakness**: pure SSMs fail at retrieval\u002Fin-context copying — removing attention layers drops retrieval accuracy to near-zero\n- **Hybrids win in production**: Jamba (AI21), Zamba2, NVIDIA Nemotron-H, IBM Granite 4.0 — small fraction of attention layers interleaved with Mamba blocks\n\n### Long Context\n- 1M+ token windows are table stakes across flagships\n- \"Lost in the middle\" and needle-in-haystack failures persist\n- RAG vs long-context is a cost\u002Ffreshness\u002Fauditability decision, not a pure capability one\n\n### Reasoning Research\n- Test-time compute scaling via PRMs\n- MCTS for reasoning\n- Self-consistency voting\n- Formal-verification systems (AlphaProof\u002FAlphaGeometry 2)\n\n---\n\n## Safety, Alignment & Governance\n\n### Alignment Approaches\n- **Constitutional AI \u002F RLAIF** (Anthropic): AI self-critique for alignment\n- **Mechanistic interpretability**: sparse autoencoders decompose superposed\u002Fpolysemantic activations; circuit tracing with cross-layer transcoders lets researchers trace and intervene on causal pathways (Anthropic, open-sourced May 2025; MIT Tech Review 2026 Breakthrough Technology)\n- Outer alignment (specification) vs inner alignment (goal generalization)\n- Reward hacking and Goodhart's Law: recurring failure modes\n\n### EU AI Act Timeline (Critical Deadlines)\n| Date | What takes effect |\n|---|---|\n| Aug 1, 2024 | In force |\n| Feb 2, 2025 | Prohibited practices + AI literacy obligations |\n| Aug 2, 2025 | GPAI (General Purpose AI) obligations |\n| **Aug 2, 2026** | **GPAI enforcement powers and fines; high-risk system obligations** |\n| Aug 2, 2027 | Legacy-GPAI compliance deadline |\n\nFines: up to €35M or 7% of global turnover (prohibited practices) — higher ceiling than GDPR. The voluntary GPAI Code of Practice offers a \"presumption of conformity\" safe harbor.\n\n**Action**: If you touch EU users, classify your system now against the AI Act risk tiers. If you fine-tune a GPAI model substantially, you may become a \"provider\" with heavier obligations.\n\n### Fairness\nDemographic parity vs equalized odds vs individual fairness are mutually incompatible (impossibility theorems). Tools: Fairlearn, AI Fairness 360.\n\n### Privacy\nDifferential privacy, federated learning, and defenses against membership-inference\u002Fmodel-inversion\u002Fpoisoning attacks.\n\n### US Framework\nNIST AI RMF (Govern\u002FMap\u002FMeasure\u002FManage) is the US reference standard.\n\n---\n\n## Durable Principles (Outlast Model Names)\n\n1. **Baseline-first**: always establish a simple baseline before reaching for complex architectures\n2. **Data quality over architecture**: garbage in, garbage out — no model overcomes bad data\n3. **Route, don't over-provision**: tiered model fleets beat one expensive frontier model\n4. **RAG before fine-tune**: try prompting and retrieval before committing to training\n5. **Guardrails and observability before launch**: build I\u002FO filtering, tracing, and cost monitoring into the initial design\n6. **Benchmark skepticism**: build internal evals on your real workflows; never select on a single leaderboard number\n7. **Open vs closed is a cost\u002Fprivacy\u002Fcapability trade-off, not an ideology**: evaluate quarterly as quality gaps narrow\n","# AI & Machine Learning Landscape — Practitioner Reference (2026)\n\n## The Three-Layer Structure of AI in 2026\n\n1. **Frontier model race** — fast-moving, narrow quality gaps between top models (~2.7% lead at the top as of March 2026)\n2. **Production engineering stack** — maturing: vLLM\u002FSGLang serving, MCP-standardized agents, contextual-retrieval RAG\n3. **Classical ML core** — stable: gradient boosting still wins most tabular problems\n\n**Inference cost collapse**: >280-fold reduction from $20.00 to $0.07 per million tokens between Nov 2022 and Oct 2024 (GPT-3.5-equivalent quality, per Stanford HAI). This makes capability cheap and **engineering discipline** the binding constraint.