Tag

#LLM

7 posts

01August 20, 2026The Agentic Loop: When AI Should Think, Act, and LoopEvery AI system you rely on — scripts, watchdogs, coding agents, Grok bots — runs the same core loop: observe, decide, act, verify. This post breaks down that loop, the two ways to put a human around it, and a practical framework for deciding when an agentic loop earns its cost and when a plain script wins.
02June 17, 2026Why AI Hasn't Replaced Engineers — And Won't Anytime SoonAI writes most of the code now, layoffs are running at over 1,000 a day, and execs are saying they hired zero new engineers this year. Yet companies are quietly rehiring the same people they let go, agents that score 80% on SWE-bench drop to 12% on real industrial tasks, and the scarce resource has shifted from typing to judgment. This is a data-backed breakdown of why the discipline survives even as the job changes.
03June 5, 2026Tokens Are Compute — Why Your LLM Bill Is Really a GPU BillMost engineers reason about LLMs in words, characters, or messages. The model sees none of that — it sees tokens, and tokens are compute someone's GPU has to run. This post traces what a token actually is, why output costs 3–10x more than input, the five-step journey of an API call, and the four cost levers most teams never pull.
04May 31, 2026PageIndex: Vectorless RAG That Reasons Through DocumentsTraditional RAG chops documents into arbitrary chunks, embeds them, and hopes cosine similarity finds the right one. PageIndex throws that out — it builds a hierarchical table-of-contents tree and lets an LLM reason its way to the right section, the way a human expert flips to the right chapter. No embeddings, no vector DB. It hit 98.7% on the FinanceBench benchmark.
05April 27, 2026Why SpaceX Is Paying $60B for CursorOn April 21, 2026, SpaceX announced an option to acquire Cursor for $60 billion later this year. This post unpacks the actual architecture that makes Cursor feel like a senior engineer — local indexing, Merkle-tree change detection, Tree-sitter AST chunking, Turbopuffer vector search, and the Composer agent model — then breaks down the three strategic reasons this deal is really about who owns the AI stack.
06April 9, 2026Claude Code Leaked: 512K Lines of AI Agent ArchitectureOn March 31, 2026, Anthropic accidentally shipped the full Claude Code source — 512,000 lines of TypeScript — inside an npm package. The resulting clean-room rewrite became the fastest-growing GitHub repo in history. Here's exactly what the architecture reveals, and what you can take from it as an engineer.
07November 5, 2025TOON vs JSON: Supercharge Your LLM Prompts & Cut Token CostsToken-Oriented Object Notation (TOON) is a compact, LLM-optimized alternative to JSON for serializing structured, mostly flat/tabular data. By removing repeated field names, quotes and redundant punctuation, TOON reduces token usage by roughly 30–60% in real-world AI workflows, leading to lower API bills, larger usable context windows, and often better model retrieval accuracy.

Loading…