#!/usr/bin/env python3 """Daily AI Quiz Generator""" import json, os, re, datetime, random, glob from pathlib import Path import anthropic SESSION_FILE = Path("data/session.json") DATE_STR = datetime.date.today().strftime("%B %d, %Y") DATE_ISO = datetime.date.today().isoformat() WEEKDAY = datetime.date.today().strftime("%A") SUPABASE_URL = "https://cokajjwirhtnvsyubepc.supabase.co" SUPABASE_KEY = "eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJzdXBhYmFzZSIsInJlZiI6ImNva2Fqandpcmh0bnZzeXViZXBjIiwicm9sZSI6ImFub24iLCJpYXQiOjE3Nzk4NDAwNDUsImV4cCI6MjA5NTQxNjA0NX0._zvclbxuixP9dvMcW-uBdA0liOLzqgcVMOyIJ457DFs" CAT_COLORS = { "Data Quality":"#7DCFB6","Pre-training":"#8FA3D8","RLHF & Fine-tuning":"#5B72B8", "Evaluation":"#5B72B8","Alignment & Safety":"#7DCFB6","Model Behavior":"#F4A261", "Interpretability":"#8FA3D8","Infrastructure":"#8FA3D8","Prompting":"#7DCFB6", "RL & Agent Behavior":"#5B72B8","Recent Research":"#F4A261","General":"#A8B0C8" } def load_session(): if SESSION_FILE.exists(): with open(SESSION_FILE) as f: return json.load(f) return {"session_number":0,"cumulative_flagged":[],"weak_areas":[], "topics_covered_recently":[],"total_sessions":0,"concept_library":[], "portfolio_balance":{"blue_chip_sessions":0,"growth_sessions":0,"speculative_sessions":0}} def save_session(s): SESSION_FILE.parent.mkdir(exist_ok=True) with open(SESSION_FILE,"w") as f: json.dump(s,f,indent=2) client = anthropic.Anthropic(api_key=os.environ["ANTHROPIC_API_KEY"]) def call_claude(system,user,max_tokens=4000): msg = client.messages.create(model="claude-sonnet-4-5",max_tokens=max_tokens, system=system,messages=[{"role":"user","content":user}]) return msg.content[0].text.strip() def extract_json(raw): text = re.sub(r'```json\s*','',raw); text = re.sub(r'```\s*','',text).strip() try: r = json.loads(text) if isinstance(r,(list,dict)): return r except: pass s,e = text.find('['),text.rfind(']') if s>=0 and e>s: chunk = text[s:e+1] try: return json.loads(chunk) except: try: return json.loads(re.sub(r',\s*([}\]])',r'\1',chunk)) except: pass return None def shuffle_answers(questions): out = [] for q in questions: opts=q.get("opts",[]); cidx=q.get("correct",0) if not opts or cidx>=len(opts): out.append(q); continue correct_text=opts[cidx]; new_opts=opts[:]; random.shuffle(new_opts) q2=dict(q); q2["opts"]=new_opts; q2["correct"]=new_opts.index(correct_text); out.append(q2) return out def get_topic_requests(): sk=os.environ.get("SUPABASE_SERVICE_KEY","") if not sk: print(" No SUPABASE_SERVICE_KEY"); return [] try: import requests as rlib r=rlib.get(f"{SUPABASE_URL}/rest/v1/topic_requests?fulfilled=eq.false&select=id,request", headers={"apikey":sk,"Authorization":f"Bearer {sk}"},timeout=10) if r.status_code==200: data=r.json(); print(f" {len(data)} requests"); return data except Exception as e: print(f" Requests error:{e}") return [] def fulfill_topic_requests(ids): sk=os.environ.get("SUPABASE_SERVICE_KEY","") if not sk or not ids: return try: import requests as rlib for rid in ids: rlib.patch(f"{SUPABASE_URL}/rest/v1/topic_requests?id=eq.{rid}", json={"fulfilled":True}, headers={"apikey":sk,"Authorization":f"Bearer {sk}","Content-Type":"application/json","Prefer":"return=minimal"}, timeout=10) print(f" Fulfilled {len(ids)}") except Exception as e: print(f" Fulfill error:{e}") def generate_questions(session,topic_requests=None): print("Generating 10 quiz questions...") weak=", ".join(session.get("weak_areas",[])[:5]) or "none" flagged=", ".join(session.get("cumulative_flagged",[])[:8]) or "none" recent=", ".join(session.get("topics_covered_recently",[])[:6]) or "none" sn=session.get("session_number",0)+1 bal=session.get("portfolio_balance",{}) topic_str="\n".join([f"- {r['request']}" for r in (topic_requests or [])]) or "None today." system=f"""Generate daily