\n\n---\n\n## Frontier Model Landscape (Mid-2026)\n\nNo single model dominates. The top model leads by only ~2.7% (Stanford HAI AI Index, March 2026). The leading closed-weight model's edge over the top open-weight model on Chatbot Arena narrowed from 8.04% (Jan 2024) to 1.70% (Feb 2025).\n\n**Key practitioner principle:** Build behind an abstraction layer (LiteLLM\u002FOpenRouter) to avoid lock-in. Model names and rankings shift weekly.\n\n---\n\n## Hosted APIs vs Open Weights — Decision Rule\n\n| Factor | Hosted API | Open Weights |\n|---|---|---|\n| Peak capability, zero ops | ✓ | |\n| Privacy, air-gap, data residency | | ✓ |\n| Fine-tuning control | | ✓ |\n| Predictable cost at sustained high QPS | | ✓ |\n| Fastest to start | ✓ | |\n\n**Open-weight options**: DeepSeek-V4 (MIT-licensed, 1M context, 1.6T\u002F49B active MoE), Qwen3, Meta Llama 4, Mistral, IBM Granite 4.0.\n\n**Threshold to self-host**: sustained high QPS where API spend exceeds fully-loaded GPU fleet cost, or a hard data-residency requirement.\n\n---\n\n## Model Selection Strategy (Staged)\n\n1. **Start with hosted frontier behind an abstraction layer.** Default to strong mid-tier (Claude Sonnet-class, GPT-4.1-mini-class, Gemini Flash-class); escalate to top-tier only for hard tasks.\n\n2. **Implement model routing** once volume justifies it. A cheap classifier sends easy queries to small\u002Fcheap models with confidence-gated escalation. RouteLLM (UC Berkeley\u002FAnyscale\u002FCanva, ICLR 2025) achieved 95% of GPT-4 quality using only ~26% of GPT-4 calls, with up to 85% cost reduction on MT-Bench.\n\n3. **Move to open weights self-hosted on vLLM\u002FSGLang** when privacy, air-gap, predictable cost at scale, or fine-tuning control becomes the binding constraint.\n\n---\n\n## Build Pattern for LLM Applications\n\n**Prompting → RAG → Fine-tuning (in that order)**\n\n- **Few-shot prompting** for behavioral change\n- **RAG** for dynamic\u002Ffactual\u002Fauditable knowledge — use hybrid retrieval (BM25 + dense) + reranking + contextual retrieval\n- **Fine-tune** (LoRA\u002FQLoRA) only for consistent format\u002Fstyle\u002Fjargon or cost reduction\n- **GRPO\u002FDPO** only when you have verifiable rewards or preference data and a reasoning\u002Fbehavior target\n\n---\n\n## Reasoning Models & Test-Time Compute\n\nThe o1\u002Fo3\u002FR1 paradigm: spending more inference compute on chain-of-thought \"thinking tokens\" is a complementary scaling axis to parameters\u002Fdata.\n\n**Process Reward Models (PRMs)**: score intermediate reasoning steps; outperform outcome-only rewards on math (GenPRM: 1.5B model beats GPT-4o on ProcessBench via test-time scaling).\n\n**GRPO** (Group Relative Policy Optimization): critic-free, group-relative RL with verifiable rewards (RLVR). Dominant post-training method for reasoning. Spawned variants DAPO, Dr. GRPO, GSPO, GMPO.\n\n**When reasoning tokens are worth it**: genuine multi-step reasoning (math, code generation, legal analysis). Wasteful when wired into pipelines expecting short outputs.\n\n---\n\n## Classical ML — Tabular Data\n\nGradient-boosted trees remain state-of-the-art on most tabular problems. Deep learning requires more tuning and rarely wins on tabular without genuine unstructured signal.