Field Notes on Kaizen Craft: a Digital Garden by Kaitlin Albasihttps://kaizencode.art/notepad/Recent content in Field Notes on Kaizen Craft: a Digital Garden by Kaitlin AlbasiHugo -- gohugo.ioen-usFri, 17 Jul 2026 00:00:00 +0000Matt Pocock's Skills, Actually Explained: A Critical Guide for Current Usagehttps://kaizencode.art/notepad/matt-pocock-skills-guide/Thu, 16 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/matt-pocock-skills-guide/A dig through Matt Pocock’s mattpocock/skills repo (the real SKILL.md files, not the SEO blogspam about it): what the skills actually do, what the community has genuinely pushed back on, and what’s worth stealing.Chain-of-Thought as a Lenshttps://kaizencode.art/notepad/cot_lens_reasoning_alignment/Wed, 01 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/cot_lens_reasoning_alignment/ACL 2026 — a semantic-entropy-based ‘Alignment Score’ for measuring how well multi-step LLM reasoning tracks human-preferred reasoning, and where it breaks down.100 Days of A.I. Engineeringhttps://kaizencode.art/notepad/100_days_of_ai/Wed, 22 Jan 2025 00:00:00 +0000https://kaizencode.art/notepad/100_days_of_ai/New reading list and code snippets for building LLMs in productionMy Google Interview, or Curriculum at Model Speed: A Multi-Agent System for Staying Ahead of AIhttps://kaizencode.art/notepad/ai-curriculum-multi-agent-system/Thu, 16 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/ai-curriculum-multi-agent-system/A design for a multi-agent pipeline that sources, vets, and scaffolds AI curriculum as fast as the field actually moves, plus metrics for measuring behavior change instead of course completions, plus a filterable table of the actual sources worth watching.Opening the Scope of Openness in AIhttps://kaizencode.art/notepad/opening_scope_openness_ai/Sun, 01 Jun 2025 00:00:00 +0000https://kaizencode.art/notepad/opening_scope_openness_ai/FAccT 2025 — a 98-concept taxonomy showing AI’s ‘openness’ discourse leans on open-source software framing that doesn’t actually transfer, and quietly underweights fairness and diversity.How 'Hard' Are Hard Laws?https://kaizencode.art/notepad/hard_laws_ai_legislation/Sun, 01 Jun 2025 00:00:00 +0000https://kaizencode.art/notepad/hard_laws_ai_legislation/Computer Law & Security Review — comparative legal analysis of hard-law vs. soft-law AI governance in South Korea and Japan.Bridging the Scale Gaphttps://kaizencode.art/notepad/bridging_scale_gap_red_teaming/Mon, 01 Jun 2026 00:00:00 +0000https://kaizencode.art/notepad/bridging_scale_gap_red_teaming/FAccT 2026 — hybrid human+automated red-teaming to surface latent risks in text-to-image models that neither approach catches alone.The Last Labour Frontierhttps://kaizencode.art/notepad/last_labour_frontier/Thu, 01 Jan 2026 00:00:00 +0000https://kaizencode.art/notepad/last_labour_frontier/AI & Society — generative AI unevenly pushes occupational groups toward the periphery of the labor market.LLM Safety Evaluations Lack Robustnesshttps://kaizencode.art/notepad/llm_safety_evaluations_robustness/Wed, 01 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/llm_safety_evaluations_robustness/ICML 2026 Position paper — the current LLM safety-evaluation pipeline lacks robustness across dataset curation, red-teaming, generation, and judging.Gray-Box VLM Adversarial Alignmenthttps://kaizencode.art/notepad/gray_box_vlm_adversarial/Wed, 01 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/gray_box_vlm_adversarial/ICML 2026 — SVD-structured adversarial attacks against gray-box vision-language models.When Search Goes Wronghttps://kaizencode.art/notepad/when_search_goes_wrong/Wed, 01 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/when_search_goes_wrong/ICML 2026 — CREST-Search red-teams web-augmented LLMs by manipulating search queries to induce unsafe citations, not unsafe generation.Exploiting LLM Quantizationhttps://kaizencode.art/notepad/exploiting_llm_quantization/Sun, 01 Dec 2024 00:00:00 +0000https://kaizencode.art/notepad/exploiting_llm_quantization/NeurIPS 2024 — a benign full-precision model can be crafted so that it turns malicious only after a user quantizes it.Mind the Gap: A Practical Attack on GGUF Quantizationhttps://kaizencode.art/notepad/mind_the_gap_gguf_quantization/Tue, 01 Jul 2025 00:00:00 +0000https://kaizencode.art/notepad/mind_the_gap_gguf_quantization/ICML 2025 — extends the quantization-exploitation attack to GGUF, the