Sam Joyce sammyjoyce

Explaining Mechanistic Interpretability to a Web Developer

The Problem

Imagine you inherited a massive web application with no documentation, no comments, and minified code. When it makes decisions (like rejecting a user login or flagging a transaction), you have no idea why. Now imagine this is handling millions of dollars in financial decisions. That's the current state of AI models.

What We're Building

Think of it as Chrome DevTools for AI models. Just like DevTools lets you inspect network requests, see which functions are called, and debug JavaScript execution, we're building tools to inspect what happens inside AI models when they make decisions.

The Web Development Analogy

Overview and Project Purpose

asinit is a Zig-based code scaffolding and update CLI designed to generate and maintain modular project codebases. It composes multiple feature “modules” (each with templates and code modifications) into a cohesive project and later applies updates to those modules even if the user has modified the code oai_citation:0‡github.com. The CLI emphasizes deterministic, offline code generation: given the same module set, it will always produce the same project structure, and it tracks generated content so it can intelligently merge upstream changes into user-edited code oai_citation:1‡github.com oai_citation:2‡github.com. In essence, asinit acts as a **modu

	--------------------------------------------------------------------------------
	-- init.lua for "react_tool_use_multi_model_example" Plugin
	--------------------------------------------------------------------------------
	-- This plugin demonstrates:
	-- 1) Multiple role-specific agents using different models:
	-- • Architect uses a cloud-based model (cloud_llm)
	-- • Reviewer uses a thinking model (thinking_llm)
	-- • Coder uses a local model (local_llm_http)
	-- 2) A ReAct-style workflow where agents can natively call tools.
	-- 3) A search tool that uses ripgrep ("rg") on the local machine.

	package bundler

	import (
	"context"
	"fmt"
	"io"
	"log"
	"net/http"
	"os"
	"os/exec"

	import { startSpan } from "@sentry/remix";
	import type { StartSpanOptions } from "@sentry/types";
	import {
	type AnyColumn,
	type AnyTable,
	type BuildQueryResult,
	type DBQueryConfig,
	type DrizzleTypeError,
	type Equal,
	type ExtractTablesWithRelations,

	import { ButtonHTMLAttributes, ReactNode } from 'react';
	import * as stylex from '@stylexjs/stylex';

	const baseStyles = stylex.create({
	base: {
	backgroundColor: "#3661DA",
	border: "1px solid #3661DA",
	color: "#fff",
	fontWeight: 600,
	fontFamily: "Inter, sans-serif",

	import { NextjsSite, StackContext, use } from "sst/constructs";
	import { DNS } from "./dns";
	import { Storage } from "./storage";
	import { Auth } from "./auth";
	import { API } from "./api";

	export function Site({ stack, app }: StackContext) {
	const dns = use(DNS);
	const api = use(API);
	const auth = use(Auth);

	import {HostedZone} from "aws-cdk-lib/aws-route53";
	import {StackContext} from "sst/constructs";

	const PRODUCTION = "domain.com";
	const DEV = "dev.domain.com";

	export function DNS(ctx: StackContext) {
	if (ctx.stack.stage === "production") {
	const zone = new HostedZone(ctx.stack, "zone", {
	zoneName: PRODUCTION,

	import { StackContext, use } from "sst/constructs";
	import { EmailIdentity, Identity } from "aws-cdk-lib/aws-ses";
	import { TxtRecord } from "aws-cdk-lib/aws-route53";
	import { DNS } from "./dns";

	export function Email(ctx: StackContext) {
	const { stage } = ctx.stack;
	if (stage !== "production" && stage !== "dev") return;
	const dns = use(DNS);
	const email = new EmailIdentity(ctx.stack, "identity", {