AI & Agent
Building intelligent systems with LLMs, autonomous agents, and multi-agent architectures. Covers Strands Agents, Amazon Bedrock, RAG, and real-world AI workflows.
18 articles
I Stopped Chasing New Model Releases. The Hallucinations Didn't Stop.
In half a year the flagship Opus shipped four generations at a flat price, and I used the newest daily — yet my hallucinations never dropped. Six things I couldn't explain, one by one.
Prompt Caching (Part 4): AWS Documents 4,096 — I Measured 1,024
AWS gives Claude Sonnet 4.5 a 4,096-token minimum cacheable prefix. Bisecting on Bedrock put the real threshold at 1,024. Three control models matched their docs; only this row was off.
Prompt Caching (Part 1): What Is Actually Cached
Prompt caching saves the server's prefill compute, not the cost of re-sending tokens. Where the KV cache comes from, why a prefix must match token for token, and why caching can cost more.
Prompt Caching (Part 2): Five Vendors, Three Clouds, One Bill
AWS documents a 4,096-token minimum for Sonnet 4.5 on Bedrock; I measured 1,024. Multipliers, TTL tiers and cross-region behaviour for five vendors — measured kept apart from documented.
Prompt Caching (Part 3): It Is On and Not Hitting — Now What
Cache invalidation almost never errors: inference succeeds, the logs are clean, only the bill grows. Four culprits by frequency, plus a cross-vendor hit-rate self-check you can run today.
Seedance 2.5 Hands-On: 725 Credits Buy 27 Seconds, and the CLI Costs Extra
A ¥69/month membership buys 725 credits — enough for 27 seconds of finished video, about ¥2.5 per second. The Dreamina CLI needs the ¥998/month premium tier, which is where my agent-driven pipeline stopped. Three blockers, one cost calculation done on what I actually paid, and four things wrong with the subtitles.
DeepSeek V4 Pro Goes GA: Third-Party Benchmarks Don't Match the Leaked Scorecard
DeepSeek V4 Pro 0813 launches as a GA release. Artificial Analysis gives it 53 points, ranking 23rd. The leaked agent benchmark table claims 87.9 on Terminal Bench — beating Opus 4.8 — but independent testing shows 79%, a 9-point gap that flips the result. Here's the full cross-check.
Inside Kimi K3's Two Architectural Pillars: KDA and AttnRes, Explained From Zero
Beginner-friendly, analogy-driven walkthrough: KDA turns attention's 'open-book exam' into 'one page of smart notes' — 75% less KV-cache memory, 6.3x faster decoding at 1M tokens. AttnRes turns the residual stream's 'running ledger' into 'a notebook with an index' — 25% better training efficiency at under 2% overhead. By the end you'll know exactly how 2.8T parameters stay standing.
Kimi K3 Is Here: 2.8 Trillion Parameters, Fully Open Source — This Time It's Different
Moonshot AI releases Kimi K3: 2.8 trillion parameters, the world's first open-source 3T-class model, 1M-token context, and native multimodality. A deep dive into the KDA and AttnRes architecture innovations, its #4 global ranking on Artificial Analysis, pricing that matches Claude Sonnet 5, and the four long-term ways an open frontier model reshapes the industry.
On the Eve of the Final, I Had fable 5 Run Another Million Simulations: Messi Kicked More Than Half the Gap Away
All six predictions from the last article hit: Spain 2-0 France, Argentina 2-1 England, and the 41.8%-probability Spain–Argentina final came true. Before the final I fed the semifinal data back to fable 5 and re-ran a million simulations: Spain's win probability dropped from 54.1% to 51.8%, Argentina's rose from 45.9% to 48.2% — Messi, leading the tournament scoring chart, took back the points the model had docked for his age. Plus the story of a photo taken 19 years ago: Messi and baby Yamal.
I'm a Messi Fan, But I Asked the Strongest Model Ever — fable 5 — to Predict the World Cup. The Answer Hurt.
The 2026 World Cup semifinals are set, so I had Claude Fable 5 build a prediction model — Elo ratings, age-curve and tournament-experience adjustments, a dedicated penalty-shootout submodel — and run 1,000,000 Monte Carlo simulations. The result is the 'Argentina Paradox': Argentina has the highest probability of reaching the final (66.5%), yet Spain is the most likely champion (37.5%). A sensitivity analysis shows Messi's form is literally the dividing line.
Slow Down in the AI Wave: A Reality Check for Employees and Bosses
At least half of today's AI anxiety is manufactured by marketing. First-hand data from PwC, MIT, METR, and DX on the 2026 AI reality: 56% of CEOs report zero ROI, 95% of enterprise GenAI pilots fail, and AI coding delivers ~10% gains, not 10x. Five grounded rules for workers, five for leaders.
Anthropic Goes Big: Claude Fable 5 and Mythos 5 Are Here
On June 9, 2026, Anthropic shipped Claude Fable 5 and Mythos 5 simultaneously: the same underlying model, differing only in guardrails. Fable 5 is SOTA on nearly every benchmark, with the lead growing on longer tasks (Stripe migrated a 50-million-line codebase in one day). Priced at $10/$50 — less than half of Mythos Preview — and already live on Amazon Bedrock. Plus: why did the hidden Mythos suddenly come out of the vault?
How Do I Explain "Agent" to My Wife?
From 'the Lark bot can't do math' to 'raising your own AI lobster' — an AI-agent explainer for normal humans. Thirteen burning questions covering LLMs, tokens, Tools, MCP, RAG, Skills, Memory, Multi-Agent systems, and 2026 model prices. The AI isn't dumb — it just hasn't been raised properly yet.
Amazon AI Strategy 2026: Why the Biggest Player Is the Least Visible
Custom chips, global infrastructure, massive investment — yet Amazon is invisible in the AI race. Here's what's really going on.
WeChat x OpenClaw: Platform Strategy in the AI Agent Era
WeChat's native OpenClaw integration signals a major shift. Why the world's largest messaging app opening up to AI agents matters.
How to Build AI Agents for Ad Creative Generation
Automate ad copywriting, image, and video production with Strands Agents and Amazon Bedrock. A practical, code-first guide.
How to Build a RAG System with LangChain and Elasticsearch
A hands-on guide to building Retrieval Augmented Generation — from vector embeddings to context-enhanced LLM answers.