Every morning at 8am PST — the 5 ML papers and posts that actually matter for AGI, written like a brilliant friend explaining them over coffee.
An autonomous agent that scrapes the entire ML frontier, computes deltas vs yesterday, picks the 5 most AGI-relevant stories, has Claude write deep explanations, and sends a beautiful dark-mode HTML email. Every day. No duplicates. No noise.
8:00 AM PST daily:
Scrape 6 sources Compute deltas Claude selects top 5
┌─────────────┐ ┌──────────────┐ ┌──────────────────┐
│ arXiv (4) │──┐ │ │ │ │
│ HuggingFace │ │ ~200 │ SQLite DB │ new │ Claude ranks │
│ Anthropic │──┼─ items ──>│ seen before?│──only──>│ by AGI signal │
│ OpenAI │ │ /day │ deduplicate │ ~50 │ picks 5 best │
│ DeepMind │ │ │ │ │ │
│ Meta AI │──┘ └──────────────┘ └────────┬─────────┘
│ Stanford │ │
│ MIT CSAIL │ v
│ Berkeley │ ┌──────────────────┐
└─────────────┘ │ Claude writes 5 │
│ deep stories │
│ Story/Logic/ │
│ Maths/Impact/AGI │
└────────┬─────────┘
│
v
┌──────────────────┐
│ Beautiful HTML │
│ email via Gmail │
│ SMTP │
└──────────────────┘
Every URL ever seen is stored in SQLite. Run it 100 days in a row — you'll never see the same paper twice. The agent only writes about what's genuinely new since yesterday.
| Source | What | Count |
|---|---|---|
| arXiv cs.LG | Machine Learning | ~30/day |
| arXiv cs.AI | Artificial Intelligence | ~30/day |
| arXiv cs.CL | Computation & Language | ~30/day |
| arXiv stat.ML | Statistics ML | ~30/day |
| HuggingFace Papers | Community-upvoted papers | ~20/day |
| Anthropic Research | Frontier lab blog | ~5 |
| OpenAI Research | Frontier lab blog | ~5 |
| DeepMind Publications | Frontier lab blog | ~10 |
| Meta AI Research | Frontier lab blog | ~10 |
| Stanford HAI | University AI news | ~10 |
| MIT CSAIL | University AI news | ~10 |
| Berkeley BAIR | University AI blog | ~5 |
Each of the 5 stories follows this structure:
- Story — Hook with a vivid analogy or surprising fact
- The Logic — What problem does this solve? How does it work?
- The Maths — Key equation or mechanism, simplified but accurate
- Impact — What changes because this exists? Be specific.
- Why this gets us closer to AGI — What capability gap does this close?
Content is scored by keyword signals:
- High (+10): self-improvement, reasoning, agents, planning, world model, scaling law, alignment, chain of thought, test-time compute, agentic
- Medium (+5): transformer, LLM, fine-tuning, evaluation, safety, interpretability, MoE, distillation
- Low (-2): image generation, video, protein, drug, climate (penalized — not AGI-relevant)
Claude then selects the final 5 from the top 20 scored candidates.
# 1. Clone
git clone https://github.com/originaonxi/agi-possible-agent.git
cd agi-possible-agent
# 2. Install
pip3 install -r requirements.txt
# 3. Set environment variables
export ANTHROPIC_API_KEY="your_key"
export SMTP_USER="your_gmail@gmail.com"
export SMTP_PASS="your_gmail_app_password"
export EMAIL_TO="recipient@email.com"
# 4. Run
python3 agent.py
# 5. Schedule daily at 8am PST
bash scheduler.sh~$0.10/day in Claude API calls (7 calls to claude-sonnet-4-20250514: 1 selector + 5 stories + 1 closing).
Anmol Chaudhary — CTO @ Aonxi, Ex-Meta, Ex-Apple