OpenAI Threatening to Ban Users for Asking Strawberry About Its Reasoning

Original Article ↗ Hacker News Discussion ↗

OpenAI’s warning that users could be banned for prompting its new “Strawberry” reasoning model to reveal its step-by-step thought process is raising both technical and governance concerns. Commenters question whether the restriction is really about safety or about protecting a fragile competitive advantage, tying it to OpenAI’s broader shift from its original “open” nonprofit mission toward a tightly controlled, profit-driven platform. Others note that such bans complicate building on the API, highlight how hard real AI alignment appears to be, and contrast OpenAI’s closed approach with more open releases from rivals like Meta.

OpenAI’s “Open” Identity and Business Model

  • Many commenters see a stark shift from the original “open AI for humanity” non‑profit vision to a closed, profit‑driven platform with some of the least open models in the industry.
  • Some argue “open” was always meant as “open to use” via API, not open source; others say this redefinition makes “open” meaningless.
  • The non‑profit / capped‑profit structure is debated: some note it’s legally common for nonprofits to own for‑profit entities; others see likely “private benefit” problems and possible fraud, referencing ongoing legal disputes.
  • Several say the real driver of secrecy is competitive advantage and valuation, not safety.

Strawberry / o1 Reasoning and Chain-of-Thought Ban

  • The “Strawberry” name is widely read as PR aimed at the meme about GPTs failing to count “r”s in “strawberry.”
  • Banning users for eliciting chain-of-thought (CoT) is seen by many as overreach and a sign they lack confidence in their alignment / safety; others think it’s about hiding an easily copyable “secret sauce.”
  • People worry about collateral damage: casual users, red‑teamers, or downstream app users might trigger bans; this is viewed as a brittle foundation for serious products and a potential attack vector.

Technical Discussion: Tokens, Counting, and Reasoning

  • Long subthread explains why models often miscount letters: they operate on subword tokens, not characters, so can’t natively “see” letters; when correct, they’re likely recalling memorized facts.
  • Others counter that this exposes limits of “reasoning” and highlights that LLMs are sophisticated interpolation systems, not symbol‑manipulating intelligences.

Prompt Engineering and Control

  • One side calls “prompt engineering” pseudoscience propped up by policy and censorship; another credits it with turning LLMs from text generators into usable “knowledge engines.”
  • Speculation appears about prompts that generate forbidden prompts, and about organizational filters controlling which questions can be asked.

Safety, Power, and Governance

  • “For your safety” is framed by some as a common facade for control; others respond that safety motives can be genuine, while still easily abused.
  • A minority expresses strong existential‑risk concerns and suggests AI development should be paused or tightly controlled, even via export controls on GPUs and research.

Ecosystem and Alternatives

  • Several defend OpenAI by noting that without its commercialization we might not have widely accessible frontier models; critics respond that similar capability would have emerged elsewhere, possibly more openly.
  • Multiple commenters report better practical results from competitors (e.g., Claude, open‑ish Meta models) and avoid OpenAI on principle.