korrents
Brad DeLong

What Brad DeLong thinks about LLMs

@brad-delong · 106 positions · On ppll ↗

Economics professor at the University of California, Berkeley and a former Deputy Assistant Secretary of the US Treasury. Author of Slouching Towards Utopia and writer of the Grasping Reality newsletter.

Brad DeLong did not write this page.

We collected these quotes from things they published elsewhere, and every quote links to where it was said. They have no account here and have not endorsed this site. Quotes are word for word; the short line under each one is our own restatement, not their wording. Their own site. Is this you? Claim it or ask us to remove it. Or tell us what is wrong here.

13 dated positions, 2026, in their own words. Our reading of what Brad DeLong has said — not written or endorsed by them.

  1. LLMs are extraordinarily good at producing text that has the shape of the right answer.

    ↗Fragile Agentic Systems Are Fragile: LAUGH OF THE DAYbraddelong.substack.com

  2. 10 days earlier
  3. When a language model describes its own state, it is just giving the most likely continuation of the conversation with some random noise mixed in.

    ↗On Split-Personality Disorder & MAMLM Agent-Harness Systems: LAUGH OF THE DAYbraddelong.substack.com

  4. 11 days earlier
  5. Whenever the answer can be checked (code runs, a URL loads, a proof verifies), these systems now seek truth reliably.

    ↗ChatGPT, Claude, Gemini, & Co.: They Are Not Brains, They Are Kernel-Smoother Functions: HOISTED FROM THE ARCHIVESbraddelong.substack.com

  6. A model’s starting map weights things largely by how often they appear in the corpus, and fine-tuning only partly corrects that. “Popular” and "true” are not the same thing.

    ↗CROSSPOST: HENRY FARRELL: Agentic AI Is Myth Set in Motionbraddelong.substack.com

  7. 2 days earlier
  8. vibe codingPython

    “Here is the number” and “here is the proper context for the number” are things that are doable at a human-scale level of reliability by MAMLMs vibe-coding python scripts and then stitching them together with the appropriate boilerplate words. That is going to change the economics profession profoundly

    ↗500,000 “Lines of Code” to Try to Track the Economy: TUESDAY TRENDS & TURMOILbraddelong.substack.com

  9. 12 days earlier
  10. That creation and transmission of meaning through the warping of language into new forms is not something an LLM, for which all there is is the text, can do.

    ↗What LLMs Can & Do Do for Me, & What They Cannot & Don’t: THURSDAY THAUMATURGYbraddelong.substack.com

  11. 2 days earlier
  12. It is not because it will give rise to a Digital God. To think that is still seems to me to be a category error of the same kind as thinking that a steamship is the same order of being as a dragon.

    ↗CROSSPOST: JOHN GRUBER & BEN THOMPSON: Meanness & Claudebraddelong.substack.com

  13. Properly trained, RAGed, and harnessed LLMs may still be stochastic parrot, but as they are enmeshed in the spheres of human cognition and public reason they increasingly do not act as stochastic parrot.

    ↗CROSSPOST: JOHN GRUBER & BEN THOMPSON: Meanness & Claudebraddelong.substack.com

  14. 4 days earlier
  15. LLMs are a newer, smaller shoggoth. They condense human writing into useful but lossy summaries, no lossier than prices, bureaucratic categories, & opinion polls.

    ↗Henry Farrell of Johns Hopkins Agora Comes to Berkeley: FOG & FLATLANDSbraddelong.substack.com

  16. 2 weeks earlier
  17. it is difficult to think of a way less helpful for coherent thought to think about “AI” than to think of it as a “Digital God”.

    ↗CROSSPOST: HENRY FARRELL: Machine God Metaphors Eat Your Brainbraddelong.substack.com

  18. That means that the important and live questions are about what metaphors—what abstraction layers, throwing away information so that the problems become things we can grasp and think about—are useful right now?

    ↗CROSSPOST: HENRY FARRELL: Machine God Metaphors Eat Your Brainbraddelong.substack.com

  19. Henry Farrell’s “MAMLMs are a socio-cultural technology” is one very useful one.

    ↗CROSSPOST: HENRY FARRELL: Machine God Metaphors Eat Your Brainbraddelong.substack.com

  20. 3 days earlier
  21. Anthropic

    The subscription tiers are still enormous loss leaders: $20/month for your first $200/month of tokens at API rates for the "Pro" tier, and $100 or $200/month for your first $1500 or $3000/month of tokens at the "Max 5x" or "Max 20x" tiers.

    ↗CROSSPOST: SHELLAC: Suspiciously Precise Floats, or, How I Got Claude’s Real Limitsbraddelong.substack.com