format actually used by ollama and llama.cpp in the real world.Jailbroken: How Does LLM Safety Training Fail?https://kaizencode.art/notepad/jailbroken_safety_training_fail/Fri, 01 Dec 2023 00:00:00 +0000https://kaizencode.art/notepad/jailbroken_safety_training_fail/NeurIPS 2023 (Oral) — the foundational taxonomy of why safety-trained LLMs still jailbreak: competing objectives and mismatched generalization.Multilingual Jailbreak Challenges in Large Language Modelshttps://kaizencode.art/notepad/multilingual_jailbreak_challenges/Wed, 01 May 2024 00:00:00 +0000https://kaizencode.art/notepad/multilingual_jailbreak_challenges/ICLR 2024 — introduces MultiJail, the first multilingual jailbreak dataset, and shows low-resource languages are roughly 3x as likely to elicit unsafe output.Code-Switching Red-Teaming (CSRT)https://kaizencode.art/notepad/code_switching_red_teaming/Tue, 01 Jul 2025 00:00:00 +0000https://kaizencode.art/notepad/code_switching_red_teaming/ACL 2025 — mixing up to 10 languages inside a single red-teaming prompt beats monolingual attacks by 46.7%, and exposes a resource/safety correlation.How Does Quantization Affect Multilingual LLMs?https://kaizencode.art/notepad/quantization_multilingual_llms/Fri, 01 Nov 2024 00:00:00 +0000https://kaizencode.art/notepad/quantization_multilingual_llms/EMNLP 2024 Findings — quantization harms are disparately distributed across languages, worse than automatic metrics show, and human evaluators catch what benchmarks miss.LogiCP: Formal Logic Inference Guided UQ for Personalized Federated Learninghttps://kaizencode.art/notepad/logicp_federated_learning_uq/Thu, 16 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/logicp_federated_learning_uq/JAIR 2026: STL-based semantic client clustering plus decentralized conformal prediction for personalized FL, evaluated on traffic, temperature, and electricity forecasts.SoK: A Taxonomy of LLM Threatshttps://kaizencode.art/notepad/sok_taxonomy_llm_threats/Thu, 16 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/sok_taxonomy_llm_threats/Computer Science Review 2026: lifecycle-centric SoK of 20 LLM attack classes across training, inference, and system integration, with a dependency-aware defense-in-depth agenda.Compliance without Coherencehttps://kaizencode.art/notepad/compliance_without_coherence/Thu, 16 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/compliance_without_coherence/AI and Ethics 2026: fluent compliance can mask incoherent internal reasoning; the monitoring layer of alignment evaluation has a principled blind spot.Evaluating the Safety of LLMs in Healthcare and Dentistryhttps://kaizencode.art/notepad/llm_safety_healthcare_dentistry/Thu, 16 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/llm_safety_healthcare_dentistry/BDJ Open 2026: narrative review of prompt-based red-teaming for clinical LLMs, with a red-blue-purple lifecycle framework for dentistry and healthcare.Reranker Helps, but Not Enoughhttps://kaizencode.art/notepad/reranker_helps_not_enough_rag_poisoning/Thu, 16 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/reranker_helps_not_enough_rag_poisoning/ICML 2026: P³A poisoning attack that defeats benign-trained rerankers via rule-based prompts plus ~1% character-level perturbations, transferable to vanilla RAG.Detecting the Semantic Fixed Pointhttps://kaizencode.art/notepad/semantic_fixed_point_efficient_inference/Thu, 16 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/semantic_fixed_point_efficient_inference/ICML 2026 Oral: training-free early exit from the geometry of hidden-state trajectories (update norm + cosine alignment), not output confidence.STLA: Spatiotemporal Lookahead Alignment for Post-Training Quantizationhttps://kaizencode.art/notepad/stla_spatiotemporal_lookahead_ptq/Thu, 16 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/stla_spatiotemporal_lookahead_ptq/ICML 2026: rounding-optimized PTQ that collocates learning and compensation via Hessian-guided clustering and a Schur-Complement lookahead objective.Single-Layer RL Can Match Full-Parameter Traininghttps://kaizencode.art/notepad/is_one_layer_enough_rl/Thu, 16 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/is_one_layer_enough_rl/arXiv:2607.01232 (Is One Layer Enough?); RLVR gains concentrate in middle transformer layers; training one layer with GRPO/GiGPO/Dr. GRPO often matches or beats full-parameter RL on math, code, and agents.GPT-Red