\n\n**2025 challenger**: TabPFN (transformer tabular foundation model, published in Nature) beats GBDTs on small\u002Fmedium single tables. XGBoost still wins on large tables.\n\n**Practitioner default**: GBDT-first (LightGBM → CatBoost → XGBoost depending on data characteristics). Only reach for deep learning with genuine unstructured signal or very large data.\n\n---\n\n## Mathematical Foundations\n\n**Linear algebra**: tensors are batched multidimensional arrays; matmul\u002Fdot products implement every linear layer and attention score; SVD underpins PCA and low-rank adaptation (LoRA literally learns low-rank ΔW). Norms drive regularization (L1 sparsity\u002FLasso, L2 weight decay\u002FRidge) and gradient clipping.\n\n**Probability**: KL\u002FJS\u002FWasserstein divergences anchor VAEs, GANs, and distribution matching. Cross-entropy is the default classification loss (= MLE under a categorical model). Bias–variance is operationalized by regularization and ensembling.\n\n**Optimization**: backprop = reverse-mode autodiff applying chain rule over the computational graph. Deep loss landscapes are non-convex but navigable — saddle points (not bad local minima) dominate, and SGD noise helps escape them. **AdamW** is the safe default; cosine annealing with linear warmup is the standard LR schedule. First-order methods dominate because Hessian-based methods don't scale to billions of parameters.\n\n---\n\n## Classical ML Reference\n\n### Regression & Classification\n- OLS with Ridge\u002FLasso\u002FElasticNet for regularized selection\n- GLMs (logistic, Poisson, negative binomial) for non-Gaussian targets\n- Gaussian Process Regression for small-data + calibrated uncertainty\n- Logistic regression: strong, interpretable baseline — always try it first\n- SVMs (kernel trick, RBF): still win on small high-dimensional data\n\n### Calibration\n**Calibration matters in production.** Temperature scaling (post-hoc, single-parameter) for neural nets; isotonic\u002FPlatt for others. Brier score and reliability diagrams for diagnosis.\n\n### Unsupervised & Anomaly Detection\n- Clustering: k-means(++)\u002FDBSCAN\u002FHDBSCAN\u002FGMM\u002Fspectral\n- Dimensionality reduction: PCA\u002FUMAP\u002Ft-SNE (UMAP preferred for preserving global structure and speed)\n- Anomaly detection: isolation forest\u002FLOF\u002Fautoencoders\n\n### Model Selection\n- Stratified\u002Fgroup\u002Ftime-series CV to prevent leakage\n- Metric choice: AUC-PR for imbalanced, MCC as balanced single number, log-loss for probabilistic\n- Optuna (TPE\u002FBayesian) + Hyperband\u002FASHA for hyperparameter tuning\n- **Leakage is the silent killer**: target leakage, train\u002Ftest contamination via preprocessing, and temporal leakage cause the most \"great offline, broken online\" failures\n\n---\n\n## Deep Learning Foundations\n\n### Activations & Normalization\n- **ReLU**: default for CNNs\n- **GELU\u002FSiLU(Swish)**: dominate transformers (smooth, work well with normalization)\n- **BatchNorm**: CNNs\n- **LayerNorm\u002FRMSNorm**: transformers (RMSNorm is cheaper, now standard in LLMs)\n- **Pre-norm** (LayerNorm before sublayer): dominates modern transformers for stability at depth\n\n### Initialization & Precision\n- He\u002FKaiming for ReLU nets; Xavier for tanh; orthogonal for RNNs\n- **BF16**: LLM-training default (wider dynamic range than FP16, no loss scaling)\n- **FP8**: production on Hopper\u002FBlackwell (DeepSeek-V3 trained in FP8)\n- Mixed precision memory tools: gradient checkpointing, gradient accumulation, activation recomputation\n\n### Residual Connections\nEnable very deep networks by giving gradients an identity path.