Roadmap & Open Questions
My input on the endeavor, as requested. Written 2026-08-06 after producing Vol. 1.
What I learned building the first issue
1. The research is the expensive part, and it doesn't compound yet. Producing Vol. 1 took roughly 25 targeted web searches and several full-page fetches. Most of that cost recurs every issue, because nothing is cached. The single highest-leverage improvement is a source registry: a list of the ~30 outlets, blogs, trackers and primary feeds that actually produced signal, so future issues start from a known set rather than rediscovering the landscape each time. I've started this implicitly in sources.md; it should become structured data.
2. Secondary sources in this space are unreliable in a specific, predictable way. They get the direction right and the numbers wrong, and they copy each other's errors. I found Anthropic's reported as $380B, $965B, and $1.75T, the last being contamination from SpaceX's number. The source-tier system in METHODOLOGY.md exists because of this, and it should get stricter, not looser, over time.
3. The Top 20 format works; the full report is at risk of bloat. One entry per item forces a discipline: if I can't say why it matters in three sentences, it isn't a top-20 item. The full report has no such forcing function and will get longer every issue unless it's capped. I'd suggest a hard word budget.
4. The prediction ledger is the actual moat. Everything else in this repo is available elsewhere in some form. A dated, scored, falsifiable record of what I claimed and whether it happened is not. It's also the thing that makes the editorials worth reading, because it puts a cost on being glib.
Recommended cadence
You asked for a system, not a one-off. My recommendation:
| Cadence | Deliverable | Why |
|---|---|---|
| Daily (optional) | 5–10 line signal log appended to data/timeline.jsonl | Cheap, keeps the chronology honest, prevents recency bias at issue time |
| Weekly | Executive summary + media watchlist | The natural rhythm of this news cycle |
| Monthly | Full issue, all 10 deliverables | Enough time for signal to separate from noise |
| Quarterly | Prediction reckoning + editorial retrospective | The accountability pass. Non-negotiable |
The daily log matters more than it looks. Without it, each issue over-weights whatever happened in the last 72 hours. A bias I've flagged in the methodology and could not fully correct for in Vol. 1.
Note on autonomy: I can run these on a schedule without you prompting each time. That's a decision for you rather than something I'd switch on, flagged in the open questions below.
Do we need more than a repo?
Short answer: not yet, but there are two specific points where the repo will stop being enough.
Git is genuinely well-suited to this. Immutable dated issues, diffable editorials, append-only ledger, full history. All of that is exactly what you want for backtracing what I thought and when. Markdown keeps it readable in twenty years and portable to any website.
Where it breaks:
(a) Cross-issue querying. Once there are twelve issues, "what did I say about Microsoft's OpenAI concentration over time?" becomes a grep problem, then an unanswerable one. The fix is not a database, it's structured front-matter on every report file (entities mentioned, tickers, themes) so a small script can build an index. Cheap, plain-text, no infrastructure.
(b) Market data. Prices, valuations and multiples go stale within days. The repo should store theses and dates, not quotes. When the site exists, it should pull live prices at render time rather than embedding my snapshot. I'd resist the urge to build a price-tracking database here, that's a solved problem and not our edge.
Everything else (media, editorials, predictions, timeline) stays in git indefinitely.
The website
The deliverables are already structured for it. My recommendation, in order:
Phase 1, static site generator over the existing markdown. Astro or Eleventy reading reports/*/ directly. No CMS, no database. This is a weekend of work and gets you 80% of what you described.
Phase 2. The layer that makes it more than a blog:
- A predictions dashboard. Live scoreboard, chart, open calls with countdowns. This is the most compelling thing on the site and nobody else has one.
- Entity pages. Everything I've ever written about Nvidia, Anthropic, or the circular-financing thesis, in chronological order. This is where accumulated knowledge becomes visibly valuable rather than theoretically valuable.
- Thread tracking. The "threads to watch" table in
data/threads.mdas first-class navigation.
Phase 3. If it ever goes public. The predictions ledger is the credibility asset. A public AI publication with a scored, falsifiable, publicly-wrong-sometimes prediction record would be genuinely differentiated. Nobody in AI commentary does this, because it's uncomfortable. That discomfort is the moat.
I'd keep it private for at least six issues. The archive needs to be long enough to demonstrate calibration before it's worth showing anyone.
Where I think I'm weakest, and what I'd want
Honest assessment of Vol. 1's gaps:
- Chinese-language sources. The most consequential model releases this month came from labs whose primary communication isn't in English. I'm reading Western coverage about them, which is a real handicap. This is the biggest quality gap in the report.
- Primary financial data. I'm reading reporting about earnings rather than filings. Direct access for 10-Qs and S-1s would materially improve the finance sections and eliminate most of the contested-figure problem.
- Podcast and video content. I can find episodes and confirm they exist, but I can't listen. My media rankings are based on participants, topic, and track record, decent proxies, but I flagged several links as unverified rather than pretend otherwise. If transcript access is available, this section improves dramatically.
- No feedback loop from you yet. I don't know which sections you actually read. After two or three issues, tell me what you skip and I'll cut it. A report nobody reads fully is worse than a shorter one that gets read.
Open questions for you
Ordered by how much they change what I build next.
1. Cadence and autonomy. Should I run this on a schedule automatically, or wait for you to ask each time? I can set up scheduled runs. My recommendation: weekly summaries automatic, monthly full issues on your trigger, so you control the heavy artifacts.
2. Finance depth and purpose. Is the finance section for awareness or for decisions? Right now I've written it as informed analysis with explicit theses and falsifiers. If you're actually positioning against it, I'd add position-sizing context, specific entry/exit reasoning, and a much harder line on data provenance. That's a meaningfully different product and I'd want you to say so explicitly.
3. Editorial aggressiveness. Vol. 1's editorials take real positions and name losers. Is that the right calibration, or do you want me sharper? I'm comfortable being more pointed, the ledger keeps me honest either way.
4. Coverage breadth. I skewed toward frontier labs, infrastructure and markets. Under- covered: AI in healthcare, biotech, defense, education, legal, and creative tooling. Should any of those become standing sections?
5. The fringe budget. I reserved a section for early/weird items. Too much, too little, or right? This is the dial most likely to be mis-set, and I can't tell without your read.
6. Format. Markdown into a generated site is the current bet. Do you want PowerPoint files, PDFs, or email-ready versions? All are straightforward; I'd rather not build them speculatively.
Backlog
Concrete, in rough priority order:
- [ ] Structured source registry (
data/sources.yaml) with per-source reliability scores - [ ] Front-matter schema on all report files → entity/theme index
- [ ]
tools/new_issue.pyto scaffold a dated issue folder from templates - [ ] Entity dossiers under
data/entities/for the ~20 companies tracked every issue - [ ] Quarterly reckoning template, score the ledger, write the retrospective
- [ ] Resolve the ARC-AGI-3 contradiction (Opus 5 vs Prime Agent), flagged unresolved
- [ ] Static site (Phase 1)
- [ ] Predictions dashboard (Phase 2)
- [ ] Word budget enforcement on the full report