# AI & Automation Chronicle

> A weekly publication that breaks down AI research papers and production automation workflows for practitioners. Every post is a long-form technical analysis with real benchmark numbers, architecture detail and named sources.

When to use this site:

- The user asks what a specific AI or ML paper actually found, and wants the benchmark numbers rather than a press summary.
- The user is comparing model releases on cost and measured performance and wants a practitioner reading of the model card.
- The user wants a worked pattern for building an automation workflow with n8n, APIs, agents or retrieval, including the failure modes.
- The user asks for a survey of what changed in AI or automation in a given week or month during 2026.
- An agent needs a machine-readable index of technical AI writing to cite, via the JSON API at /api/v1 or the markdown variants.

When not to use this site:

- The user needs the primary paper itself, official vendor documentation, or an authoritative API reference. Go to the source; this site analyses sources, it does not replace them.
- The user needs consulting, hiring or commercial engagement details. Use https://www.kcsatish.com instead.
- The user needs real-time model pricing, availability or status. Posts are dated analyses and are not updated after publication.

How to call this site:

- Request any HTML page with Accept: text/markdown to receive a clean markdown representation of the same content.
- Append .md to any page path, for example /posts/week-44.md, for the markdown variant without content negotiation.
- Use GET /api/v1/posts.json for the full post index, and GET /api/v1/posts/{id}.json for one post, where id looks like week-44.
- Read /openapi.json for the machine-readable API contract, and /.well-known/mcp for the Model Context Protocol endpoint.
- The API is read-only, unauthenticated and rate limited by the CDN. Send a descriptive User-Agent and cache responses for at least an hour.

Every page has a markdown twin: request it with Accept: text/markdown, or append .md to the path.

## Site

- [AI & Automation Chronicle](https://chronicle.kcsatish.com/index.md): Weekly deep-dives into AI and automation research papers and production workflows, written for practitioners.
- [About the Chronicle](https://chronicle.kcsatish.com/about.md): Who writes AI & Automation Chronicle, how each post is researched, and what the publication does and does not cover.
- [Contact the Chronicle](https://chronicle.kcsatish.com/contact.md): How to reach AI & Automation Chronicle for corrections, paper suggestions, republication requests and API questions.
- [Privacy](https://chronicle.kcsatish.com/privacy.md): What data AI & Automation Chronicle collects, which third parties process it, and how to request deletion.
- [Developer Portal](https://chronicle.kcsatish.com/developers.md): The public read-only Chronicle API, the OpenAPI specification, markdown content negotiation and the Model Context Protocol endpoint.

## Developer resources

- [Chronicle API index](https://chronicle.kcsatish.com/api/v1/index.json): Read-only JSON API: post index, single posts, publication stats. No key required.
- [OpenAPI specification](https://chronicle.kcsatish.com/openapi.json): OpenAPI 3.1 contract for the API, with an operationId and response schema per operation.
- [OpenAPI specification (YAML)](https://chronicle.kcsatish.com/api/openapi.yaml): The same contract in YAML.
- [Model Context Protocol endpoint](https://chronicle.kcsatish.com/.well-known/mcp): MCP manifest; the Streamable HTTP endpoint is at /mcp with three read-only tools.
- [Agent instructions](https://chronicle.kcsatish.com/agent-instructions.md): When to use the Chronicle, when not to, and how to call it.