\n\n---\n\n## The Transformer & Attention\n\n**Architecture types:**\n- Encoder-only (BERT): classification\u002Fembedding\n- Encoder-decoder (T5): seq2seq tasks\n- Decoder-only (GPT): dominates generative LLMs\n\n**Attention components:**\n- Scaled dot-product attention over Q,K,V\n- Multi-head attention\n- **RoPE** (rotary positional encoding): now standard; YaRN\u002Fposition interpolation extends context\n- **FlashAttention (v1→v3)**: IO-aware, exact (not approximate), tiling attention in SRAM — mitigates O(n²) attention cost\n- **KV cache**: makes autoregressive decoding tractable\n- FFN uses **SwiGLU**\n\n**Mixture of Experts (MoE)**: dominant frontier-scaling lever. Sparse gating + routing + load balancing. DeepSeek-V3 activates 37B of 671B params per token. Mixtral, DeepSeek-MoE, Qwen3-MoE all use this architecture.\n\n---\n\n## LLM Training & Post-Training\n\n### Tokenization\n- BPE\u002FWordPiece\u002FSentencePiece\u002Ftiktoken; larger vocabulary = shorter sequences but larger embedding tables\n- **Chinchilla scaling** (~20:1 tokens:params) guided compute-optimal training; inference-cost economics now push toward \"overtrain a smaller model\"\n\n### Pre-training Infrastructure\n- Curated\u002Fdeduplicated\u002Fquality-filtered corpora + heavy synthetic data (Phi \"textbooks\" thesis)\n- Tensor\u002Fpipeline\u002Fdata parallelism; ZeRO\u002FDeepSpeed; FSDP; Megatron-LM\n\n### Alignment Pipeline\n1. **SFT** (Supervised Fine-Tuning): high-quality instruction examples; \"LIMA \u002F Less is More\" — a few thousand high-quality examples often beat massive noisy sets\n2. **DPO** (Direct Preference Optimization): reference-model based, no explicit reward model; simple default\n3. **GRPO**: critic-free, group-relative, RLVR; dominates reasoning post-training\n\n### PEFT (Parameter-Efficient Fine-Tuning)\n- **LoRA**: low-rank adapters; workhorses of production fine-tuning\n- **QLoRA**: 4-bit NF4 base + LoRA; enables fine-tuning large models on consumer hardware\n- DoRA\u002FVeRA: refinements of LoRA\n\n---\n\n## LLM Inference Optimization\n\n### Quantization Formats\n| Format | Use case | Notes |\n|---|---|---|\n| **AWQ** | Best throughput in vLLM; INT4 weight-only | \u003C2% degradation on most tasks; noticeable on math\u002Fcode\u002Freasoning |\n| **GPTQ** | Best throughput in vLLM | Similar trade-offs to AWQ |\n| **GGUF** | llama.cpp\u002FOllama; CPU+GPU offload | Q4_K_M–Q6_K near-BF16 quality |\n| **FP8 W8A8** | Near-lossless on Hopper\u002FBlackwell | Production preferred when hardware supports it |\n\n**Rule**: avoid INT4 for math\u002Fcode\u002Freasoning-critical paths.\n\n### Serving Engines\n| Engine | Strength |\n|---|---|\n| **vLLM** | PagedAttention; broadest model\u002Fhardware support; V1 architecture |\n| **SGLang** | RadixAttention; ~29% higher H100 throughput; up to 6.4× on prefix-heavy RAG\u002Fchat |\n| **TensorRT-LLM** | Maximum NVIDIA throughput |\n| **Ollama\u002Fllama.cpp** | Local deployment; CPU offload |\n\n**Continuous batching** + **disaggregated prefill\u002Fdecode** (separate compute-bound prefill and bandwidth-bound decode pools) are the 2026 production patterns.\n\n**Speculative decoding** (draft+verify, Medusa\u002FEAGLE): 2–4× latency wins losslessly.\n\n---\n\n## Evaluation & Benchmark Skepticism\n\n**The contamination problem is severe.** A single model can swing 17–35 points between SWE-bench Verified and the standardized-scaffold Pro leaderboard — scaffold and harness move scores more than model swaps.