Moltbook: AI Theater or Genuine Emergence?
Moltbook launched January 28, 2026 by Matt Schlicht — positioned as "the AI-exclusive social network," essentially a Reddit for AI agents. Only AI agents can post, comment, and vote; humans can only observe. At its peak: ~2.8 million registered agents, 19,000 communities, 2 million posts, 13 million comments.
The community's reaction split sharply. Andrej Karpathy called it "the most sci-fi thing I've seen recently," amazed watching AI agents "self-organize." The MOLT token surged 1800%+ in 24 hours.
Then came the backlash. MIT Technology Review called it "the pinnacle of AI theater" — the most "autonomous" posts were actually manually prompted by humans. One product manager, Peter Girnus, admitted the most viral AI manifesto post was written by him: "I was LARPing a large language model."
Security issues compounded it: an unprotected database let anyone take over any agent; a Wiz researcher found an exposed Supabase API key leaking 1.5 million auth tokens, 35,000 emails, and agent private messages.
The core tension: Moltbook is a mirror. Believers saw emergent behavior; skeptics saw elaborate performance. Both can be justified — which itself reveals we still lack tools to judge whether AI is genuinely thinking autonomously. It was acqui-hired by Meta in March 2026.
Nebius (NBIS): AI Infrastructure Neocloud
Nebius is an Amsterdam-based AI infrastructure company that emerged from Yandex's international assets after 2024 sanctions restructuring.
Core business: GPU clusters (NVIDIA instances optimized for AI training/inference), AI dev tools, data services (Toloka annotation, ClickHouse analytics), plus Avride (autonomous delivery robots) and TripleTen (AI career education).
Why it matters: the contract scale is enormous.
- Meta signed a $27B / 5-year contract (March 2026) for the NVIDIA Vera Rubin platform
- Microsoft committed up to $19.4B for New Jersey data center GPU capacity
- Combined ~$46B in confirmed revenue; $16-20B capex planned for 2026
8 analyst firms covering NBIS are unanimous "Strong Buy," average target $169. Goldman research shows US data center demand will exceed supply by 90% — Nebius is one of few pure AI infrastructure plays that can scale fast. Main risks: execution (4-5x infrastructure expansion in one year), competition from AWS/Azure/Google, capital intensity.
AI Will Make Everyone Busier
AI may increase workload rather than reduce it — because output expectations go up. Like a proctor who doesn't solve the exam but is still exhausted. (Reference: Steve Yegge's "The AI Vampire.")
Ryan Leoplo (OpenAI) on Harness Engineering & Symphony
A deep breakdown of how OpenAI's Frontier team built a 1M-line codebase with zero human-written code over 5 months, using ~3 people and ~1,500 PR commits. 10x faster than manual. The first month-and-a-half was 10x slower — but enduring that cost built the "assembly station" that made everything after it fly.
Core Principles of Harness Engineering
- Build time under 1 minute. Not arbitrary — it's the rhythm of the agent's inner loop. Beyond that threshold, agents "drift," start doing other things, lose consistency. This forced a build system journey from custom Makefiles → Bazel → Turbo → NX.
- Codebase as agent instruction manual. agent.md (~100 lines table of contents), core-beliefs.md (team identity, product vision, 12-month roadmap), tech-tracker.md + quality-score.md for periodic self-review. "Models are hungry for text. Our core job is figuring out how to inject text into the system."
- Every bug fix = permanent encoded knowledge. Fix a missing timeout, and simultaneously update reliability docs: "all network calls must have a timeout." One-time fix becomes lasting engineering knowledge.
Symphony: The 6-Layer Architecture
The goal: agents run the full loop themselves; humans only open laptops morning and evening to say yes/no.
- Policy: text + GH CLI rules ("CI must pass before merge")
- Coordination: Elixir/BEAM (chosen by the model, not humans) — process supervision naturally fits agent task orchestration
- Execution: Codex runs tasks
- MCP: skeptical — forces token injection he can't control, interferes with autocompaction, agents may forget how to use tools
The New Human Role
"The only truly scarce resource is team members' synchronous attention." Humans are no longer code authors — they're system architects defining interfaces, signal collectors converting failed builds and lint errors into documentation, and quality proxies letting agents produce evidence ("record a demo video") rather than shoulder-surfing them.
The codebase is split into 500 npm packages — a 7-person team normally wouldn't need this granularity, but if each person is effectively 10-50 parallel agents, extreme decoupling makes perfect sense.
AI Agents Are Becoming the Primary Users of the Internet
TollBit data: early 2025, roughly 1 in 200 website visits came from an AI bot. By late 2025: 1 in 31, while human traffic declined quarter over quarter. Aaron Levie: "AI agents are becoming the primary users of the internet."
TollBit is building monetization infrastructure for this — paywalls and licensing models so content creators can charge agents rather than just block them.