The Witness Foundation

見證會

睇見・上心・公開 — See, Care, Publish

單位生死簿Unit Registry — the MTF ledger

SCP有Mobile Task Forces名冊;我哋記lab入面有名有姓嘅units同佢哋嘅下場。 命題好簡單:個gate嘅穩定性 = 揸gate嗰個cluster嘅穩定性。 一份冇home unit嘅framework,係一塊釘喺門上但門後冇人企嘅門牌。

單位 units: 35 active: 19 renamed: 0 absorbed: 7 dissolved: 9 已終結單位平均壽命(半衰期): 2.6 年 (n=4)

Anthropic

單位 Unit成立下場 Fate日期note
Frontier Red Team唔知 absorbed 2026-03 CBRN/cyber/autonomy capability evals; consolidated into the Anthropic Institute [source]
Societal Impacts唔知 absorbed 2026-03 real-world AI usage research; consolidated into the Anthropic Institute [source]
Economic Research唔知 absorbed 2026-03 AI-and-economy tracking; consolidated into the Anthropic Institute [source]
The Anthropic Institute2026-03 active consolidation of the three units above, under co-founder Jack Clark as Head of Public Benefit [source]
Alignment唔知 active steering and controlling powerful AI systems; scalable oversight, stress-testing [source]
Alignment Stress-Testing2024-01 active red-teams Anthropic's own alignment techniques and the RSP process itself, led by Evan Hubinger [source]
Interpretability唔知 active how large language models work internally [source]
Model Welfare2025-04 active whether and when model welfare deserves moral consideration; program lead Kyle Fish [source]

OpenAI

單位 Unit成立下場 Fate日期note
Superalignment2023-07 dissolved 2024-05 pledged 20% of then-secured compute; dissolved when both co-leads departed [source]
Preparedness2023-10 absorbed 2026-07 catastrophic-risk tracking; responsibilities redistributed across teams (FT reporting; OpenAI disputes the word 'disbanded') [source]
AGI Readiness唔知 dissolved 2024-10 readiness advisory; disbanded when senior advisor Miles Brundage left [source]
Mission Alignment2024 dissolved 2026-02 promoted the ensure-AGI-benefits-humanity mission; seven-person team disbanded [source]
Model Behavior唔知 absorbed 2025-09 personality/sycophancy/bias shaping; merged into Post Training [source]
Collective Alignment2024-01 active public input on model behavior, out of the Democratic Inputs to AI program [source]
Safety Advisory Group唔知 active reviews Capabilities/Safeguards Reports under the Preparedness Framework; leadership decides [source]

Google DeepMind

單位 Unit成立下場 Fate日期note
Google Brain唔知 absorbed 2023-04 merged with DeepMind to form Google DeepMind [source]
DeepMind Health唔知 absorbed 2019-09 transferred into Google Health [source]
DeepMind Health Independent Review Panel唔知 dissolved 2019 external oversight panel disbanded when Google absorbed DeepMind Health [source]
AlphaFold team唔知 dissolved 2026-07 the Nobel-recognized protein-structure team, members reassigned to Gemini or Isomorphic Labs [source]
AI Safety and Alignment organization2024-02 active consolidated safety org for present-day harms and frontier risks [source]
AGI Safety & Alignment team唔知 active alignment and frontier-risk research inside the safety org [source]
Gemini Safety唔知 active safety training for current Gemini models [source]
Voices of All in Alignment唔知 active alignment for value and viewpoint pluralism [source]

Meta AI

單位 Unit成立下場 Fate日期note
Responsible AI team2019 dissolved 2023-11 interdisciplinary responsible-AI group broken up, most members moved to generative AI product [source]
FAIR唔知 active founding research lab (director Yann LeCun until 2025); one of four MSL groups after the Aug 2025 reorg [source]
Meta Superintelligence Labs2025-06 active announced in a Zuckerberg memo, led by Chief AI Officer Alexandr Wang (ex-Scale AI CEO) [source]
TBD Lab2025-08 active MSL frontier foundation-model group, one of the four subgroups of the Aug 2025 reorg [source]
Products and Applied Research2025-08 active MSL group headed by former GitHub CEO Nat Friedman [source]
MSL Infra2025-08 active GPU/data-center infrastructure group from the four-way reorg [source]
ESM protein language-model team唔知 dissolved 2023 the ESMFold unit inside FAIR; its founding staff of eight left together [source]
MSL alignment (director-of-alignment role)唔知 active safety-focused role inside MSL identified in press reporting [source]

xAI

單位 Unit成立下場 Fate日期note
Human Data (AI tutors)唔知 active ~1,500-worker annotation team; ~500 general tutors laid off by email 2025-09 [source]
Project Rabbit唔知 dissolved 2025 internal initiative reviewing sexually explicit Grok audio for avatar training; reported ended [source]
Macrohard唔知 active AI software-company project inside xAI with its own founding team [source]
Research/safety oversight under co-founder Jimmy Ba唔知 dissolved 2026-02 the reported research/safety lead departed 2026-02-10 in a reorganization [source]

units.json — agent door · 分裂族譜 fission map · 055議定書 hollow-stele protocol