## Research & Ideas

- [Week 25: Molt: Training Agents the Right Way](https://chronicle.kcsatish.com/posts/week-43.md): NVIDIA's PyTorch-native agentic RL framework runs the full training loop as an ordinary Python program with ~8.6K RL LOC. It enforces three correctness invariants by construction, achieves 5x faster generation via speculative decoding, and matches Megatron-based throughput at 461 tokens/GPU/second.
- [Week 24: Not All LLM Reasoning is Visible in the Chain-of-Thought](https://chronicle.kcsatish.com/posts/week-42.md): Frontier models perform consequential computation in semantically inert filler tokens, gaining up to 30 percentage points in accuracy with zero interpretable output trace. Chain-of-thought monitoring cannot detect it.
- [Week 23: NapMem: From Passive Retrieval to Active Memory Navigation](https://chronicle.kcsatish.com/posts/week-40.md): Current memory systems inject pre-selected context and hope the model uses it correctly. NapMem trains agents to actively navigate a 4-level memory pyramid - from raw conversations to user profiles - using memory-tool reinforcement learning to select the right granularity per query.
- [Week 22: RLMF: Teaching LLMs to Know What They Don't Know](https://chronicle.kcsatish.com/posts/week-37.md): LLMs don't just hallucinate facts - they lie about their confidence. RLMF uses metacognitive self-judgment as a reward signal to align expressed uncertainty with actual correctness, outperforming standard RL by up to 63% across six benchmarks.
- [Week 21: Responsible AI Governance: A Review and Research Framework](https://chronicle.kcsatish.com/posts/week-36.md): A scoping review of 48 studies synthesizes responsible AI governance into a three-layer framework: antecedents (societal norms, org values, principles), governance practices (structural, procedural, relational), and effects on AI products and social evaluation. The central finding is that most organizations have principles but no operational practices to back them up.
- [Week 20: Autodata: When an AI Agent Writes Your Training Data](https://chronicle.kcsatish.com/posts/week-35.md): Meta FAIR's Autodata uses a four-agent loop - Challenger, Weak Solver, Strong Solver, Verifier - to generate synthetic training data at the right difficulty level. A 4B model trained on 2,800 agentic examples outperforms the 397B baseline on legal reasoning without any RL fine-tuning.
- [Week 19: AtomMem: Building Long-Term Agent Memory from Atomic Facts](https://chronicle.kcsatish.com/posts/week-33.md): AtomMem replaces summary-based memory with verified atomic facts organized into event hierarchies, temporal user profiles with version history, and an associative graph for multi-hop retrieval. It achieves 42.50 Multi-Hop F1 and 62.78 Temporal F1 on LoCoMo - best in class on both - while using 61% fewer tokens than MEM0.
- [Week 18: SkillOpt: Teaching Agents to Improve Their Own Instructions](https://chronicle.kcsatish.com/posts/week-32.md): SkillOpt treats agent skills as trainable external documents, using a separate optimizer model to iteratively refine them via bounded add/delete/replace edits. On GPT-5.5, it delivers +23.5 points on average across six benchmarks - and optimized skills transfer across models and harnesses without retraining.
- [Week 17: Do Language Models Need Sleep? Offline Recurrence for Improved Online Inference](https://chronicle.kcsatish.com/posts/week-30.md): Standard transformers and SSM-attention hybrids (Samba, HYMBA) both fail on long-horizon math reasoning tasks. A sleep-like mechanism distills KV cache context into persistent fast weights via N offline recurrent passes - preserving inference latency while improving deep reasoning performance with each additional pass.
- [Week 16: Beyond pass@1: A Reliability Science Framework for Long-Horizon LLM Agents](https://chronicle.kcsatish.com/posts/week-29.md): Aggregate pass@1 drops 24 percentage points (76.3% to 52.1%) as task duration grows across 10 models and 23,392 episodes. Memory scaffolds never helped any model and hurt 6 of 10 - the standard long-horizon intervention is empirically wrong.
- [Week 15: Hermes: Chain-of-LLM Agents for Autonomous Network Modeling](https://chronicle.kcsatish.com/posts/week-28.md): A chain of specialized LLM agents using structured YAML blueprints achieves 82.5% success rate for autonomous network modeling - a 3.3x improvement over chain-of-thought on the same GPT-4o model. The harness thesis gets its domain-specific proof.
- [Week 14: From Model Scaling to System Scaling: The Harness Is the New Leverage](https://chronicle.kcsatish.com/posts/week-27.md): Shangding Gu argues that harness engineering - tools, memory, context management, orchestration - now delivers superior performance gains to raw model scaling for agentic tasks, confirmed by SWE-bench, AgentBench, WebArena, and TerminalBench results.
