After my neighbor's brain hemorrhage, neither money nor AI could take responsibility for her dog - only human trust could. This got me thinking about what we're actually losing to automation.
Have you ever built a house? The most important person is the site manager you have to trust and constantly communicate with. An architect and a mason are not enough. So where is the role intersecting the craft, feasibility and the actual circumstances in AI?
I wrote a guide on using category theory concepts to compose AI tool calls more effectively. It shows how treating functions as typed morphisms with contracts enables both sequential (monadic) and parallel (applicative) composition patterns in GPT-4's function-calling API.
AI chat models can call small “tool” functions. This post shows how to compose those tools (chaining G→M, e.g., generate → tone-adjust) and how to decompose a user request into subtasks (e.g., flights + hotels), with short Python snippets. It also touches on iterative refinement as a kind of fixed-point convergence—why repeated generate→edit cycles tend to stabilize.
When LLMs orchestrate tool calls, they routinely generate malformed invocations like search_web({}) that crash at runtime. There's no type safety, no convergence guarantee, and no provenance tracking.
This tutorial explores how AI models can solve complex, multi-step queries by composing and decomposing function calls - similar to how we break down problems ourselves.
The approach demonstrates how an AI can:
- Decompose a complex task into subtasks
- Execute each part through separate function calls
- Compose results into a final, structured answer
The article connects this to fixed point theory from mathematics, showing how iterative refinement converges to a stable "consensus" result. Includes practical code examples using OpenAI's API.
A beginner-friendly guide explaining how modern AI models like GPT-4 can call functions to retrieve live data and take actions, rather than just generating text. Includes hands-on Python examples with OpenAI's function calling API, plus insights on safety guardrails that prevent misuse. Learn why this makes AI responses more accurate and structured than pure text generation.
New analysis of 70+ blockchain-AI systems reveals most "decentralized" AI still centralizes compute off-chain, using tokens mainly for coordination. Only 10% of participatory AI projects give stakeholders real control over models. The conceptual frameworks exist - from ETHOS governance to federated taxonomies - but the gap between theory and practice exposes decentralization as mostly camouflage for concentrated power.
Zero-knowledge proofs now verify 13-billion parameter AI models in under 15 minutes with 200KB proofs, while GPU enclaves achieve 99% native performance. This deep dive covers the cryptographic infrastructure enabling trustless, privacy-preserving AI at scale - from zkML and Byzantine-robust federated learning to production TEEs cutting inference costs by 90%.
Research shows humans form deep emotional bonds with objects through memory externalization, identity construction, and sentimental value - mechanisms AI systems are now inadvertently (or deliberately) exploiting. This comprehensive review across psychology, AI, and HCI reveals how understanding object attachment theory could transform AI design from exploitative to genuinely supportive, while warning about the manipulation risks when vulnerable users anthropomorphize AI during social isolation.