quiz questions for a data quality manager working in AI model training. Smart but NOT an engineer. Today: Session {sn}, {DATE_STR} ({WEEKDAY}) Portfolio: blue_chip={bal.get('blue_chip_sessions',0)}, growth={bal.get('growth_sessions',0)}, speculative={bal.get('speculative_sessions',0)} User requests: {topic_str} Weak areas: {weak} | Flagged: {flagged} | Recent topics: {recent} Mix: 4 data quality, 2 RLHF, 2 evaluation/alignment, 1 prompting/RL, 1 recent research, 1 recent product launches (prioritize frontier model releases, new AI apps gaining traction - e.g. new Claude/GPT/Gemini releases, viral AI tools). For each question, add an "applied_to" field: a short phrase showing where this concept applies in real work (e.g. "Annotation pipelines", "RLHF training", "Eval design", "Prompt engineering"). Return ONLY valid JSON array of 10 objects, no markdown: [{{"section":"...","concept":"2-6 words","q":"question","opts":["A","B","C","D"],"correct":0,"explanation":"2-3 sentences","difficulty":"foundational|practitioner|advanced","isNews":false,"article_url":"ar5iv link or empty","applied_to":"short real-world context"}}] For article_url: if referencing an arxiv paper use https://ar5iv.org/abs/PAPER_ID format. Otherwise leave empty string.""" for attempt in range(3): raw=call_claude(system,f"Generate 10 questions for session {sn}. JSON only.") parsed=extract_json(raw) if parsed and isinstance(parsed,list) and len(parsed)>=8: valid=[q for q in parsed if q.get("q") and isinstance(q.get("opts"),list) and len(q.get("opts",[]))==4 and isinstance(q.get("correct"),int) and q.get("concept")] if len(valid)>=8: print(f" Got {len(valid)} questions attempt {attempt+1}") return shuffle_answers(valid[:10]) print(f" Attempt {attempt+1} failed") print(" Using fallback"); return shuffle_answers(get_fallback()) def get_fallback(): return [ {"section":"Data Quality","concept":"Inter-annotator agreement","q":"Your annotators agree 55% of the time on a preference task. What does this most directly signal?","opts":["You need more annotators","The task definition is unclear — the model will learn noise","One annotator is an outlier","The model is too good to rate"],"correct":1,"explanation":"Low inter-annotator agreement almost always means the task is not well-defined. If skilled humans cannot agree, the model learns noise instead of signal.","difficulty":"foundational","isNews":False,"article_url":""}, {"section":"RLHF & Fine-tuning","concept":"KL penalty","q":"What does the KL divergence penalty prevent during RLHF?","opts":["Reward model overfitting","Policy drifting too far from SFT model","Gradient explosion","Data memorization"],"correct":1,"explanation":"The KL penalty keeps the RL policy anchored to the supervised model. Without it, the model exploits reward model weaknesses.","difficulty":"practitioner","isNews":False,"article_url":""}, {"section":"Data Quality","concept":"Synthetic data tradeoffs","q":"The biggest risk when using synthetic training data is:","opts":["Always lower quality","Oversamples easy cases, misses hard edge cases","Causes plagiarism","Style too uniform"],"correct":1,"explanation":"Synthetic data reflects what the generating model handles well, systematically missing hard edge cases where robustness matters most.","difficulty":"practitioner","isNews":False,"article_url":""}, {"section":"Evaluation","concept":"LLM-as-judge verbosity bias","q":"Verbosity bias in LLM judges means:","opts":["Prefers shorter answers","Rates longer answers higher regardless of quality","Penalizes formatting","Favors newer models"],"correct":1,"explanation":"Verbosity bias causes LLM judges to prefer longer responses regardless of actual quality, corrupting your evaluation signal.","difficulty":"practitioner","isNews":False,"article_url":""}, {"section":"Prompting","concept":"Few-shot