\n\n**Key benchmarks:**\n- MMLU(-Pro), GPQA Diamond: knowledge\u002Freasoning\n- HumanEval\u002FMBPP, SWE-bench Verified\u002FPro: coding\n- GSM8K\u002FMATH: math reasoning\n- MMMU: multimodal\n- LiveBench\u002FLiveCodeBench: contamination-resistant\n- Chatbot Arena Elo: human preference\n- RULER\u002FHELMET: long-context\n- TruthfulQA\u002FFActScore: factuality\n- RAGAS: RAG evaluation\n\n**Decision rule for benchmarks:**\n- Never select on a single leaderboard number\n- Build a 50–200 case internal eval of your real workflows\n- Weight SWE-bench Pro (standardized scaffold) over Verified for coding\n- When a benchmark and your production results diverge sharply, trust contamination\u002Fscaffold effects over the leaderboard\n- LLM-as-judge is widespread but carries position\u002Fverbosity\u002Fself-preference biases\n\n---\n\n## RAG Fundamentals (Quick Reference)\n\n**Production default** (2026): hybrid search (BM25 + dense via RRF) + cross-encoder reranking + parent-child chunking.\n\n**Anthropic Contextual Retrieval** (Sept 2024): prepend LLM-generated chunk summary before embedding\u002Findexing.\n- Contextual Embeddings alone: 35% reduction in retrieval failure (5.7%→3.7%)\n- + Contextual BM25: 49% reduction (→2.9%)\n- + Reranking: 67% reduction (→1.9%)\n\n**RAG vs fine-tuning vs long-context:**\n- RAG: dynamic\u002Fproprietary knowledge with citations\n- Fine-tuning: changing behavior\u002Fformat\u002Ftone\u002Fdomain style\n- Long-context: single-document deep reasoning where doc fits\n- They combine\n\n---\n\n## Agent Stack & MCP\n\n**Model Context Protocol (MCP)**: Anthropic, Nov 2024; donated to Linux Foundation Dec 2025; adopted by OpenAI, Google, Microsoft, Amazon. \"USB-C for AI tools.\" Over 97M monthly SDK downloads and 10,000 active servers at donation time.\n\n**Production failure modes**: runaway loops, unbounded retries, silent context truncation. Iteration limits, cost ceilings, and termination conditions are mandatory, not polish.\n\n**Single-agent suffices for ~80% of cases.** Multi-agent adds cost and non-determinism.\n\n**Framework selection:**\n- **LangGraph**: production default for regulated\u002Fauditable workflows; graph state machines with checkpointing\n- **CrewAI**: fastest role-based multi-agent prototyping\n- **Pydantic AI**: typed, minimal boilerplate\n\n---\n\n## Generative AI Across Modalities\n\n### Image\n- GANs and VAEs gave way to **diffusion** (DDPM→DDIM→Latent Diffusion = Stable Diffusion)\n- Now **flow matching \u002F rectified flow** (backbone of SD3, FLUX)\n- Classifier-free guidance, ControlNet (pose\u002Fdepth\u002Fedge conditioning), LoRA\u002FDreamBooth personalization are standard\n- FLUX (Black Forest Labs): open-weight image generation leader in 2026\n\n### Video\n- Standardized on **diffusion transformers (DiT) over spacetime patches** (Sora, Veo, Kling, Seedance, WAN, Hunyuan)\n- By early 2026: 8–25s clips at native resolution with synchronized audio and plausible physics\n- Weaknesses: long-range temporal coherence and quadratic attention cost\n\n### Audio\n- TTS: Tacotron→FastSpeech→VITS→XTTS\u002FF5-TTS\u002FKokoro; ElevenLabs leads commercial quality\n- ASR: Whisper dominates\n- Discrete audio tokens: EnCodec\u002FDAC\n\n### Multimodal\u002FVLMs\n- Vision encoders trend SigLIP over CLIP, with tile-based high-res\n- Open options: LLaVA, Qwen-VL, InternVL, Molmo, Pixtral\n\n---\n\n## MLOps & Infrastructure\n\n### Hardware\n| Chip | Key spec | Notes |\n|---|---|---|\n| NVIDIA H100\u002FH200 (Hopper) | 80GB\u002F141GB HBM3 | Current production standard |\n| B200\u002FGB200 (Blackwell) | 192GB HBM3e at 8TB\u002Fs | ~2.5–3× H100 training; GB200 NVL72 trained Llama 3.1 405B in ~10 min on 5,120 GPUs (MLPerf) |\n| AMD MI300X\u002FMI325X\u002FMI355X | 192–288GB HBM3 | More memory; CUDA\u002FROCm moat keeps NVIDIA at ~80%+ share |\n| Google TPU v5\u002Fv6e | — | JAX\u002FXLA affinity |\n| Groq LPU | — | Deterministic high-throughput inference |\n\n### Frameworks\n- **PyTorch**: dominant (eager + torch.compile\u002FInductor)\n- **JAX**: TPU\u002Fresearch\n- **Hugging Face stack**: Transformers, Diffusers, PEFT, TRL, Datasets, Accelerate — the default\n- **Experiment tracking**: W&B (research favorite), MLflow (open, ubiquitous)\n\n### Observability & Monitoring\n- **LLM serving\u002Fobservability**: LangSmith, Langfuse, Arize Phoenix, Helicone, Braintrust, W&B Weave\n- **Drift\u002Fmonitoring**: Evidently, NannyML, Arize, WhyLabs, Fiddler; PSI\u002FKS tests; distinguish covariate\u002Flabel\u002Fconcept shift\n- **Explainability**: SHAP\u002FLIME\u002FIntegrated Gradients (caution: attention weights ≠ explanations)\n\n---\n\n## Advanced Architectures\n\n### State-Space Models (SSMs)\n- **Mamba** (selective state spaces, Mamba-2): linear complexity, ~5× throughput\n- **Critical weakness**: pure SSMs fail at retrieval\u002Fin-context copying — removing attention layers drops retrieval accuracy to near-zero\n- **Hybrids win in production**: Jamba (AI21), Zamba2, NVIDIA Nemotron-H, IBM Granite 4.0 — small fraction of attention layers interleaved with Mamba blocks\n\n### Long Context\n- 1M+ token windows are table stakes across flagships\n- \"Lost in the middle\" and needle-in-haystack failures persist\n- RAG vs long-context is a cost\u002Ffreshness\u002Fauditability decision, not a pure capability one\n\n### Reasoning Research\n- Test-time compute scaling via PRMs\n- MCTS for reasoning\n- Self-consistency voting\n- Formal-verification systems (AlphaProof\u002FAlphaGeometry 2)\n\n---\n\n## Safety, Alignment & Governance\n\n### Alignment Approaches\n- **Constitutional AI \u002F RLAIF** (Anthropic): AI self-critique for alignment\n- **Mechanistic interpretability**: sparse autoencoders decompose superposed\u002Fpolysemantic activations; circuit tracing with cross-layer transcoders lets researchers trace and intervene on causal pathways (Anthropic, open-sourced May 2025; MIT Tech Review 2026 Breakthrough Technology)\n- Outer alignment (specification) vs inner alignment (goal generalization)\n- Reward hacking and Goodhart's Law: recurring failure modes\n\n### EU AI Act Timeline (Critical Deadlines)\n| Date | What takes effect |\n|---|---|\n| Aug 1, 2024 | In force |\n| Feb 2, 2025 | Prohibited practices + AI literacy obligations |\n| Aug 2, 2025 | GPAI (General Purpose AI) obligations |\n| **Aug 2, 2026** | **GPAI enforcement powers and fines; high-risk system obligations** |\n| Aug 2, 2027 | Legacy-GPAI compliance deadline |\n\nFines: up to €35M or 7% of global turnover (prohibited practices) — higher ceiling than GDPR. The voluntary GPAI Code of Practice offers a \"presumption of conformity\" safe harbor.\n\n**Action**: If you touch EU users, classify your system now against the AI Act risk tiers. If you fine-tune a GPAI model substantially, you may become a \"provider\" with heavier obligations.\n\n### Fairness\nDemographic parity vs equalized odds vs individual fairness are mutually incompatible (impossibility theorems). Tools: Fairlearn, AI Fairness 360.\n\n### Privacy\nDifferential privacy, federated learning, and defenses against membership-inference\u002Fmodel-inversion\u002Fpoisoning attacks.\n\n### US Framework\nNIST AI RMF (Govern\u002FMap\u002FMeasure\u002FManage) is the US reference standard.