Founder shared 2026-07-05 ("noise or juice?"). Author = real credentialed operator: Stanford → Scale AI → DeepMind → OpenAI research, now building an agent-native startup. 2.8M impressions, 17k bookmarks (~3:1 bookmark:like = a save-this reference piece, not a hot take).
Thesis
AI aces anything with a loss function (gradeable, school-style problems). Durable career value lives in the un-gradeable: selecting which problems to solve, and allocating scarce time/tokens/capital toward them. "The valuable work of the next decade is everything that can't be graded within the span of model training."
Argument skeleton
- Optimize for the truly scarce — capital is cheap now; time, real relationships, reputation are the bottleneck.
- Problem-finding > problem-solving — his firm DELETED Leetcode/system-design from interviews (uncorrelated with performance); now tests how fast you grok a new environment, spot the problem worth solving, execute in constraints. The differentiator isn't solving — it's cheap solving (fewer tokens/time).
- Work the most ambitious form of the problem — bitter lesson applied to careers; simple systems are free now, so pick frontier problems + real shot.
- Sprint the last mile — median output is "what an agent produces from a sloppy prompt"; value is polish/taste/iteration. Often restart from scratch with the next model rather than refine.
- Raise xG AND conversion (soccer metaphor, decoration) — position to see opportunities + actually convert.
- Research is breakable-into — models build intuitions/evals; "being a researcher is a mentality, not an occupation."
Verdict: JUICE (lightly diluted)
Beats generic AI platitudes because it's mechanistic and costed by someone who hires for it. "AI is good at anything with a loss function → value is the un-gradeable" is a sharper reformulation of "focus on judgment." Beat #2 has real proof (deleted Leetcode, replaced with problem-selection). Weak spots: soccer metaphor is decoration; beats 1/3 restate known startup wisdom; name-drop density (turned down Anthropic-at-50, Cursor-at-2) is flex-as-credibility. Net: reference-grade, not throwaway.
RDCO relevance
- Founder's DSA / discovery-scoping role (direct hit): "problem-selection + token/time allocation over raw solving" IS the role framing (R on discovery/scoping) and the COO-agent unhobbling logic. Beat #2 mirrors his daily work — value is identifying the right problem then routing agent effort. Career-lever context for [[project_l5_north_star_strategic_direction]].
- Sanity Check reframe seed (clears no-derivative bar): NOT "cover Chen." The original lens for a data/AI audience: "audit which of your data/analytics tasks are gradeable (AI eats them) vs un-gradeable (your moat)." Worth a research-brief, not an auto-draft.
Related: [[2026-07-05-dan-koe-most-profitable-skill-human-nature]] (same evening, adjacent "what skill to bet on in the AI era" genre)