// Bookmark detail / reading view function BookmarkDetail({ bookmarkId, onBack }) { const b = SAMPLE.bookmarks.find(x => x.id === bookmarkId) || SAMPLE.bookmarks[0]; const collection = SAMPLE.collections.find(c => c.id === b.collection); return (
{/* Hero */}
{collection?.name} ยท Saved {b.time} ago

{b.title}

{b.url} {/* Image placeholder */}
[og:image โ€” {b.domain}]
{/* AI Summary */}
AI Summary

{b.ai_summary}

Key points
    {[ "Scaling alone has flatlined since GPT-5; quality and post-training drive gains.", "Three case studies of recent models that improved without parameter increases.", "Implication: differentiation is about data, RL, and architectural priors.", ].map((pt, i) => (
  • {i+1} {pt}
    • ))}
{/* Tags + actions row */}
Tags {b.tags.map(t => #{t})}
{/* Article body placeholder */}
Article

The frontier of capability has, quietly, stopped responding the way it used to. For a decade, every order-of-magnitude increase in compute begat a corresponding bump on the benchmarks. That bargain has thinned. {b.desc}

What changed isn't a single thing โ€” it's that the cheap wins are gone. Models trained on the open web have memorized most of it. Models trained with RLHF have plateaued where human raters can't reliably distinguish answers. The remaining problems require taste โ€” a sense of what to optimize for that humans themselves struggle to articulate.

"We are not running out of compute. We are running out of obviously good things to do with it."

The next chapter, then, is less about size and more about shape. About picking the right losses, the right environments, the right curricula. About what it means to "post-train" an intelligence rather than to grow one.

{/* Related */}
Related in your library
{SAMPLE.bookmarks.filter(x => x.id !== b.id && x.collection === b.collection).slice(0, 3).map((rel, i) => ( {}} onClick={()=>{}} /> ))}
); } window.BookmarkDetail = BookmarkDetail;