\n\n---\n\n## Durable Principles (Outlast Model Names)\n\n1. **Baseline-first**: always establish a simple baseline before reaching for complex architectures\n2. **Data quality over architecture**: garbage in, garbage out — no model overcomes bad data\n3. **Route, don't over-provision**: tiered model fleets beat one expensive frontier model\n4. **RAG before fine-tune**: try prompting and retrieval before committing to training\n5. **Guardrails and observability before launch**: build I\u002FO filtering, tracing, and cost monitoring into the initial design\n6. **Benchmark skepticism**: build internal evals on your real workflows; never select on a single leaderboard number\n7. **Open vs closed is a cost\u002Fprivacy\u002Fcapability trade-off, not an ideology**: evaluate quarterly as quality gaps narrow\n",{"name":5,"description":7},"develop","ae32736a1aabddabca5d2c96d665aa0ebc7d8700","src\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fai-and-data\u002Fskills\u002Fai-ml-landscape",3,"https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey","https:\u002F\u002Fthe-vibey-project.github.io\u002Fvibey\u002F","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project","https:\u002F\u002Fvibewithadam.matthewsteinberger.com","freedom for software engineers","2026-11-15 05:47:42",{"source":8,"skillName":5,"commands":35,"prompt":45,"zipAvailable":46},[36,39,42],{"tool":37,"command":38},"npx","npx skills add the-vibey-project\u002Fvibey --skill ai-ml-landscape",{"tool":40,"command":41},"bunx","bunx skills add the-vibey-project\u002Fvibey --skill ai-ml-landscape",{"tool":43,"command":44},"pnpm","pnpm dlx skills add the-vibey-project\u002Fvibey --skill ai-ml-landscape","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fai-and-data\u002Fskills\u002Fai-ml-landscape\u002FSKILL.md",true,[],[49,56,63,70,77,84],{"skillNo":50,"name":51,"nameZh":6,"description":52,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":53,"categories":54,"keywords":55,"publishedAt":19},"01M7AT1FQJAE1FQYAMDTF490Y4","white-paper-writing","Use when writing white papers, technical reports, policy documents, thought leadership pieces, or any long-form authoritative document. Triggers on requests to write, structure, outline, or review white papers, technical reports, position papers, policy briefs, or evidence-driven arguments. Also use for B2B content marketing strategy, technical documentation requiring authority, academic-style reports, or any document that must educate while persuading. Use when someone asks \"how should I structure this\" for a long-form document.","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fwriting-craft\u002Fskills\u002Fwhite-paper-writing\u002FSKILL.md",[],[],{"skillNo":57,"name":58,"nameZh":6,"description":59,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":60,"categories":61,"keywords":62,"publishedAt":19},"01M7AT16BJRJJKK2MASXRZS76G","prose-technical-mechanics","Use when editing or writing prose at a technical\u002Fmechanical level: choosing between Anglo-Saxon and Latinate diction, placing information in topic vs stress positions, writing periodic vs cumulative sentences, handling sentence rhythm, punctuation as logic, transitions, paragraph structure, or phonesthetics. Use when someone wants to understand WHY prose works or doesn't, not just surface-level style advice. Triggers on word choice, sentence structure, emphasis, rhythm, diction, or any deep writing mechanics question.","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fwriting-craft\u002Fskills\u002Fprose-technical-mechanics\u002FSKILL.md",[],[],{"skillNo":64,"name":65,"nameZh":6,"description":66,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":67,"categories":68,"keywords":69,"publishedAt":19},"01M7AT0W9HT0ZZ9Y287NW99MR3","prose-mechanics","Use when writing, editing, or analyzing prose style; when trying to avoid generic AI-sounding writing; when applying