Just say no to OpenAI and Anthropic

Original Article ↗ Hacker News Discussion ↗

Critics of OpenAI and Anthropic argue that relying on closed “frontier labs” concentrates power, erodes privacy, and entrenches opaque data-retention practices, even when providers advertise zero-data-retention options. Many commenters advocate shifting to open-weight models and local or regional hosting as a form of harm reduction, citing concerns over surveillance, climate impact of datacenters, corporate governance, and the quasi-religious rhetoric around AGI. Others push back that all major players have similar incentive problems, stressing the difficulty of deciding whom to trust and calling for stronger regulation and more democratic control over advanced AI systems.

Big AI, Politics, and “Big Tech 2.0”

  • Many see AI repeating the social-media/Big Tech pattern: concentrated power, weak regulation, and media/political systems captured by money and hype.
  • Some frame current leadership and media as opportunistic “grift” machines where specifics (crypto, AI) matter less than extraction.
  • Others push back, arguing the last decades have been broadly good and that pessimism is overstated, though they acknowledge inequality and dysfunction.

HN Culture, Astroturfing, and Polarization

  • Several note that HN once skewed skeptical of Uber-style law-flouting, but only turned against such companies after scandals.
  • There are complaints of astroturfing and a strong US bias in AI discussions, especially against non‑US or open models.
  • Emotional attachment to favorite labs/models (and open vs closed) is suspected of driving some of the heat.

Privacy, Zero Data Retention, and Trust

  • Central debate: whether frontier labs can ever be “private” when default policies include logging and short‑term retention for abuse/security.
  • Zero Data Retention (ZDR) exists via some cloud providers and intermediaries, but:
    • It’s inconsistently available across models and vendors.
    • Abuse-monitoring exceptions and flagged‑prompt retention undermine “true” ZDR.
    • Enterprise ZDR is often gated; some see this as privileging large customers.
  • API data is generally claimed not to be used for training, but short‑term retention is common. Some argue no company’s promises should be taken at face value; others say distrust then also applies to smaller “privacy” wrappers.

Cost and Access Tradeoffs

  • Individual subscriptions to frontier labs are described as “insanely good deals,” especially for heavy users, making it hard to justify moving to paid open‑weight inference.
  • Others report comparable or lower costs using efficient open models via third‑party providers, depending heavily on token volume, caching, and parallel agents.
  • There is concern that low prices from frontier labs may be temporary and lock users in.

Model Quality: Frontier vs Open

  • Some say newer open or Chinese models (e.g., DeepSeek, GLM) are “basically on par” with top proprietary models for coding.
  • Others strongly disagree, reporting clear quality gaps, especially on large or complex codebases; they suspect satisfied users are doing simpler work where quality differences matter less.
  • Overall: no consensus; experiences vary by workload.

Ethics, Risk, and Motives of Frontier Labs

  • A major thread disputes whether leaders of US frontier labs are reckless “lunatics,” cynical grifters, or sincere but myopic technologists.
  • Critics argue:
    • These companies knowingly accelerate climate impact and job displacement.
    • Some insiders publicly assign nontrivial probabilities to human‑extinction scenarios yet continue to scale.
    • High compensation and the ability to quit raise moral expectations on non‑janitorial staff.
  • Defenders call for basic human charity, noting not “every human” at these labs behaves badly and that fanatic rhetoric is counterproductive.
  • Some accept that founders think they are building a quasi‑“machine god” or ASI, and a few openly welcome that, prioritizing a post‑human “next phase of life” over current human welfare.
  • Others find this worldview morally horrifying, explicitly seeing it as hostile to the 8 billion people alive today.
  • Multiple commenters call for strict regulation, personal liability for catastrophic harms, and skepticism of any entity (state or corporate) wielding such power.

Alternatives: Open Weights, Local, and Non‑US Providers

  • Many advocate open‑weight models on personal or European hardware, citing:
    • Privacy and control.
    • Competition and ecosystem diversity.
    • Reduced dependence on US frontier labs.
  • Chinese open‑weight models are praised for openness and cost, but some flag the CCP/state‑control risk as a serious moral dilemma.
  • Harm‑reduction framing is common: prefer local, then open‑weight via trusted national providers, and treat payments to major US labs as “worst‑case” where unavoidable.
  • Some emphasize that open‑weight models can be used via neutral providers, avoiding sending money to the original lab.

Broader Societal Anxieties

  • The AI debate ties into worries about wealth inequality, institutional corruption, environmental damage from datacenters, and possible civilizational decline.
  • Others remain optimistic, seeing recent decades as historically good and believing society can course‑correct after experimenting with bad options.
  • A minority suggests simply not using AI at all is a valid and under‑discussed choice.