Every thesis essay I publish stakes a position on how the world will go. This page records those positions in falsifiable form — specific outcomes, dated resolution windows, and what would prove me wrong.
The point isn’t to be right. It’s to be calibrated: if I’m only right 60% of the time, I want that to be visible. Predictions written down can’t be retroactively edited into wins. Predictions remembered can.
I update the Status column as evidence comes in. Resolved predictions keep their original text — corrections happen in the Notes column, not by rewriting history.
Status overview
| ID | Essay | Resolve by | Status |
|---|---|---|---|
| P1 | Agent Topology Follows Task, Not Template | 2029-Q1 | Open |
| P2 | The $100B Monologue | 2029-Q1 | Open |
| P3 | Why AI Has to Create Jobs | 2029-12-31 | Open |
| P4 | The OpenClaw Playbook | 2027-12-31 | Open |
Status legend: Open (waiting), Right (resolved correctly), Wrong (resolved against), Partial (mixed), Voided (premise invalidated).
P1 — Agent topology will go dynamic
Source essay: Agent Topology Follows Task, Not Template (April 2026) Written: 2026-04-19 Resolve by: 2029-Q1 Status: Open
Prediction. By end of 2028, at least one major agent framework — LangGraph, AutoGen, or OpenAI’s Agents SDK — will ship query-adaptive / dynamic topology as the recommended default execution mode, with fixed Planner → Executor → Critic relegated to a legacy import path.
Wrong if. All three of LangGraph, AutoGen, and OpenAI’s Agents SDK still recommend static hierarchical topologies as the default execution mode by Q1 2029.
How I’ll check. Read each framework’s official “getting started” tutorial in Q1 2029. Default = whatever appears in the first runnable code example. Dynamic counts if the topology is computed by the framework or an LLM at runtime, not declared by the user.
Notes.
- (none yet)
P2 — Latent reasoning will displace the visible monologue
Source essay: The $100B Monologue (February 2026) Written: 2026-02 Resolve by: 2029-Q1 Status: Open
Prediction. By end of 2028, at least one frontier lab — OpenAI, Anthropic, or Google DeepMind — will ship a production reasoning mode that matches or beats its visible-CoT counterpart on hard reasoning benchmarks (AIME, GPQA-Diamond, FrontierMath) while emitting under 30% the thinking tokens, or replaces token-level deliberation with latent-state reasoning entirely.
Wrong if. All three frontier labs continue to bill users for thousands of monologue tokens per hard query through Q1 2029, with no production mode at the <30%-tokens or latent-reasoning bar.
How I’ll check. Compare published model cards / pricing pages in Q1 2029. Token counts visible to API users. Benchmark scores from third-party leaderboards (not vendor-reported only).
Notes.
- (none yet)
P3 — AI either creates entry-level jobs or forces policy
Source essay: Why AI Has to Create Jobs—or Fail Spectacularly (September 2025) Written: 2025-09-12 Resolve by: 2029-12-31 Status: Open
Prediction. By end of 2029, US labor force participation for 22–26 year-olds will either recover above the 2024 baseline (suggesting AI created replacement entry paths) or trigger explicit federal labor-policy intervention — retraining mandate, automation tax, or expanded UBI pilot. The silent default — AI hollowing out entry-level work with no policy response — will not survive four more years.
Wrong if. End of 2029 shows sub-2024 youth (22–26) labor force participation AND no federal AI-labor policy enacted (no automation tax, no federal retraining mandate, no expanded UBI pilot beyond pre-2025 scope).
How I’ll check. BLS labor force participation series (22–26 age cohort). Cross-reference enacted federal legislation by 2029-12-31.
Notes.
- (none yet)
P4 — Personal AI gets won by indie / open source, not VC
Source essay: The OpenClaw Playbook (February 2026) Written: 2026-02 Resolve by: 2027-12-31 Status: Open
Prediction. By end of 2027, the most-used personal AI tool in the US (by DAU) will be either open-source community-led (OpenClaw, Ollama-class) or a single-developer indie effort — not a VC-backed personal-AI startup (Pi, Rabbit-class, Humane-class).
Wrong if. A venture-funded personal-AI company holds the #1 DAU spot in the US at end of 2027.
How I’ll check. Sensor Tower / SimilarWeb DAU rankings for personal AI assistant category at end of Q4 2027. “Personal AI” defined as: assistant tools running on consumer device or chat platform, not generic AI chatbots from foundation-model providers (ChatGPT, Claude.ai, Gemini app are out of scope — they’re foundation-model UX, not personal-AI products).
Notes.
- (none yet)