diff --git a/web/src/glossary.tsx b/web/src/glossary.tsx
new file mode 100644
index 0000000..adbbd30
--- /dev/null
+++ b/web/src/glossary.tsx
@@ -0,0 +1,52 @@
+import { useState, useEffect } from "react";
+
+// temporary hack to get glossary working
+const GLOSSARY_JSON = {"chain of thought prompting":{"term":"chain of thought prompting","pageid":"8EL7","contents":"<p>Chain-of-thought prompting is a technique which makes a language model generate intermediate reasoning steps in its output.</p>\n"},"chain-of-thought":{"term":"chain-of-thought","pageid":"8EL7","contents":"<p>Chain-of-thought prompting is a technique which makes a language model generate intermediate reasoning steps in its output.</p>\n"},"goodhart's law":{"term":"goodhart's law","pageid":"8185","contents":"<p>Goodhart’s law states that when a measure becomes a target, it ceases to be a good measure.</p>\n"},"the big g,":{"term":"the big g,","pageid":"8185","contents":"<p>Goodhart’s law states that when a measure becomes a target, it ceases to be a good measure.</p>\n"},"terminal goals":{"term":"terminal goals","pageid":"","contents":"<p>Goals which are valued as ends in themselves, rather than as instrumental to something else.</p>\n"},"terminal goal":{"term":"terminal goal","pageid":"","contents":"<p>Goals which are valued as ends in themselves, rather than as instrumental to something else.</p>\n"},"orthogonality thesis":{"term":"orthogonality thesis","pageid":"6568","contents":"<p>The thesis that any level of intelligence is compatible with any terminal goals. This implies that intelligence alone is not enough to make a system moral.</p>\n"},"instrumental convergence":{"term":"instrumental convergence","pageid":"897I","contents":"<p>Instrumental convergence is the idea that different AI agents, each with distinct terminal goals, will end up adopting many of the same instrumental goals.</p>\n"},"instrumentally convergent goals":{"term":"instrumentally convergent goals","pageid":"897I","contents":"<p>Instrumental convergence is the idea that different AI agents, each with distinct terminal goals, will end up adopting many of the same instrumental goals.</p>\n"},"llm":{"term":"llm","pageid":"","contents":"<p>A large language model is an AI  model which has been trained on a large body of text, in order to produce texts in a human-like way.</p>\n"},"large language model":{"term":"large language model","pageid":"","contents":"<p>A large language model is an AI  model which has been trained on a large body of text, in order to produce texts in a human-like way.</p>\n"},"goal misgeneralization":{"term":"goal misgeneralization","pageid":"","contents":"<p>pursuing a different goal during deployment from the one that was pursued during training due to distribution shift</p>\n"},"interpretability":{"term":"interpretability","pageid":"8241","contents":"<p>Interpretability is an area of alignment research that aims to make machine learning systems easier for humans to understand.</p>\n"},"existential risk":{"term":"existential risk","pageid":"89LL","contents":"<p>risks that threaten the destruction of humanity's long-term potential, including human extinction</p>\n"}}
+
+type GlossaryItem = {
+    term: string;
+    pageid: string;
+    contents: string;
+};
+
+// A component which wraps a paragraph and injects glossary terms into it as
+// hoverable pop-up links. The text is immediately rendered normally, but after
+// the glossary is loaded (which happens once per page, asynchronously), the
+// glossary terms are replaced with elements.
+export const GlossaryP: React.FC<{content: string}> = ({content}) => {
+    const [glossary, setGlossary] = useState<Map<string, GlossaryItem> | null>(null);
+    const [glossaryRegex, setGlossaryRegex] = useState<RegExp | null>(null);
+    
+    useEffect(() => {
+        if (glossary === null) {
+            const glossary = new Map(Object.entries(GLOSSARY_JSON));
+            setGlossary(glossary);
+            setGlossaryRegex(new RegExp(Array.from(glossary.keys()).join("|"), "gim"));
+        }
+    }, [glossary]);
+
+    // If the glossary hasn't loaded yet, just render the text normally.
+    if (glossary == null || glossaryRegex == null) {
+        return <span dangerouslySetInnerHTML={{__html: content}} />;
+    }
+
+    // Otherwise, replace glossary terms with links. We can do this in
+    // O(n * sum of term lengths) by finding String.prototype.indexOf of
+    // each term in the glossary (since that'd probably be backed by KMP)
+    // but I think it should be faster to compile a regex state machine
+    // once and use that instead.
+
+    return <span dangerouslySetInnerHTML={{__html: content.replace(glossaryRegex!, (match) => {
+        const item = glossary.get(match.toLowerCase());
+        if (item == undefined) return match;
+
+        const hover_content = item.contents;
+        const pageid = item.pageid;
+        return "AAAAAAAAA";
+
+    })}} />;
+}
+
+
+
+
diff --git a/web/src/pages/index.tsx b/web/src/pages/index.tsx
index 9a18f48..a268d9f 100644
--- a/web/src/pages/index.tsx
+++ b/web/src/pages/index.tsx
@@ -10,6 +10,7 @@ import Image from 'next/image';
 import Header from "../header";
 import { SearchBox, Followup } from "../searchbox";
 import logo from "../logo.svg"
+import { GlossaryP } from "~/glossary";
 
 type Citation = {
     title: string;
@@ -202,7 +203,7 @@ const ShowAssistantEntry: React.FC<{entry: AssistantEntry}> = ({entry}) => {
                 response.split("\n").map(paragraph => ( <p> {
                     paragraph.split(in_text_citation_regex).map((text, i) => {
                         if (i % 2 === 0) {
-                            return text.trim();
+                            return <GlossaryP content={text.trim()} />;
                         }
                         i = parseInt(text) - 1;
                         if (!citations.has(i)) return `[${text}]`;