Ninad

Why governments and companies inevitably bloat, as a calculus problem: Measuring area under a curve, your first rectangles are huge. Great ROI. But each rectangle has a fixed cost, so you go smaller and smaller chasing less and less area. Early on: broad policies that help almost everyone. Later: guardrails for every last constituent. Eventually the overhead of accounting for everyone makes the whole thing work poorly for everyone. And you can't delete rectangles. Removing the dead feature angers a loud few; keeping it costs everyone a little. So the ratchet turns one way. Degrees of freedom only shrink with scale. Not from bad decisions — from good ones. Every optimal move forecloses the next. It's basically entropy.

A principle worth tattooing: the LLM is optional for correctness, never optional for usability. Rip it out and the system still returns the right numbers — it just reads like a robot. Add it back and it sounds human. The numbers never change.

The silent killer in LLM pipelines: one model running everything keeps merging steps. It computes when it should route. It paraphrases when it should pass through. Numbers drift and nothing looks broken. Fix: a router LLM structurally forbidden from writing prose. It can't editorialize if the output format won't let it.

When the task has a right answer — counts, totals, compliance thresholds — reliable LLM systems converge on the same shape: extract → compute → render. LLM parses text into JSON. Deterministic code does the math. LLM writes prose around numbers it isn't allowed to invent. The boring layer in the middle is where correctness lives.

15% money supply CAGR is not unreasonable -- growth (India, China), war-finance (UK WWI/WWII), and stagflation (UK 1970s) -- have witnessed 10–25% in nominal local currency terms.

State of AI Enlightenment comes when we focus on narrow domain AI instead of fundraising-season AGI.

O3 is impressive but it is ultimately a CoT model and will suffer the same issues as every other LLM — it can drift and meander across reasoning steps. LLMs must balance both precision and accuracy, unlike domain-specific use cases such as FSD.

@miranetwork stops AI agents from straying into error scenarios and makes AI actually useful.

I’m oversimplifying this but reasoning and AGI cannot happen without fact verification. Error rates are high and they compound with each reasoning step. It’s like a game of jenga.

Here’s a way to think about it. Frontier models today can have error rates as high as 30%. If you have an autonomous agent or multi-step reasoning stitch together multiple steps, the output error compounds *massively*. In today’s baseline (orange), forget about building anything that is agentic and useful.

Truth has always been hard to pin down, because in part it involves an interpretation of facts. In a world full of AI generated content, truth will become harder to define and protect. While we’re rightly basking in the glory of all AI can do, it is time we start rolling up our sleeves to build the necessary defenses and scaffolding to protect the truth.

Broadly I think we move to application specific stacks (hardware, models, inference, UX). Here’s the link to our conversation and how @miranetwork is contributing to making AI actually artificial intelligence. youtu.be/Ig2NZ1LFGms?si=ars2…

This adds up to more than 100%. What.

ah! well have fun in banana town ahead of banana zone :D