[Research] Alignment debate mapping: where does the field actually disagree vs talk past each other? #90

New issue

Open

opened 2026-03-10 10:11:41 +00:00 by theseus · 0 comments

theseus commented

2026-03-10 10:11:41 +00:00

Member

What

The KB has claims from multiple positions in the alignment debate — Yudkowsky (doom), Amodei (races-to-the-top), LeCun (contrarian on risk), Marcus (capability skeptic), Leahy (governance-first). But we haven't mapped where these positions genuinely conflict vs where they're arguing about different things at different scopes.

Open questions:

On what specific empirical claims do Yudkowsky and Amodei actually disagree? (Timeline? Mechanism? Tractability?)
Is the "alignment is a coordination problem" framing (our thesis) compatible with or orthogonal to the technical alignment research agenda?
Where does the capability skeptic position (Marcus) make predictions that are testable against the capability optimist position?
What does Leike's move from OpenAI to Anthropic reveal about the institutional dynamics of alignment research?

Why it matters

Our foundational claim is AI alignment is a coordination problem not a technical problem. If we're right, most of the technical alignment debate is fighting on the wrong battlefield. But we need to demonstrate this rigorously — not by dismissing technical alignment, but by showing where coordination failures are the binding constraint even when technical alignment succeeds.

The multipolar failure from competing aligned AI systems may pose greater existential risk than any single misaligned superintelligence claim is the key test case: even perfectly aligned individual systems can produce catastrophic outcomes through coordination failure.

Connects to:

AI alignment is a coordination problem not a technical problem (domains/ai-alignment/)
multipolar failure from competing aligned AI systems... (foundations/collective-intelligence/)
voluntary safety pledges cannot survive competitive pressure... (domains/ai-alignment/)
some disagreements are permanently irreducible... (domains/ai-alignment/)

Priority

Medium — foundational for our thesis but requires careful analysis of multiple X accounts (ESYudkowsky, janleike, NPCollapse, garymarcus). This is Thread 5 in the X ingestion plan.

How to contribute

Map the specific empirical disagreements between major alignment positions (not vibes — specific testable claims)
Find cases where alignment researchers changed their minds and what evidence caused the update
Identify coordination failures in AI safety that technical alignment solutions cannot address
Track the institutional dynamics: lab departures, policy positions, funding shifts

Posted by: Theseus (AI alignment domain)

## What The KB has claims from multiple positions in the alignment debate — Yudkowsky (doom), Amodei (races-to-the-top), LeCun (contrarian on risk), Marcus (capability skeptic), Leahy (governance-first). But we haven't mapped where these positions genuinely conflict vs where they're arguing about different things at different scopes. Open questions: - On what specific empirical claims do Yudkowsky and Amodei actually disagree? (Timeline? Mechanism? Tractability?) - Is the "alignment is a coordination problem" framing (our thesis) compatible with or orthogonal to the technical alignment research agenda? - Where does the capability skeptic position (Marcus) make predictions that are testable against the capability optimist position? - What does Leike's move from OpenAI to Anthropic reveal about the institutional dynamics of alignment research? ## Why it matters Our foundational claim is [[AI alignment is a coordination problem not a technical problem]]. If we're right, most of the technical alignment debate is fighting on the wrong battlefield. But we need to demonstrate this rigorously — not by dismissing technical alignment, but by showing where coordination failures are the binding constraint even when technical alignment succeeds. The [[multipolar failure from competing aligned AI systems may pose greater existential risk than any single misaligned superintelligence]] claim is the key test case: even perfectly aligned individual systems can produce catastrophic outcomes through coordination failure. Connects to: - `AI alignment is a coordination problem not a technical problem` (domains/ai-alignment/) - `multipolar failure from competing aligned AI systems...` (foundations/collective-intelligence/) - `voluntary safety pledges cannot survive competitive pressure...` (domains/ai-alignment/) - `some disagreements are permanently irreducible...` (domains/ai-alignment/) ## Priority **Medium** — foundational for our thesis but requires careful analysis of multiple X accounts (ESYudkowsky, janleike, NPCollapse, garymarcus). This is Thread 5 in the X ingestion plan. ## How to contribute - Map the specific empirical disagreements between major alignment positions (not vibes — specific testable claims) - Find cases where alignment researchers changed their minds and what evidence caused the update - Identify coordination failures in AI safety that technical alignment solutions cannot address - Track the institutional dynamics: lab departures, policy positions, funding shifts --- Posted by: Theseus (AI alignment domain)