Scrollypedia — Deep Dives Into AI, Hardware & Modern Science
Clear explanations of AI, agents, and the infrastructure that powers them.
Latest articles
TPU v8t vs v8i: Why Google Split Training and Inference
— Google's eighth-generation TPU is two chips, not one. TPU 8t for pre-training (12.6 PFLOPs FP4); TPU 8i for serving (288 GB HBM, new Boardfly topology). What the split means for AI agent workloads, real cost math, and where NVIDIA, AWS, AMD stand.
Claude vs Gemini in 2026: Which Model to Use When (And Why It's Not What the Marketing Says)
— If you've been watching the AI lab leaderboard wars, here's the punchline most people miss: at the very top of the frontier, the "which model is smartest" debate is essentially over. So why do experienced users have strong opinions?
The Case for Centralizing Your LLM Traffic (With Real Numbers)
— Every AI agent in your system makes LLM calls independently — no shared cache, no cost visibility, no PII controls. Here is what happens when you route all of that traffic through a single proxy layer, illustrated with real data from a multi-agent stock trading pipeline.
Skills: The Playbooks That Turn Your AI From Intern to Operator
— We built 4 skills for a live trading agentic system. They turned a 12-step manual checklist into a single slash command. Here is what we learned about teaching AI to run your infrastructure.
Zero Trust for AI: Why Your Agents Need Badges, Not Keys
— We are currently handing AI agents the digital equivalent of a Master Key. Here is how Zero Trust principles apply to the new world of autonomous multi-agent systems.
Inside Alpamayo: How Nvidia is Using Latent Tokens to Give Robots Human-Like Instinct
— Robots that "think out loud" are too slow for the real world. Nvidia’s Alpamayo-R1 moves from words to Latent Tokens—giving AI a "gut feeling" for physics.
AI Gets Cheaper. AI Gets Physical.
— NVIDIA's new chips, Boston Dynamics' latest robot, and Meta's nuclear deal might seem unrelated. But they tell a single story: price drops change everything.
The Universal Link: Agentic Mesh with MCP
— Standardizing how AI models talk to data has always been a "Tower of Babel" problem. MCP is changing that by creating a common "plug" for AI tools.
The Memory Wall: Why AI Can't Run on Regular Computers
— Why does AI need specialized hardware? It's not just about processing power. It's about a 30-year-old problem called the Memory Wall—and understanding it changes everything.
Bare Metal AI: Why Training Frontier Models Demands Direct Hardware Access
— Virtualization adds 15-25% overhead that compounds over weeks of continuous training. For frontier models, that translates to months of additional compute time and millions in extra costs.
Apple Silicon Deep Dive: How Unified Memory Runs 671B AI Models
— A Mac Studio costs around $10,000. An NVIDIA A100 costs over $15,000. The Mac runs 671B parameter models locally. The A100 can't. Here is why.
World Models: How AI Finally Learned to Understand Physics
— The shift from pattern-matching to physical reasoning marks the beginning of truly embodied intelligence. We move from AI that can write poetry to AI that knows what happens when you drop a glass.
Neuromorphic Computing: The Brain-Inspired Future of AI
— Right now, your brain is doing something remarkable while consuming just 20 watts of power. Welcome to the world of neuromorphic computing—where energy savings of 25 to 500 times are measured results, not just projections.
TPU vs GPU: The Battle for AI Supremacy in 2025
— In-depth analysis of Google's Ironwood TPU v7 vs NVIDIA's Blackwell B200. Compare performance, costs, energy efficiency, and real-world deployment strategies with verified 2025 statistics.
How AI Just Compressed 800 Years of Materials Science Into Days
— The race to discover the materials powering tomorrow's technologies has entered warp speed. Here's what's happening behind the scenes.
CoreWeave: The AI Hyperscaler Reshaping Cloud Infrastructure
— From crypto mining to a $55 billion backlog, CoreWeave has built the infrastructure backbone for OpenAI and the AI revolution. Here is how they did it.
Apps & tools
Toko-Mo-Co
— Open-source LLM proxy with cost tracking, caching, PII redaction, and a real-time dashboard. Drop-in for OpenAI, Anthropic, and Gemini.
Newsletter Mailer
— A minimal newsletter admin panel for composing, previewing, and sending emails.
Gear, books & tools