prompting","q":"Few-shot prompting is most valuable when:","opts":["Model too large for zero-shot","Task format needs precise constraint by example","You want more creativity","Factual accuracy required"],"correct":1,"explanation":"Few-shot prompting excels at constraining output format and style, showing the model exactly what good output looks like.","difficulty":"foundational","isNews":False,"article_url":""}, {"section":"RL & Agent Behavior","concept":"Process reward models","q":"Process reward models differ from outcome reward models in that they:","opts":["Use human preference pairs","Reward each reasoning step not just final answer","Verify proofs","Only work for code"],"correct":1,"explanation":"Process reward models provide step-level feedback, helping models learn good reasoning chains.","difficulty":"advanced","isNews":False,"article_url":""}, {"section":"Alignment & Safety","concept":"Constitutional AI","q":"Constitutional AI reduces annotation costs by:","opts":["Crowdsourcing","Using principles and AI self-critique to generate preferences","Averaging multiple models","Rule-based filters"],"correct":1,"explanation":"Constitutional AI uses explicit principles and AI self-critique to generate preference labels at scale.","difficulty":"practitioner","isNews":False,"article_url":""}, {"section":"Model Behavior","concept":"In-context learning","q":"What makes in-context learning remarkable from an ML perspective?","opts":["Proves format memorization","Model adapts via forward-pass with no weight updates","Only works in-distribution","Model uses retrieval"],"correct":1,"explanation":"In-context learning adapts to new tasks through the forward pass with no gradient updates.","difficulty":"foundational","isNews":False,"article_url":""}, {"section":"Data Quality","concept":"Deduplication","q":"Beyond saving compute, deduplication matters because:","opts":["Duplicates slow tokenization","Models upweight repeated content increasing memorization","Improves annotation consistency","Prevents contamination"],"correct":1,"explanation":"Deduplication prevents models from over-learning repeated content, which skews distributions.","difficulty":"practitioner","isNews":False,"article_url":""}, {"section":"Interpretability","concept":"Superposition","q":"Superposition in neural networks means:","opts":["Multiple heads attend same token","Neurons encode multiple features via sparsity","Models store duplicates","Attention processes multiple contexts"],"correct":1,"explanation":"Superposition lets networks store more features than dimensions by encoding sparse features per neuron.","difficulty":"advanced","isNews":False,"article_url":""} ] def generate_concept_explanations(session,questions): print("Updating concept library...") existing={c["concept"]:c for c in session.get("concept_library",[])} new_concepts=[] for q in questions: concept=q.get("concept",""); cat=q.get("section","") if concept and concept not in existing: new_concepts.append({"concept":concept,"category":cat,"article_url":q.get("article_url","")}) if new_concepts: print(f" Generating explanations for {len(new_concepts)} concepts...") for c in new_concepts: try: sys=f'Explain "{c["concept"]}" ({c["category"]}) to a data quality manager in AI training. Smart, not an engineer.\nWrite 3 paragraphs: (1) what it is and why it matters for data quality, (2) nuances/failure modes, (3) concrete analogy.\nWrap 1-2 key terms in tags. Warm tone. No bullets. Start directly.' raw=call_claude(sys,"Write the explanation.",max_tokens=600) paras=[p.strip() for p in raw.strip().split("\n\n") if p.strip()][:3] existing[c["concept"]]={"concept":c["concept"],"category":c["category"], "color":CAT_COLORS.get(c["category"],"#8FA3D8"), "explanation":"".join(f"

