> "If humans never do it themselves, how would they know what is good?" > > Not a philosophy thesis prompt — this is the scalpel thrown by Silicon Valley's sharpest capital narrative in June 2026.
1. "Cognition as a Commodity": A Narrative Revolution Redefining Labor
Sequoia Capital's AI Ascent 2026 summit (April 20, 2026, San Francisco) was not an ordinary investment forum. Per the author, it was a narrative surgery redefining "human value."
When speakers say "99.9% of human cognitive work can be handed to machines," the audience is not sci-fi readers but decision-makers controlling $85B+ in assets, directing ~60% of new capital into AI (Value Add VC, June 2026). This is not prediction — it is declaration: the arrival of a new commodity — cognition.
In the industrial age, commodities were coal, steel, oil. In the information age: data, attention, algorithms. Now, the claim is that thinking itself can be packaged, priced, and traded. When a company can run without human thought, "hiring people" shifts from "necessary cost" to "optional luxury."
2. The Numbers Behind the Capital Surge
Devin's 8 days vs. Mercedes' 8 months
From Cognition's official blog:
- Project: Mercedes-Benz COBOL system modernization
- Codebase: 200,000 lines
- Traditional timeline: 8 months
- Devin's timeline: 8 days
- ~60% of new capital going to AI (Value Add VC, June 2026)
- Anthropic valuation: $170B four months ago → $350B now (36Kr, Jan 2026)
- Cognition valuation: $26B, ARR $492M (June 2026)
- Clients: Citi, Goldman Sachs, Mercedes-Benz, Dell, Santander, US Army, US Navy, NASA-JPL
- After only 10–15 minutes of AI interaction, independent task performance drops significantly
- Math: AI-assisted group independent success rate 0.57 vs. control 0.73 (P<0.001)
- Skip rate: AI-assisted 0.20 vs. control 0.11 (P=0.031)
- Reading comprehension: AI-assisted 0.76 vs. control 0.89 (P=0.007)
- Someone who has never written code cannot judge AI code quality.
- Someone who has never done research cannot judge AI research rigor.
- Someone who has never thought cannot judge the correctness of AI thinking.
A 30:1 time compression. Paired with another figure: 89% of Cognition's internal code is now committed by AI (up from 13% in December 2025 to 89% in May 2026). Roles have flipped.
A zero-employee company with $10M annual revenue
Polsia — a company claiming to have almost no human employees, run by AI agents handling coding, research, marketing, customer service, and ads — reports annual revenue near $10 million and just closed a $30 million round (Jawlah, May 2026). The traditional "team → product → revenue" chain is compressed into "AI agents → revenue." This is a qualitative change in organizational form.
The capital numbers
The capital consensus: AI is not an investment theme; AI is the new infrastructure — like the internet 20 years ago.
3. Cognitive Atrophy: A Shrinking Brain Underway
The causal hammer from MIT/Oxford
In April 2026, a joint Oxford/MIT/CMU/UCLA study (N=1,222) landed:
> "AI Assistance Reduces Persistence and Hurts Independent Performance"
Causal findings (not correlations):
The sharpest conclusion:
> "People who used AI to get answers directly performed 36% worse than those who never used AI (Cohen's d=-0.66)."
Two mechanisms, one dead loop
1. Hedonic-adaptation reference-point shift: AI finishes in seconds → perception of how long tasks "should" take is distorted → work without AI feels abnormally arduous. 2. Loss of metacognitive calibration: never solving independently → not knowing what you can do → losing the motivation to persist.
These mechanisms are self-reinforcing: each outsourcing lowers the probability of the next independent attempt.
Adolescents: developmental blocking, not atrophy
> "If adolescents outsource skills they have not yet learned to AI, this is not skill atrophy — it is developmental blocking."
A child who has never solved an equation independently will never know if they can. AI is not replacing their labor — it is replacing their developmental process.
4. The Core Paradox: Capital Contradicting Itself
Two statements from the summit narrative, placed together, produce a logical explosion:
> A. "Humans only need to review AI's output." > B. "AI's output will exceed human comprehension."
If A holds, humans must understand AI output to review it. If B holds, they cannot.
Conclusion: humans will review what they cannot understand, and take responsibility for outcomes they do not comprehend. This is not future tense. Cognition's clients (Citi, Goldman Sachs, NASA-JPL) already use Devin on code they cannot quickly verify themselves. Review becomes ritual, not a substantive safety net.
5. The Final Scalpel
"Good" is a concept requiring a frame of reference — a boundary between "good enough" and "not good enough" that is built by doing it yourself.
6. A Non-Conclusion Conclusion
The "is AI good or bad" debate is over — capital voted with money, technology with lines of code, users with usage hours. The real question: do humans realize what they are giving up in this vote?
The MIT/Oxford study gives a timescale: 10 minutes. Sequoia gives: now — 60% of capital, $26B valuations, zero-employee companies. Not a trailer; a released film.
> "Don't worry — even if the world forgets, I'll remember for you." > > But if humans no longer experience, think, or make mistakes for themselves — what is left worth remembering?
Reference Index
| Data point | Source | Verification | |--------|------|----------| | Devin: 8 days vs. Mercedes' 8-month project | Cognition official blog | ✅ Verified | | Devin writes 89% of Cognition's code | Multiple outlets (2025.12 → 2026.5) | ✅ Verified | | Cognition $26B valuation / $492M ARR | Multiple outlets (2026.6) | ✅ Verified | | Zero-employee company, ~$10M revenue | Polsia, Jawlah (2026.5) | ✅ Verified | | Sequoia ~60% of new capital to AI | Value Add VC (2026.6) | ✅ Verified | | Anthropic $170B → $350B | 36Kr / FT (2026.1) | ✅ Verified | | MIT/Oxford 10-minute cognitive atrophy | arXiv:2604.04721 (2026.4) | ✅ Verified | | "99.9% of cognitive work to machines" | Summit talk / paraphrase | ⚠️ Exact source not found |
*Data verified as of 2026-06-25.*