Greg Brockman interview [video]
Perception of OpenAI Leadership and “Grift”
- Several see the company as having betrayed its original nonprofit, “for humanity” mission, now dominated by money and power.
- Some argue that if firing a possibly “grifty” CEO would “kill the company,” that itself signals a flawed, personality‑dependent organization.
- Others defend keeping a controversial but effective leader as “the devil you know” and see this as rational in a high‑stakes, hype‑driven field.
Brockman Diary and Billionaire Ethics Debate
- The leaked diary line “what will take me to $1B?” triggers debate:
- One side: wanting $1B is normal and morally neutral; many would use it for security or philanthropy.
- Other side: extreme wealth is inherently exploitative and inconsistent with claiming moral high ground.
- Some note the diary excerpts came via legal discovery; opinions differ on whether they reveal fraud, hypocrisy, or just ambition.
Nonprofit-to-For‑Profit Transition & Governance
- Simplified account from the thread:
- Founded as a nonprofit; later concluded they needed massive compute and funding.
- Created a capped‑profit subsidiary in 2019; nonprofit transferred IP (valued ~$60M) and received capped returns and residual rights.
- Large investments from a major tech company followed; later recapitalized into a public benefit corporation with the nonprofit reportedly holding
26% equity ($200B on paper).
- Some see this as clever mission financing; others as mission drift and a precedent that nonprofits can pivot to enrich insiders.
- Broader criticism of nonprofits: often used for tax arbitrage, political influence, or sham foundations; calls for tighter regulation.
Technical Plan and AI Progress
- “Three‑step technical plan” summarized as: (1) solve reinforcement learning, (2) solve unsupervised learning, (3) tackle increasingly complex tasks.
- Commenters point out that pretraining is actually self‑supervised, not unsupervised, and that OpenAI “accidentally” hit its own goals via large‑scale pretraining + RLHF.
Training Data, Copyright, and Openness
- Strong dispute over whether mass ingestion of copyrighted books and media is “theft” or analogous to a robot reading library books.
- Some stress scale and piracy (e.g., using sites like Anna’s Archive) as ethically and legally distinct from human reading.
- Others argue that human culture should not be privatized, and that current US labs are rent‑seeking on globally created knowledge.
- Chinese labs releasing open‑weights are contrasted with US firms’ closed APIs.
Altman Firing, Petition, and Power Dynamics
- Employees’ pro‑CEO petition (hosted on Google Docs) sparks questions about peer pressure and career risk for dissenters.
- Board’s brief ouster of the CEO and rapid collapse under pressure is seen as a lesson in being outmaneuvered by capital and internal politics.
- Unclear motivations of key scientific leadership (e.g., abrupt shifts from firing to supporting the CEO) remain a focal curiosity.
Dependence on OpenAI and Competitive Landscape
- Builders on the API describe governance drama as exposing fragility of depending on a single vendor.
- Some claim Anthropic is now the “most important” or best for coding; others counter that this reflects hype or narrow benchmarks.
- Google/DeepMind credited for foundational research (e.g., transformers) but criticized for squandered lead.
Views on AGI and LLM Limits
- Skeptics argue current “glued‑together text predictors” are fundamentally not AGI, citing complexity arguments (Shannon vs. Kolmogorov) and inability to reason from first principles.
- Others leave room for uncertainty: LLMs may be part of an eventual AGI stack, though current “agents” and buzz look like a bubble phase.
- Some reject the premise altogether and tune out once “AGI” is mentioned.
Meta: Corporate Drama vs. Real Tech
- A subset finds this kind of leadership/board gossip boring “reality TV,” preferring technical content.
- Others note that for most of the world “tech” now primarily means money, power, and corporate intrigue, not engineering details.