OpenAI delays launch of open-weight model

Original Article ↗ Hacker News Discussion ↗

DeepSeek, Costs, and “Openness”

  • Some compare DeepSeek and Qwen favorably to US labs, seeing them as cheaper, more efficient, and more open, dubbing US firms “money and compute eaters.”
  • The oft-quoted “$5M DeepSeek cost” is heavily disputed: multiple commenters note this only reflects GPU hours for a final run, excluding salaries, facilities, and prior experiments. Estimates for total cost stretch to many tens of millions.
  • Others point out DeepSeek relied partly on outputs from closed models (e.g., OpenAI APIs) and may have violated TOS; the true level of “FLOSS” and independence is questioned.
  • Debate over which US companies are really open: Google is praised for T5/FLAN and Gemma, but others note Gemini is closed and Gemma isn’t OSI-open; similarly, some push back on calling DeepSeek fully open.

OpenAI’s Delayed Open-Weights Model and Competition

  • Several speculate the delay is performance-driven: OpenAI may not want to ship a model that looks weak next to strong new open-weight releases like Kimi K2, or is just “middle of the pack.”
  • Others suggest they’re reallocating effort to beating Grok 4 and larger rivals, or doing last-minute “benchmark hacking.”
  • Some users hope for a ~20B-parameter open model suitable for local use; rumors of “multiple H100s to run it” suggest something much larger and less accessible.
  • There’s a broader sentiment that OpenAI’s post-GPT-4 models are no longer clearly ahead, and that talent churn plus commoditized “genius engineers” weaken its edge.

“Safety Tests”: Genuine Concern or Marketing?

  • A long subthread questions whether “safety tests” are mostly PR and regulatory theater—primarily about censoring offensive content, avoiding PR disasters, and protecting providers.
  • Critics argue LLMs are just “machines that talk,” comparable to books, PDFs, or generic computers; harms should be handled like any other speech or tool misuse.
  • Others insist safety is substantive: LLMs can give dangerous medical instructions, worsen mental health crises, amplify hate, or embed bias in automated decisions. Tool use raises stakes further.
  • There’s tension between:
    • Viewing big-lab “AI safety” as a way to lobby for regulation that hobbles smaller competitors, and
    • Acknowledging that top labs do invest in dedicated safety teams and show clear pre/post-alignment differences.
  • Multiple comments note that open-weight models can be easily uncensored via fine-tuning or “jailbreaks,” undermining provider-imposed safeguards.

Motives, Licensing, and Broader Cynicism

  • Many doubt “safety testing” is the real cause of the delay; alternative theories include distancing from Grok’s MechaHitler incident and simple PR timing.
  • Questions remain about what license OpenAI will use (Meta-style restricted vs truly open), and what actual business benefit it gets from releasing open weights.
  • Several comments express generalized cynicism: joking about “ClosedAI,” “hedging humanity” via prediction markets, and the declining seriousness of Twitter/X as a venue for such announcements.