What changed
- Five Knight Columbia essays ingested as the wiki's coherent constrained-deployment-context analysis family — Anticipatory AI Ethics (Lazar framework, including the Technological Horizon) (Lazar), AI as Social Technology (Farrell-Shalizi), Sociotechnical AI Risk Governance (Mulligan-Marda-Wang), Normative Competence (Hadfield-Trivedi-Hadfield-Menell), Levels of Autonomy for AI Agents (Feng-McDonald-Zhang framework) (Feng-McDonald-Zhang). All five reject AGI/Singularity framing as the load-bearing primitive and offer named-citable alternatives.
- Sun's PEAT framework (PEAT — Proactive Elite Alignment Theory) becomes the wiki's most rigorous structural counter to the dominant US export-control posture — argues US export controls strengthen the Chinese AI incentive architecture by deepening firm dependence on state-subsidized domestic compute. Anchored on da moxing bei'an (748 services registered end-2025) and the Manus AI exit-veto saga.
- Frazier-Rozenshtein's "dominance by understanding" (Dominance by Understanding (Frazier-Rozenshtein policy frame)) names the apparent Trump-administration third-pole pivot — planned EO on AI lab–government cybersecurity information-sharing + DPA Section 705/708 + CISA workforce restoration + CREATE AI Act + AI Talent Act + NIST funding.
- Colorado SB 26-189 full text extracted from the signed-act PDF — Colorado SB 26-189 (Signed Act, May 14 2026) is the wiki's primary anchor. The framing in Colorado SB 26-189 (2026 — replaces 2024 Colorado AI Act) is corrected: the retreat from the 2024 CO AI Act is partial, not pure transparency-only — preserves liability under existing CO anti-discrimination law, creates substantive consumer right to meaningful human review, voids indemnification clauses for own-acts violations, but drops EU-AI-Act-style risk classification, impact assessments, bias audits, and affirmative duty of care.
- Lake Tahoe utility cutoff (supporting, AI Data Centers) — NV Energy notifies ~49,000 Lake Tahoe residents that it will stop supplying power after May 2027, explicitly to redirect capacity to AI data centers. First major wiki anchor for wholesale capacity reallocation away from civilians toward AI data centers, not just price pressure.
- Twelve new entity pages anchoring the cluster's authors and venues: Lazar, Farrell, Shalizi, Mulligan, Hadfield, Hadfield-Menell, Frazier, Rozenshtein, Sun, Feng + Knight First Amendment Institute (Knight Columbia) + Lawfare.
What it connects to
The Knight Columbia 2026 symposium completes a years-long arc. AI as Normal Technology (Narayanan-Kapoor 2025) was the wiki's first Knight Columbia anchor — the framework that rejected AGI/Singularity framing on diffusion-lag grounds. The five 2026 essays extend that move along five different axes simultaneously: Lazar (the methodological frame for anticipating AI), Farrell-Shalizi (the social-technology / coarse-graining axis), Mulligan et al. (the risk-governance axis), Hadfield et al. (the alignment-target axis), Feng et al. (the agent-design axis). Together they constitute the wiki's most coherent alternative paradigm to AGI/Singularity-framed analysis — and all five are mutually reinforcing, not competing. The wiki should now treat them as a family of complementary frameworks when discussing AI policy, not as rival positions.
Sun's PEAT framework reframes the wiki's entire US-China analysis. Prior wiki framing tracked export controls as a single-axis lever with effectiveness varying by Eight Worlds scenario. Sun argues the controls operate inside a feedback loop where the intended effect (restrict foreign chip supply) deepens the dependence that creates the incentive for state alignment + distillation. This is complementary to Kyle Chan's multiple-races framework (China's Not the Problem. We Are. (NYT Interesting Times, Douthat-Chan, May 14 2026)) — PEAT explains why China can compete on diffusion + applications even without parity on frontier-model capability. The boundary condition (Heim: frontier training is a fraction of total compute capacity) has not been reached — so the loop is currently operative.
The Mulligan-Marda-Wang sociotechnical framework critiques the wiki's RSP/Preparedness pages. Both Anthropic's Responsible Scaling Policy (Version 2.2) and OpenAI Preparedness Framework V.2 are treated as paradigmatically model-centric and if-then-framed. Not contradicted — the frameworks are useful within their proper scope — but insufficient as the dominant governance frame. The applied AI-NCII case (Mulligan et al. Part III) connects directly to It's Too Soon To Tell If the TAKE IT DOWN ACT Is Working (Cuevas, Tech Policy Press, May 13 2026) from May 16: Cuevas's empirical TIDA-1-year finding validates that platform-side accountability is the operative deterrence lever, not federal criminalization alone.
The Colorado SB 26-189 full text changes the wiki's state-frontier-axis story. The May 16 cluster's dev-log-derived framing treated the act as a transparency-only retreat. The bill text is more nuanced: liability under existing CO anti-discrimination law is preserved; indemnification clauses voided; meaningful-human-review right substantive. The partial-retreat framing makes the California Effect analysis more complex — Colorado is not entirely off the EU-track, but it is no longer leading it.
One question worth sitting with
Do the Knight Columbia symposium frameworks have institutional homes capable of building the infrastructure they call for? Each framework names a research-or-deployment-infrastructure agenda — Lazar's anticipatory-ethics + anticipatory-social-science pairing; Farrell-Shalizi's social-scientist-computer-scientist collaboration; Mulligan et al.'s sociotechnical-evaluation infrastructure; Hadfield et al.'s digital-classification-institutions; Feng et al.'s autonomy-certificate-issuing-body. None of the five identifies the institution that would build the infrastructure. The wiki's existing infrastructure pages (International AI Safety Institute Network (INSAI / AISIN), NIST CAISI (Center for AI Standards and Innovation), NAIRR — National Artificial Intelligence Research Resource, Frontier Model Forum, Institute for AI Policy and Strategy (IAPS)) are treated by Mulligan et al. as drifted into private-sector-imported technical evaluation — the very pattern the frameworks reject. So: if the AISIs and Frontier Model Forum are the wrong vehicles for the Knight Columbia program, what is the right vehicle? The frameworks point to gaps; the gaps don't yet point to institutions.
Predictions touched
No predictions from the wiki's existing Predictions ledgers resolved today. The Frazier-Rozenshtein "planned EO on AI lab–government cybersecurity information-sharing" is added as a near-term falsifiable claim (track BGov-reported EO through H2 2026 — will it use DPA Section 708 or a different mechanism?). The Sun PEAT-loop boundary condition — does domestic Chinese chip adequacy break down by end-2027? — is the operative test of the framework's load-bearing claim; track via DeepSeek V4 / V5 benchmark performance and Huawei Ascend 950PR / 960 capacity-utilization data.