Mark Zuckerberg attacks 'closed' AI rivals as Meta returns to open models

Perceived motives and strategy

• Many see Meta’s “return to open models” as a business move, not idealism: Meta is behind frontier labs, so “open” is used to commoditize rivals’ advantages and weaken closed incumbents.
• Others argue this fits Meta’s long-standing pattern of open-sourcing non-core tech (PyTorch, React, LLaMA family, SAM, etc.), so it’s not purely desperation.
• Some expect a future rug‑pull: Meta will stay “open” only while it’s losing; if it ever has a clear lead, models may become closed.

“Open weights” vs true open source

• Strong pushback on calling LLaMA-style releases “open source”:
  – Weights are published, but training data and full training code are not.
  – Licenses often restrict commercial/competitive use.
• Several liken this to shipping binaries without source: useful, but not FOSS.
• Others say weights are where the practical value is (self‑hosting, finetuning), and insisting on fully open data is unrealistic given copyright risk.

Competition, regulation, and geopolitics

• Some welcome Meta’s stance as a counterweight to calls from OpenAI/Anthropic for heavy regulation and potential regulatory capture.
• Chinese open‑weights models (e.g., Kimi, GLM) are cited as proof that open frontier models can exist, but also as examples of “open weights under contract,” not fully free.
• Opinions split on export controls: several support chip controls on China but oppose restricting open models; others see controls as a way to entrench Western incumbents.

Trust, ethics, and Meta’s track record

• A large subset rejects giving Meta moral credit, citing: early contempt for user privacy, addictive design, political manipulation, harms to children, and cases like Myanmar.
• Some argue that even if motives are bad, open models that dilute centralized power are still a net positive.
• Others say Meta’s broader behavior means any “individual empowerment” rhetoric should be treated with extreme skepticism.

Safety, misuse, and societal risk

• Concerns raised that open powerful models will supercharge scams, spear‑phishing, deepfake identity theft, and cyberattacks; some compare to making nuclear tech cheap and ubiquitous.
• Counter‑view: these abuses will happen with or without open weights; at least openness prevents a few firms from monopolizing both capabilities and abuse revenues.
• Debate over AGI: some want faster progress and see resource limits as natural speed bumps; others think AGI will destroy economies and democracy and should not be hurried.

Technical and economic considerations

• Disagreement over whether frontier models are effectively “closed” due to compute cost.
  – One side: ordinary users can’t run Kimi‑class models, so openness is academic.
  – Other side: renting GPUs, group use, and future cheaper hardware make open weights materially useful, especially for SMEs and as a fallback if vendors rug‑pull.
• Open models are seen as pressuring inference prices downward and preventing monopoly pricing, but there’s confusion about sustainable business models for open‑weight labs.

Reactions to Zuckerberg’s “personal superintelligence” vision

• Some like the idea of capable personal agents handling drudgery (admin, insurance quotes, therapy search, legal triage).
• Others find the vision dystopian or hollow:
  – It assumes analysis/optimization, not systemic constraints, are the main human problems.
  – A life where an agent does “everything” feels depressing and disempowering.
  – Meta’s own history of addictive, manipulative products undercuts claims about “empowerment.”
• His critique of AI doomers and centralized control resonates with some, but is derided as hypocritical given Meta’s walled gardens and surveillance‑driven ad business.