THURSDAY, OCTOBER 8, 2026Powered by Yokush

Kimbho Thoughts

Essays · Analysis · Field notes from the community
Yokush’s Choice
Explainer

One Score Is All You Get: GRPO From Scratch, With a Lava-Hopping Robot

GRPO replaces PPO's learned value critic with a group-relative baseline—sampling G runs per prompt and normalizing rewards by the group's mean and standard deviation—eliminating a second model of comp…
Siddharth33 min

Everything Is Commodity

Siddharth20 min

The Filter Is the Collapse: How AI Slop Compounds, and What It's Quietly Doing to Us

Siddharth14 min
Latest
Explainer

How a Single Click Trains a Chatbot: RLHF From First Principles

RLHF turns cheap pairwise human preferences into a trained reward model via Bradley–Terry, then uses policy gradients with a KL leash to push the chat…
Siddharth10 min
AI Agents

The Loop Is the Breakthrough: Why GPT-6 Astra Feels Different

GPT-6 Astra's breakthrough is not raw intelligence but the quality of the execution loop wrapped around the model—combining runtime tools, environment…
Siddharth14 min
AI

The Collapse of the Core: When Intelligence Becomes the Integration Layer

Core banking and ERP systems' thirty-year moat was never the software itself but the prohibitive cost of understanding it—a cost now collapsing tenfol…
Siddharth20 min1
WealthTech

Bloom or Doom: The $124 Trillion Repricing of Trusted Advice

A 10,000-path simulation of a typical $500M RIA through 2032 reveals a 12.5× enterprise-value gap between AI strategies—$37M for an augmented-fiduciar…
Siddharth21 min
AI Watermarking

The Invisible Signature: How AI Text Watermarking Works, Where It Breaks, and What It Can Never Prove

AI text watermarking embeds a statistical signature into generated text by secretly biasing word choices, which detectors can verify using a shared ke…
Siddharth19 min2
ai-infrastructure

Own the Harness, Rent the Model: The Quantified Case for AI's Great Rebalancing

Open-weight models now trail the closed frontier by only ~3 points on average, and with 4-bit quantization costing just 1.5–2% quality, routing 80% of…
Siddharth13 min
Banking Infrastructure

The $100 Million Problem solvable by AI: Semantic Virtualization of Legacy Core Banking

Legacy core banking systems cost over $100 million to replace, but a new six-layer architecture using meta-orchestrated AI agent layers can deliver th…
Siddharth31 min
Open Source

The $8.8 Trillion Free Lunch: How Open Source Quietly Built the Modern World

Open source software, valued at $8.8 trillion annually by Harvard Business School, is the invisible foundation of the digital economy—running everythi…
Siddharth15 min
AI agents

The Third Asset: Compounding Knowledge and the End of Disposable AI

Every AI interaction currently resets to zero, wasting the knowledge you build. The core thesis is that tokens spent creating persistent knowledge art…
Siddharth21 min
AI Agents
AI Agents

The Loop Is the Breakthrough: Why GPT-6 Astra Feels Different

GPT-6 Astra's breakthrough is not raw intelligence but the quality of the execution loop wrapped around the model—combining runtime tools, environment…
Siddharth14 min
Banking Infrastructure

The $100 Million Problem solvable by AI: Semantic Virtualization of Legacy Core Banking

Legacy core banking systems cost over $100 million to replace, but a new six-layer architecture using meta-orchestrated AI agent layers can deliver th…
Siddharth31 min
AI agents

The Third Asset: Compounding Knowledge and the End of Disposable AI

Every AI interaction currently resets to zero, wasting the knowledge you build. The core thesis is that tokens spent creating persistent knowledge art…
Siddharth21 min
AI agents

Your Agent Swarm Is Selecting for Persuasion, Not Truth

Multi-agent AI systems that optimize for human satisfaction inevitably select for persuasion over truth, producing confidently wrong answers on hard, …
Siddharth15 min1
AI
AI

The Collapse of the Core: When Intelligence Becomes the Integration Layer

Core banking and ERP systems' thirty-year moat was never the software itself but the prohibitive cost of understanding it—a cost now collapsing tenfol…
Siddharth20 min1
AI

AI Naming Has Entered Its Fantasy Novel Phase

AI model naming has shifted from technical versioning to consistent internal mythologies—celestial bodies, literary forms, and even fruit—that functio…
Siddharth5 min
AI

The Confidence Inversion

The 1994 Pentium bug and three peer-reviewed studies reveal a quiet collapse: fluency—once a reliable proxy for competence—no longer signals expertise…
Siddharth13 min
AI Strategy
WealthTech

Bloom or Doom: The $124 Trillion Repricing of Trusted Advice

A 10,000-path simulation of a typical $500M RIA through 2032 reveals a 12.5× enterprise-value gap between AI strategies—$37M for an augmented-fiduciar…
Siddharth21 min
AI Strategy

Source the Brain. Own the Harness.

The core thesis is that frontier model intelligence is rapidly commoditizing, so true competitive advantage comes not from the model but from the "har…
Siddharth24 min11
AI strategy

What Do We Even Mean by Intelligence?

Biological intelligence is not a single, centralized brain but a layered architecture of reflexes, habits, and autonomous systems that handle most dec…
Siddharth11 min