Ilya Sutskever: We're moving from the age of scaling to the age of research

Meaning of “age of research” vs “age of scaling”

• Many read the title as: brute-force scaling is running out of cost-effective gains; future progress requires new ideas.
• Several argue scaling is physically and economically constrained: power, chips, data center capex, and data availability are hitting limits; price/performance isn’t improving fast enough.
• Others insist there is still substantial room to scale (more compute, larger runs, more simulations) and point to ongoing improvements and analyses that say scaling can continue for years—though with diminishing returns.

Skepticism about SSI and the business story

• A recurring “translation” of the interview: “The old scaling cow is out of milk; please fund our new research cow.”
• Strong skepticism that a company with no product, very vague public roadmap, and long timelines (“5–20 years”) justifies tens of billions in valuation.
• Critics emphasize lack of a revenue story and see this as another manifestation of ZIRP-era thinking and VC FOMO: investors bet on every plausible AGI team to avoid missing the winner.
• Defenders note that if you believe AGI is possible, backing top frontier researchers is rational; if not, the whole sector is a shovel-seller’s gold rush anyway.

Moats, secrecy, and IP

• Debate over whether secret training tricks (data curation, shuffling, architectures) meaningfully differentiate labs, or whether most insights are quickly rediscovered.
• Some argue secrecy is partly about safety (avoiding an arms race) and partly about maintaining a competitive edge; others see it as incompatible with claims of building “safe superintelligence.”
• Reputation and brand are seen as giving a growth boost (hiring, media, early users) but not a true moat.

Technical limits: generalization and intelligence

• Several commenters agree with the claim that current models generalize much worse than humans despite massive data.
• Examples: failures on simple tasks (letter counting, inconsistent code fixes), sycophantic outputs (overpraising certain figures), and inability to reason about what’s important in a text.
• Long subthreads debate:
  • Whether “intelligence” is even well-defined;
  • How evolution and inherited structure give humans extreme sample efficiency;
  • Whether LLMs are just sophisticated compressors of text vs anything like brains.

Economic impact and integration

• Many note the models feel “smarter than their economic impact”: the bottleneck is integration into workflows and products, not raw capability.
• Expectation that the next few years will be about:
  • Better engineering (agents, tools, product integration),
  • Local/smaller models and efficiency,
  • Figuring out viable business models rather than chasing ever-bigger training runs.
• Some see “age of research” as a euphemism for an impending AI winter; others think we’re entering an “era of engineering” and digestion rather than collapse.