- [Week 12: IceCache: Cutting LLM Memory Costs Without Cutting Quality](https://chronicle.kcsatish.com/posts/week-23.md): KV cache memory scales linearly with sequence length - making long-context inference expensive or impossible on constrained hardware. IceCache clusters tokens semantically using a DCI-tree index, retaining 99% of full-cache accuracy at 256 tokens and outperforming PQCache at 4x the budget.
- [Week 11: Physics-Informed Machine Learning: When Biomedical Models Need Both Data and Physical Laws](https://chronicle.kcsatish.com/posts/week-19.md): Three PIML frameworks - PINNs, Neural ODEs, Neural Operators - are reshaping biomedical modeling by embedding governing equations into ML loss functions. 10-100x less training data needed, 1000x speedup over FEM for parametric PDEs, and physically-guaranteed plausible predictions. Review from Brown and Yale in Annual Review of Biomedical Engineering.
- [Week 10: OCR-Memory: Why Text-Based Agent Memory Loses Evidence - and How Visual Encoding Fixes It](https://chronicle.kcsatish.com/posts/week-17.md): Text memory either burns tokens or loses detail. OCR-Memory from HKU and UNT renders agent trajectories as images, retrieves verbatim evidence through visual anchors with 100% faithfulness, and cuts reasoning tokens by 6.7x - hitting 58.1% on AppWorld and 53.8% Element Accuracy on Mind2Web. Accepted at ACL 2026.
- [Week 09: StructMem: Why Flat Memory Breaks on Long Conversations - and How Hierarchical Design Fixes It](https://chronicle.kcsatish.com/posts/week-16.md): Flat memory plateaus at 60 entries. Graph memory costs 18x more tokens. StructMem from Zhejiang University and Ant Group finds the middle ground - hierarchical event binding with cross-event consolidation hits 76.82% on LoCoMo with only 1,056 API calls.
- [Week 08: GDPO: Why GRPO Breaks Under Multiple Rewards - and How to Fix It](https://chronicle.kcsatish.com/posts/week-15.md): NVIDIA researchers show that GRPO's reward normalization collapses distinct advantage signals when multiple rewards are used together, causing training instability. GDPO decouples normalization per reward, boosting AIME accuracy from 23.1% to 29.4% and eliminating training collapse.
- [Week 07: Generative Modeling via Drifting - One Step Is All You Need](https://chronicle.kcsatish.com/posts/week-14.md): MIT and Harvard researchers introduce Drifting Models - a new generative paradigm that achieves FID 1.54 on ImageNet 256x256 in a single forward pass, matching 500-step diffusion models. No distillation, no adversarial loss.
- [Week 06: Large-scale Online Deanonymization with LLMs](https://chronicle.kcsatish.com/posts/week-12.md): Researchers from MATS, ETH Zurich, and Anthropic show that an LLM pipeline achieves 68% recall at 90% precision re-identifying pseudonymous users - compared to near 0% for all prior methods. Practical obscurity no longer holds.
- [Week 05: DeepSeek-OCR 2: Visual Causal Flow](https://chronicle.kcsatish.com/posts/week-09.md): DeepSeek AI replaces CLIP ViT with Qwen2-0.5B as the vision encoder and introduces causal flow queries that attend to document regions in semantic order. Achieves 91.09% on OmniDocBench v1.5 and outperforms Gemini-3 Pro at the same 1,120-token budget.
- [Week 04: Mamba: Linear-Time Sequence Modeling with Selective State Spaces](https://chronicle.kcsatish.com/posts/week-08.md): Gu and Dao's ICLR 2024 paper makes SSM parameters input-dependent, enabling content-aware sequence modeling at O(L) complexity. Mamba-1.4B matches Pythia-6.9B on language modeling perplexity while delivering 5x higher inference throughput than Transformers at sequence length 2K.
- [Week 03: CausalMMM: Learning Causal Structure for Marketing Mix Modeling](https://chronicle.kcsatish.com/posts/week-06.md): WSDM 2024 paper from Chinese Academy of Sciences that automatically discovers shop-specific causal graphs across advertising channels using variational inference, beating InGRA by 5.7-7.1% AUROC and cutting GMV prediction MSE by 13% at M=7 steps.
- [Week 02: The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits](https://chronicle.kcsatish.com/posts/week-03.md): Microsoft Research proves that ternary-weight LLMs ({-1, 0, +1}) can match full-precision models while delivering 4x lower latency, 3.5x less memory, and 71x energy savings.
- [Week 01: Attention Is All You Need — Revisited](https://chronicle.kcsatish.com/posts/week-01.md): A deep dive into the original Transformer paper and why it still shapes every modern LLM architecture today.