{p}

" for p in paras), "times_seen":0,"marked_learned":False,"added_date":DATE_ISO,"article_url":c.get("article_url",""),"is_new":True} print(f" ✓ {c['concept']}") except Exception as e: print(f" ✗ {c['concept']}: {e}") for q in questions: c=q.get("concept","") if c in existing: existing[c]["times_seen"]=existing[c].get("times_seen",0)+1 bad=["explanation will appear","check back","coming tomorrow","next session","coming soon"] result=[c for c in existing.values() if len(c.get("explanation","").lower())>=50 and not any(p in c.get("explanation","").lower() for p in bad)] print(f" Library: {len(result)} concepts") return result def update_session(session,questions): session["session_number"]=session.get("session_number",0)+1 session["last_updated"]=DATE_ISO session["total_sessions"]=session.get("total_sessions",0)+1 topics=list(set([q["section"] for q in questions])) session["topics_covered_recently"]=(topics+session.get("topics_covered_recently",[]))[:12] adv=sum(1 for q in questions if q.get("difficulty")=="advanced") found=sum(1 for q in questions if q.get("difficulty")=="foundational") bal=session.get("portfolio_balance",{"blue_chip_sessions":0,"growth_sessions":0,"speculative_sessions":0}) if adv>=3: bal["speculative_sessions"]=bal.get("speculative_sessions",0)+1 elif found>=4: bal["blue_chip_sessions"]=bal.get("blue_chip_sessions",0)+1 else: bal["growth_sessions"]=bal.get("growth_sessions",0)+1 session["portfolio_balance"]=bal return session def save_quiz_file(questions, session_num): """Save today's quiz to Supabase daily_quizzes table.""" sk = os.environ.get("SUPABASE_SERVICE_KEY","") if not sk: print(" No SUPABASE_SERVICE_KEY — skipping quiz save to Supabase") return try: import requests as rlib url = f"{SUPABASE_URL}/rest/v1/daily_quizzes" headers = { "apikey": sk, "Authorization": f"Bearer {sk}", "Content-Type": "application/json", "Prefer": "resolution=merge-duplicates" } payload = { "date": DATE_ISO, "session_number": session_num, "questions": questions } r = rlib.post(url, json=payload, headers=headers, timeout=15) if r.status_code in (200, 201): print(f" Quiz saved to Supabase for {DATE_ISO}") else: print(f" Quiz save error: {r.status_code} {r.text[:100]}") except Exception as e: print(f" Quiz save error: {e}") def write_html(path,content): with open(path,"w",encoding="utf-8") as f: f.write(content) def css_base(): return """@import url('https://fonts.googleapis.com/css2?family=Bebas+Neue&family=Unbounded:wght@700;900&family=DM+Sans:opsz,wght@9..40,400;9..40,500;9..40,600&display=swap'); :root{--bg:#F2EDE4;--white:#FFF;--ink:#2D2926;--muted:#A8B0C8;--coral:#5B72B8;--teal:#7DCFB6;--yolk:#F4A261;--plum:#8FA3D8;--sky:#5B72B8;--border:rgba(45,41,38,.10);--nav-h:56px} *,*::before,*::after{box-sizing:border-box;margin:0;padding:0} body{background:var(--bg);color:var(--ink);font-family:'DM Sans',sans-serif;min-height:100vh;-webkit-font-smoothing:antialiased}""" def build_index(questions, concept_library, session_num): qs = json.dumps(questions, ensure_ascii=False) return f""" Distilled · {DATE_STR}
Daily AI Quiz

Welcome back.

Sign in to access your quiz and track your progress.

No account? Sign up
DISTILLED
Session {session_num} · {DATE_STR}

Time to learn something.

10 questions on AI training & data quality, calibrated to your role and history.

Loading your progress...
Open concept library →
Session {session_num} · {DATE_STR} 0 / 0
Question 1 of 1010%
🌟 0%
0
Correct
0
Wrong
0
Flagged
—
—
—
Sessions
—
Avg score
—
Streak
Recent sessions
No sessions yet — take your first quiz!
Areas to work on
Complete a quiz to see your weak areas.
Leaderboard
Loading...
Get the app

Daily AI Quiz
on any device.

No app store needed. Add it to your home screen in seconds.

✓ Already installed — you're good to go!
iPhone & iPad
Must use Safari
1
Open this site in Safari
2
Tap Share ⬆ at the bottom of the screen
3
Tap "Add to Home Screen"
4
Tap Add — done!
Android
Chrome browser
1
Tap ⋮ menu in Chrome
2
Tap "Add to Home Screen"
3
Tap Add — done!
Desktop
Chrome or Edge
1
Look for ⊕ install icon in the address bar
2
Click it and select Install
3
App opens as a standalone window
""" def build_learn(concept_library, session_num): bad=["explanation will appear","check back","coming tomorrow","next session","coming soon"] concepts=[c for c in concept_library if len(c.get("explanation","").lower())>=50 and not any(p in c.get("explanation","").lower() for p in bad)] cj=json.dumps(concepts,ensure_ascii=False) total=len(concepts) if total==0: body="""
🌱
Building your library

Concepts appear here after each quiz session.

Take today's quiz
Start building your library
→
""" else: body=f"""
Progress0%
🎉
All done!

Repetition is how they stick.

Take today's quiz
Put these concepts to the test
→
Request a topic
Add to tomorrow's quiz
→
""" return f""" Concept Library · Daily AI Quiz
← Quiz 0 learned

Concept Library

Flip each card or scroll through all at once.

