Four new cards from three operators. Ethan Mollick on how to design agentic systems that preserve human judgment. Tomasz Tunguz twice: once on the compounding economics of the AI model factory and once on how frontier AI markets are shifting from token-based utility to access-controlled infrastructure. Kevin Indig on why freshness outweighs every other ranking factor for listicle content.
Designing agentic systems that keep humans in the right decisions
Ethan Mollick published "Agency and Agents" on August 31. His argument is not that agents are too autonomous. It is that they become autonomous in the wrong places. A system that handles every interesting choice and routes only failures and edge cases to humans slowly removes humans from all the work that builds judgment and motivation.
The proposed alternative, which Mollick and Lilach Mollick call the Twilight Factory, deploys facilitator agents whose job is to identify when a decision should go back to a person. The four scenarios: authorization (external contact, spending, sensitive access), expertise (specialist knowledge where AI capabilities are uneven), variance (injecting human diversity to counter AI output homogeneity), and interest (decisions that are engaging enough that a human should make them, per Sid Meier's game-design principle). The last scenario is the one most organizations ignore.
"If agents make every interesting decision and leave people with approvals, the exceptions, and the failures, we will have automated the wrong half of the job."
This is a design principle, not a policy position. It requires that product teams specify which decisions are "interesting" before building the automation. Agentic systems need a facilitator layer that routes decisions back to humans in four specific scenarios.
- Ethan Mollick, Agentic systems need a facilitator layer that routes decisions back to humans in four specific scenarios. Fully automated agent systems erode human skill by handling every interesting decision; facilitator agents should route four specific decision types back to humans.
Two angles on frontier AI economics
Tomasz Tunguz published two pieces in the last six days that together describe the structural economics of frontier AI. They are best read together because they cover opposite ends of the same market.
The first, from August 27, is about production economics. AI model labs convert megawatts to intelligence at a 3-5x margin. Tunguz cites SemiAnalysis analyst Dylan Patel on Anthropic's specific numbers.
"The base cost of compute tends to be around 10 or 13 or $15 million per megawatt. In the case of Anthropic, the revenue has gone as high as $50 million per megawatt."
The margin funds the next training run. Each run produces a more parameter-efficient model, which lowers inference cost per output unit while sustaining or improving the margin. Jason Warner of Poolside names the structural asset: "The Model Factory is the compounding asset." Anthropic's trajectory confirms the pattern: -94% gross margin in 2024, 40-50% in 2025, $10.9B revenue with $559M operating profit in 2026. AI model labs generate 3-5x revenue per megawatt over cost, and the margin compounds into the next training run.
The second piece, from August 31, describes what happens at the demand side when that production advantage is converted into access control. Frontier AI is segmenting into closed camps. Anthropic restricted Mythos 5 to "Project Glasswing," a named program that rations access to trusted partners. OpenAI released government-facing GPT-5.6 variants to select partners before any public availability. Z.ai requires security review for hosts exceeding $10B in aggregate revenue. Fable went dark outside the US for three weeks following Commerce Department restrictions. The scarcity is no longer price. It is permission. Frontier AI has shifted from token-based utility to access-controlled infrastructure where permission is the new scarcity.
The two pieces connect: the model factory economics produce the capability differentiation that makes access worth rationing, and the access control programs create the enterprise relationships that sustain the revenue margin.
- Tomasz Tunguz, AI model labs generate 3-5x revenue per megawatt over cost, and the margin compounds into the next training run. AI model companies convert $10-15M per megawatt compute cost into up to $50M per megawatt in revenue, and the margin funds a compounding training loop.
- Tomasz Tunguz, Frontier AI has shifted from token-based utility to access-controlled infrastructure where permission is the new scarcity. Frontier AI markets have shifted from token-based utility to access-controlled infrastructure, where partner program approval is the new scarcity.
The one signal that predicts listicle ranking
Kevin Indig published his listicle ranking study in Growth Memo on August 31. He analyzed 83 manually reviewed pages across 8 industries. Freshness was the dominant signal. Pages with dates from the previous two years appeared in top-3 positions 66.6% of the time versus 57.3% for positions 8-10. Old or unreadable dates correlated with 56% lower odds of ranking in the top 3.
The numbered format also matters. Listicles framed as "17 Best Tools" outperformed editorial roundups in 56.8% of direct matchups, averaging position 3.77 versus 4.59. Author bios and schema markup showed no independent lift.
The practical implication is simple. A stale, high-quality listicle with an old date is likely underranking for structural reasons, not content reasons. Updating the date is the first test. Freshness is the single strongest predictor of listicle ranking, outweighing author signals and schema markup.
- Kevin Indig, Freshness is the single strongest predictor of listicle ranking, outweighing author signals and schema markup. Freshness is the single strongest predictor of listicle ranking success, with recent-year dates appearing in top-3 positions at 66.6% vs 57.3% for positions 8-10.