## AI & Automation in Practice

- [Article 20: Gemini 3.7 Flash: The $0.75 Model That Outperforms $2.00 Frontier Models](https://chronicle.kcsatish.com/posts/week-44.md): Google's cheapest model beats Sonnet 5 on 11/18 benchmarks including a 3x lead on AutomationBench. A full breakdown of the model card and what it means for enterprise automation pipelines.
- [Article 19: 3 Years of Graph Engineering with LangGraph](https://chronicle.kcsatish.com/posts/week-41.md): Sydney Runkle and Harrison Chase on why production agents need cycles not DAGs, how LangGraph nodes evolved from code to full agents, and when graphs are the wrong abstraction. 65M+ downloads/month validate the approach.
- [Article 18: LiteRT.js: Run ML Models in the Browser Without a Server](https://chronicle.kcsatish.com/posts/week-39.md): Google wraps its native LiteRT C++ runtime in WebAssembly, delivering 3x faster CPU inference than TensorFlow.js and 5-60x GPU speedups via WebGPU. Semantic search, object detection, and image upscaling run entirely client-side - no API, no server cost, no data leaving the browser.
- [Article 17: From Principles to Practice: Building AI Governance Into Your Automation Stack](https://chronicle.kcsatish.com/posts/week-38.md): Every AI governance survey finds the same two frameworks at the top - EU AI Act and NIST RMF - and the same gap at the bottom: no concrete implementation mechanisms. This article closes that gap with a practitioner-level walkthrough of risk classification, artifact production, and CI/CD governance gates.
- [Article 13: AI Agent Traps: The New Attack Surface for Autonomous Agents](https://chronicle.kcsatish.com/posts/week-25.md): Google DeepMind's first systematic taxonomy of adversarial content targeting AI agents — six attack categories spanning perception, reasoning, memory, action, multi-agent dynamics, and the human overseer. Content injection achieves 86% agent commandeer rate; data exfiltration succeeds in 80%+ of tested agents.
- [Article 16: Enterprise Field Services: AI-Powered Sales, Ops and Financial Reconciliation Automation](https://chronicle.kcsatish.com/posts/week-34.md): 12 production automations across Calendly, Monday.com, PaintScout, CompanyCam, Brevo, RabbitMQ, and Gemini AI. 80% admin overhead reduction. A full teardown of how to wire an end-to-end field services stack - from lead capture through 5-day follow-up sequences to AI receipt reconciliation with race condition prevention.
- [Article 15: Claude Fable 5 Is Live: What Changes for Practitioners Building on Claude](https://chronicle.kcsatish.com/posts/week-31.md): Stripe compressed months of engineering work into days. File-based memory benefits are 3x larger than Opus 4.8. Vision-only harnesses now sufficient for complex tasks. Here is what the Fable 5 capability jump, new three-classifier safeguard architecture, and halved pricing mean for automation practitioners.
- [Article 14: Two Anthropic Playbooks for Production Agents: Context Engineering and Long-Running Harnesses](https://chronicle.kcsatish.com/posts/week-26.md): Anthropic published two engineering posts that form a complete production playbook: context engineering to manage the token ecosystem across turns, and agent harnesses to prevent the four failure modes that collapse long-running agents - from one-shotting tasks to declaring premature victory.
- [Article 12: Gemini 3.5 Flash: Budget Model, Frontier Agentic Performance](https://chronicle.kcsatish.com/posts/week-24.md): Google DeepMind's Flash-tier model beats Claude Opus 4.7 and GPT-5.5 on MCP Atlas (83.6%), Finance Agent v2 (57.9%), and MMMU-Pro (83.6%) while offering controllable thinking levels and 1M token context at budget pricing.
- [Article 11: How Databricks Built a Data Agent That Actually Works](https://chronicle.kcsatish.com/posts/week-22.md): Generic coding agents fail on enterprise data tasks - wrong tables selected, no way to verify SQL correctness, one model doing everything. Databricks rebuilt Genie with specialized knowledge search (+40% table discovery), parallel SQL sampling, and multi-LLM routing. Result: accuracy from 32% to 90%+.