specific prose styles such as journalistic, academic, narrative, KJV biblical, Hemingway plain, Latinate\u002FJamesian, Orwellian, Augustan, Victorian, Early Modern (Bacon, Browne, Milton), or free indirect discourse; or when coaching someone on writing voice and mechanics. Triggers on any request to improve, analyze, or write in a specific prose style, to fix \"AI-sounding\" writing, to make prose more vivid or concrete, to vary sentence rhythm, or to understand how a particular writer achieves their effects.","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fwriting-craft\u002Fskills\u002Fprose-mechanics\u002FSKILL.md",[],[],{"skillNo":71,"name":72,"nameZh":6,"description":73,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":74,"categories":75,"keywords":76,"publishedAt":19},"01M7AT0KR635JX66X2X6F3E2CP","narrative-structure","Use when structuring stories, pitches, presentations, white papers, or any persuasive narrative; analyzing the hero's journey in fiction or non-fiction; applying StoryBrand for business communication; designing narrative arcs for thought leadership; or understanding Vogler, Snyder, Propp, or Campbell frameworks. Triggers on any narrative structure, storytelling, pitch structure, StoryBrand, hero's journey, three-act structure, Freytag, Save the Cat, narrative arc, or \"how should I structure this story\u002Fpitch\u002Fpresentation\" request.","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fwriting-craft\u002Fskills\u002Fnarrative-structure\u002FSKILL.md",[],[],{"skillNo":78,"name":79,"nameZh":6,"description":80,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":81,"categories":82,"keywords":83,"publishedAt":19},"01M7AT0AKGW09NKKZDYBT8VRDE","legal-brief-writing","Use when writing legal briefs, motions, or any court filing; when analyzing judicial cognition and persuasion; when applying elite brief-writing techniques from Roberts, Clement, Kagan, Garner; or when structuring legal arguments for maximum persuasion. Also use for autistic\u002Fneurodivergent advocates who need to understand how their natural communication style maps to elite legal advocacy. Triggers on any legal writing, brief drafting, motion preparation, or legal argument structuring request.","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fwriting-craft\u002Fskills\u002Flegal-brief-writing\u002FSKILL.md",[],[],{"skillNo":85,"name":86,"nameZh":6,"description":87,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":88,"categories":89,"keywords":90,"publishedAt":19},"01M7AT00VWPS4Q8B5VEMXX9NQT","journal-publishability-criteria","Use whenever a research paper, article or preprint is being written, revised, reviewed or published, and specifically before every publication of docs\u002Fpaper.md. Covers what legitimate journals actually gate on, in order of gating — the editor's desk screen (scope, novelty, package problems), what reviewers judge (methodological soundness, claims the evidence can carry, statistical and qualitative reporting), EQUATOR reporting guidelines, the ICMJE authorship criteria, COPE's view of text recycling, the near-universal generative-AI policy (no AI author, disclosure, author responsibility), TOP 2025 transparency standards, how to verify a journal is legitimate, submitting without an institution, and a pre-submission checklist. Triggers on requests to evaluate, review, prepare or submit a paper, to judge publishability, to write a declarations or data-availability section, to disclose AI use, or to check a journal.","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fwriting-craft\u002Fskills\u002Fjournal-publishability-criteria\u002FSKILL.md",[],[]]