The LLM warnings Google fired Timnit Gebru over have all come true

Original Article ↗ Hacker News Discussion ↗

Warnings from the 2021 “stochastic parrots” paper on large language models — about biased outputs, environmental costs, un-auditable training data, and consolidation of power in a few tech firms — are revisited in light of today’s AI systems. Commenters disagree on how well those predictions have held up: some cite real-world examples of biased or unsafe ML systems and opaque training corpora, while others argue modern LLMs show genuine utility and partial mitigation of bias, and question both the quality of the Tumblr article and the narrative that its author, Timnit Gebru, was “fired for being right.” Underneath is a broader debate about how much harm current AI is actually causing, whether critics must also propose solutions, and who should control the values these models encode.

Scope and validity of the “Stochastic Parrots” warnings

  • Commenters list five main warnings: misleading fluency without understanding, bias amplification, environmental cost, non‑auditable datasets, and centralization of power.
  • Some argue that bias, environmental cost, and centralization have clearly materialized; others say the first two are unconvincing, the fourth’s impact is unclear, and the fifth is just another instance of existing tech monopolies.
  • Several note that many cited bias incidents (hiring, healthcare triage, Apple Card) predate LLMs and involved other ML systems, so using them as proof about LLMs is questioned.

Bias, discrimination, and red‑teaming

  • Many agree large models encode social biases; examples include sexist outputs, role assumptions, and ideological tuning (e.g., “anti‑woke” models).
  • Disagreement on significance: some see LLM bias as central and dangerous; others say it’s mostly irrelevant for many day‑to‑day use cases and can be mitigated with RLHF.
  • Debate over what counts as “discrimination,” proxy variables, and whether some statistical differences (e.g. insurance pricing) should be legal or moral inputs.
  • Red‑team vs blue‑team analogy used to argue that critics don’t need to present solutions to be valuable; others say criticism without concrete remedies isn’t worth continued funding.

Understanding vs “stochastic parrots”

  • Some claim recent models show clear competence (theorem proving, complex tasks), so dismissing them as shallow parrots has “aged poorly.”
  • Others argue these capabilities might still be achievable without “understanding,” challenging human exceptionalism rather than validating machine comprehension.
  • Multiple commenters note the word “stochastic” is doing less normative work now; many important human and natural processes are stochastic too.

Environment, data auditing, and model collapse

  • Environmental harms from large‑scale training are widely accepted, though how to balance them with benefits is unresolved.
  • Non‑auditable training sets (e.g., containing abusive or illegal material) are seen as a serious but somewhat under‑specified risk; suggested documentation practices are mentioned but not deeply discussed.
  • Concerns about feedback loops from AI‑generated data degrading low‑resource languages appear but are viewed by some as vague or under‑substantiated.

Centralization, open models, and power

  • Some view the concentration of model training capacity in a few firms as a major new form of cultural and linguistic control.
  • Others think this will self‑limit as compute becomes cheaper and open/self‑hosted models reach “good enough” capability.
  • There is concern that whoever controls these systems also controls which biases are embedded or removed.

Gebru/Google dispute and meta‑discussion

  • Strong disagreement over whether the researcher was “fired” or effectively resigned after issuing demands; the terminology is contested.
  • Some see the episode as evidence that big tech wants ethical credibility without true academic freedom.
  • Several criticize the Tumblr post itself as under‑sourced, possibly AI‑written clickbait, and not the rigorous retrospective the topic deserves.
  • Meta‑concerns include HN flagging behavior, low comment counts, and the quality of debate around bias and “wokeness.”