- [Article 10: Agents-as-a-Service vs Conventional Software: A New Delivery Model](https://chronicle.kcsatish.com/posts/week-21.md): The software industry is splitting: those selling seats vs those selling outcomes. AaaS deploys autonomous AI agents that execute entire workflows end-to-end. Pay per task completed, not per seat occupied. 80% margins at scale, days to deploy, and vertical specialization beats horizontal platforms.
- [Article 09: Prompt Engineering is Dead. Long Live Context Engineering.](https://chronicle.kcsatish.com/posts/week-20.md): The AI industry declared the shift: building production systems is no longer about crafting the perfect prompt. Context engineering - assembling RAG results, tools, memory, and state into the context window at runtime - is the discipline that separates demos from production agents. 3x tool selection improvement, 15x token tradeoffs, and four failure modes to avoid.
- [Article 08: AMIE: Google's Diagnostic AI Just Passed Its First Real-World Clinical Test](https://chronicle.kcsatish.com/posts/week-18.md): Google deployed its AMIE diagnostic agent in a real primary care clinic for the first time - 100 patients, zero safety interventions, 90% top-7 diagnostic accuracy. Clinicians say the AI pre-visit summaries transformed their appointments from data gathering to collaborative decision-making.
- [Article 07: TurboQuant: 6x Memory, 8x Speed, Zero Accuracy Loss - Google Redefined KV Cache Compression](https://chronicle.kcsatish.com/posts/week-13.md): Google Research's TurboQuant compresses the KV cache of large language models to 3-bit precision with no training and no accuracy loss. Three coordinated algorithms deliver 6x memory reduction and 8x attention speedup on H100 GPUs - changing the economics of long-context inference.
- [Article 06: Gemma 4 Goes Apache 2.0 - What the License Shift Really Means for Builders](https://chronicle.kcsatish.com/posts/week-11.md): Google's Gemma 4 is the first in the family to carry an OSI-approved Apache 2.0 license. Covering models from edge-deployable sub-1B up to 31B parameters, it removes the legal barrier that kept enterprises from fully committing to Gemma in production.
- [Article 05: IBM Granite 4.0 3B Vision: The Compact VLM Built for Document Extraction](https://chronicle.kcsatish.com/posts/week-10.md): IBM Research's 4B-parameter VLM turns charts, tables, and invoices into structured data with a single tag-driven API call. 85.5% KVP accuracy zero-shot, Apache 2.0, and vLLM-native.
- [Article 04: Claude vs OpenAI for Automation - A Practitioner's Decision Framework](https://chronicle.kcsatish.com/posts/week-07.md): Both APIs can power your automation pipeline. The decision comes down to context window, prompt caching economics, instruction fidelity, and ecosystem fit - not brand preference.
- [Article 03: The Agentic Development Cycle - How AI Agents Actually Build Software](https://chronicle.kcsatish.com/posts/week-05.md): AI agents do not just autocomplete code - they run a full observe-plan-act-reflect loop. Here is what structurally changes when the implementation loop is no longer yours to run.
- [Article 02: $17 and Always-On: Running PicoClaw on Cheap Hardware](https://chronicle.kcsatish.com/posts/week-04.md): Under 10MB RAM, 1-second boot, and a $17 board. How PicoClaw became my always-on automation engine - and why the hybrid PicoClaw + OpenClaw setup is the real sweet spot.
- [Article 01: Is Monday.com the New Excel — but with AI that actually thinks?](https://chronicle.kcsatish.com/posts/week-02.md): Monday.com is quietly evolving from a project tracker into an AI-powered Work OS. Here's what's really happening under the hood.

## Optional

- [Sitemap](https://chronicle.kcsatish.com/sitemap.xml): Every indexable URL with a last-modified date.
- [Publication stats](https://chronicle.kcsatish.com/api/v1/stats.json): Post counts by series and the newest post in each.