Updated {DATE_STR} · {total} concepts · Session {session_num} {body}
""" def build_topics(session_num): return f""" Request a Topic · Daily AI Quiz
← Quiz

Request a Topic

Tell us what you want to learn next. Your request will be included in tomorrow morning's quiz.

Ideas to get you started
Process reward models Prompt injection attacks RLVR and verifiable rewards Data flywheel strategy Annotation calibration Mixture of experts
What do you want to learn?

Be as specific or broad as you like — a concept, a question, or an area you want to go deeper on.

Takes effect tomorrow morning
✓ Request saved! It'll be included in tomorrow's quiz.
Please sign in to the quiz app first to save topic requests.
Back to quiz
Continue learning
→
""" def build_archive(session_num): return f""" Past Quizzes · Daily AI Quiz
← Quiz
🔒
Sign in to access

Past quizzes are available to signed-in members only.

Go to sign in →

Past Quizzes

Browse and take any previous session. Great for catching up.

Loading sessions...
—
Archive session
0 / 0
Question 1 of 1010%
🌟 0%
0
Correct
0
Wrong
0
Flagged
Today's quiz →
""" def build_about(): return f""" About · Distilled
← Quiz
Welcome to

DISTILLED

A daily knowledge training tool for people working in and around AI. Built to close the gap between reading about AI and actually understanding it.

The idea
Learning by not being able to guess

Most AI content is written to be understood — which means you can follow along without actually learning. Distilled works differently. Each question is designed so you cannot deduce the answer from surface-level reading. You have to know the concept.

The quiz tracks where you go wrong, flags concepts you struggle with, and adjusts future sessions to revisit your gaps. Over time, it builds a personalized curriculum based on your actual knowledge — not what you think you know.

Every session also adds to your concept library: a growing reference of explained ideas you can flip through, organized by topic, with real-world context for each one.

How to use it
Getting the most out of each session
1
Take the daily quiz — 10 questions every morning. Select your answer, then submit before seeing if you're right. Sit with your choice.
2
Flag what confuses you — anything that trips you up, mark it. Flagged concepts get revisited in future sessions.
3
Read the explanations — after submitting, every question shows a 2-3 sentence explanation written for practitioners, not engineers.
4
Browse the library — your concept library grows with every session. Flip through flashcards or scroll through all by category.
5
Request topics — if there's something specific you want to cover, add it. It'll appear in tomorrow's quiz.
6
Catch up with the archive — missed a session? Browse all past quizzes and take them anytime. Your scores are tracked separately.
What we cover
10 content areas, updated daily
Data Quality RLHF & Fine-tuning Evaluation Alignment & Safety Model Behavior Interpretability Infrastructure Prompting RL & Agent Behavior Recent Research Product Launches

Each session mixes foundational concepts with practitioner-level questions and frontier research. The balance shifts based on your history and portfolio — think of it as a learning portfolio: 40% durable fundamentals, 40% gaining-importance ideas, 20% frontier speculation.

Install the app
Add Distilled to your home screen
✓ Already installed — you're good to go!
iPhone & iPad
Must use Safari
1
Open this site in Safari
2
Tap Share ⬆ at the bottom
3
Tap "Add to Home Screen"
4
Tap Add
Android
Chrome browser
1
Tap ⋮ menu in Chrome
2
Tap "Add to Home Screen"
3
Tap Add
Desktop
Chrome or Edge
1
Look for ⊕ install icon in address bar
2
Click it and select Install
DIS
TILLED
Download logo
SVG · For home screen & sharing
↓
""" def main(): print(f"=== Daily AI Quiz · {DATE_STR} ===") session=load_session() print(f"Session #{session.get('session_number',0)+1}") topic_requests=get_topic_requests() request_ids=[r['id'] for r in topic_requests] questions=generate_questions(session,topic_requests) print(f"Generated {len(questions)} questions") if request_ids: fulfill_topic_requests(request_ids) concept_library=generate_concept_explanations(session,questions) session=update_session(session,questions) session["concept_library"]=concept_library save_session(session) print("Session saved") sn=session["session_number"] write_html("index.html",build_index(questions,concept_library,sn)) print("index.html written") write_html("learn.html",build_learn(concept_library,sn)) print("learn.html written") write_html("topics.html",build_topics(sn)) print("topics.html written") write_html("archive.html",build_archive(sn)) print("archive.html written") write_html("about.html",build_about()) print("about.html written") save_quiz_file(questions, sn) print("=== Done ===") if __name__=="__main__": main()