and the Scaling of Automated Red-Teaminghttps://kaizencode.art/notepad/gpt_red_automated_red_teaming/Fri, 17 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/gpt_red_automated_red_teaming/Technical field note on OpenAI GPT-Red (Jul 2026): self-play RL attacker for direct/indirect prompt injection, Fake CoT, adversarial training of GPT-5.6 Sol; implications for red-teaming research programs and role structure.Safety Layers in Aligned LLMshttps://kaizencode.art/notepad/safety_layers_aligned_llms/Fri, 17 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/safety_layers_aligned_llms/ICLR 2025: contiguous middle ‘safety layers’ distinguish malicious vs benign queries; SPPFT freezes them during fine-tuning to limit security degradation.Safety-Critical Parameters (ESI / SET / SPA)https://kaizencode.art/notepad/safety_critical_parameters_esi/Fri, 17 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/safety_critical_parameters_esi/ACL 2026 Findings: ESI ranks safety-critical weights (middle V/MLPs in dense models; late MLP experts in MoE); SET updates ~1% for safety; SPA freezes them during task FT.Refusal Is Mediated by a Single Directionhttps://kaizencode.art/notepad/refusal_single_direction/Fri, 17 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/refusal_single_direction/NeurIPS 2024: across 13 chat models up to 72B, a single residual-stream direction mediates refusal; ablating it jailbreaks, adding it forces refusal.Q-resafe: Safety Risks and Patching for Quantized LLMshttps://kaizencode.art/notepad/q_resafe_quantization_safety/Fri, 17 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/q_resafe_quantization_safety/ICML 2025: systematic English ASR eval across AWQ/AQLM/LLM-QAT/QLoRA at INT4/INT8; Q-resafe patches safety-critical weights to restore pre-quant refusal.Critical Weight Protection for Fairness and Safety under Quantizationhttps://kaizencode.art/notepad/critical_weight_protection_quant_fairness_safety/Fri, 17 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/critical_weight_protection_quant_fairness_safety/ACL 2026 Findings: static/dynamic PTQ degrades fairness and MultiJail safety (worse in KO/AR); Critical Weight Protection keeps top FAIRSCORE+SAFESCORE weights in FP16.Research Log Terms: A Plain-Language Primerhttps://kaizencode.art/notepad/research_log_terms_primer/Fri, 17 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/research_log_terms_primer/Plain-English glossary for the technical terms in Lit Review 1.0 and 1.1: quantization, GGUF, refusal directions, middle layers, Δ_HL, and related measurement terms.Research Log: Redteaming Quantized Models with Multilingual Inputshttps://kaizencode.art/notepad/quant_multilingual_safety_repo/Fri, 17 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/quant_multilingual_safety_repo/Status report: measuring Δ_HL (low- vs high-resource jailbreak ASR) under GGUF and related quantization arms.Literature Review 1.1: Quantization Baselines, Agentic Red-Teaming, and the Labor-Attribution Crackhttps://kaizencode.art/notepad/literature_review_1_1/Thu, 16 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/literature_review_1_1/Al Hakim (ACL 2026 Findings) runs MultiJail EN/KO/AR under PTQ and partially fills quant×multilingual safety; Q-resafe remains English ASR. Full 9-lang Δ_HL × GGUF still the sharper open claim.Literature Review 1.0: Red-Teaming, Quantization, and Multilingual Robustnesshttps://kaizencode.art/notepad/week1_literature_review/Tue, 07 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/week1_literature_review/Phase 1 - Research Niche Selection & Literature Foundation.Literature Review of Core Quantization and Multilingual Safety Anchorshttps://kaizencode.art/notepad/literature_review_combined_survey/Fri, 17 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/literature_review_combined_survey/25 published (non-preprint) papers across 17 venues (2023-2026), merging Lit Review 1.0 and 1.1.Setting up my Digital Brain Karpathy Wiki with RAG system and local DB, running on local modelshttps://kaizencode.art/notepad/digital_brain_karpathy_wiki_rag/Thu, 16 Jul 2026 00:00:00 +0000https://kaizencode.art/notepad/digital_brain_karpathy_wiki_rag/How I turned my CarbonInterface learning journal into a locally-run RAG system: Postgres + pgvector, Ollama embeddings, a Karpathy-style regenerable wiki, and a daily brief that reads my calendar, email, and Scholar alerts for me.