О чём мораторий реально был

В марте 2023 Future of Life Institute опубликовал open letter подписанный тысячами researchers, engineers, executives (Musk, Bengio, Wozniak) призывающий к шестимесячному pause AI training runs "more powerful than GPT-4". Letter утверждал AI labs locked в out-of-control race develop and deploy systems whose risks были poorly understood.

Reception была loud, polarized, brief. Через месяцы conversation moved on. К late 2024 letter был artifact earlier discourse.

Три года позже — c shipping GPT-5, Claude Opus 4.X, Gemini Ultra, и волны open-weight models — стоит посмотреть что предлагали, что случилось вместо, что бизнесам забрать.

Что просили

Три specific просьбы:

  1. Шестимесячный pause на training runs more capable чем GPT-4. Ни одна lab не complied. Anthropic, OpenAI, Google DeepMind, Meta AI, Mistral continued training всё more capable systems через 2023-2024.

  2. Development of shared safety protocols. Это partially happened — labs published responsible-scaling policies, frontier-model safety frameworks. Protocols vary в rigor и не externally verified.

  3. Regulatory intervention. Это happened more substantially. EU AI Act passed in 2024 и entered enforcement phases. US executive orders established baseline requirements. UK launched AI Safety Institute.

Actual outcome: no pause, partial industry self-regulation, significant external regulation.

Что бизнесам забрать, в 2026

Three lessons applicable operating decisions:

1. Capability projections были directionally right, timeline off-uneven

2023-е forecasts где AI was к 2026 close на raw capability и off на commercial deployment.

2. Regulatory shifts move slower чем capability shifts, но catch up

EU AI Act took 4 years от proposal к enforcement. Capability frontier moved more в those 4 years чем previous 20.

Practical implication: если building AI-dependent product — watch regulatory pipeline.

3. Provider concentration risk реален и растёт

Три major foundation-model providers (OpenAI, Anthropic, Google) handle majority high-end commercial AI workloads.

Для бизнесов building на AI: maintain optionality. Architect to swap providers.

Что got better

  • Model safety policies real и increasingly substantive
  • Independent evaluation capacity grew
  • Red-teaming practices matured
  • Catastrophic misuse mostly avoided

Что got worse

  • Concentration в small number labs now structural feature
  • Compute costs frontier training grew faster чем expected
  • Deployment velocity consistently outpaced safety-research velocity
  • Job displacement в knowledge work faster чем expected в specific domains

Что это means для нашей работы

Marketing MIX uses AI tools daily across practice. Также routinely turn down work requiring more AI capability чем currently dependable. Intersection "what AI can do" и "what we'll stake client's outcome on" moving target, но в 2026 line roughly:

  • Yes: AI for drafting, research, translation, code generation с senior review, content variation, brainstorming, customer support augmentation.
  • No: AI for autonomous decision-making в regulated contexts, AI for content shipping без human review, AI for any task где failure would damage client's reputation.

Maksym Stepanenko. Последний обзор: 2026-05-13.