Three new cards from three operators. Two threads run across them. First: platform operators are using design and detection to reshape who wins at distribution and consent. Second: local AI infrastructure reached an efficiency threshold that makes cloud-first architecture expensive on everyday workloads.
Platform design as the new bottleneck: detection and consent
Kevin Indig published "Slop Antibodies" on August 17 documenting how LinkedIn, YouTube, Reddit, TikTok, and Substack deploy AI content detection that suppresses algorithmic reach without deleting posts. The binding constraint for AI-heavy content is no longer production cost. It is distribution access. A Pangram Labs scan of more than one million LinkedIn posts from April through June 2026 found 41% of long-form posts were fully AI-generated. LinkedIn's detection runs at 94% accuracy and caps flagged posts to the creator's immediate network. Indig's conclusion: production efficiency gains from AI do not offset the distribution losses from suppression. Platform anti-slop detection shifts content scarcity from production to distribution reach.
Eric Seufert's August 18 analysis of Apple's App Tracking Transparency ruling extends this theme from content suppression to consent architecture. Germany's Bundeskartellamt found that Apple used "personalization" in its own consent prompts while requiring third-party apps to tell users they would be "tracked across apps and websites owned by other companies." The word choice is the mechanism. "Personalization" encourages consent; "track you" discourages it. Apple's own advertising collected the same data as competitors but achieved structurally higher opt-in rates because the choice architecture was unequal. The regulator ordered Apple to neutralize prompt designs and allow developers to bundle consent requests into a single prompt. Apple's ATT used asymmetric consent language to systematically advantage its own advertising opt-in rates over third-party competitors.
- Kevin Indig, Platform anti-slop detection shifts content scarcity from production to distribution reach. Platform detection systems cap AI content at the immediate network, making distribution access the binding constraint for content at scale.
- Eric Seufert, Apple's ATT used asymmetric consent language to systematically advantage its own advertising opt-in rates over third-party competitors. Apple's ATT used asymmetric prompt language to advantage its own advertising opt-in rates over third-party competitors, a design choice Germany's Bundeskartellamt found abused market dominance.
Local AI reaches the efficiency threshold: 89% of queries at 20% of the energy
Tomasz Tunguz wrote on August 21 about a Stanford and Together AI research paper titled "Intelligence per Watt." Intelligence-per-watt efficiency for local models rose 5.3x from 2023 to 2026, driven by algorithm improvements (3.1x) and hardware gains (1.7x). The best local model's win-or-tie rate against frontier cloud models rose from 23.2% in 2023 to 71.3% in 2025. The operational implication is a routing layer: send 89% of everyday queries to local models and reserve frontier cloud models for the 11% that require deeper reasoning or current knowledge. Measured savings against an all-cloud baseline: energy cut 80%, compute cut 77%, cost cut 74%. Tunguz frames this as the minimill moment for AI infrastructure. Decentralized, cost-efficient local inference is now cheaper for the high-volume majority of workloads. Routing everyday queries to local models cuts energy 80%, compute 77%, and cost 74% while matching frontier quality on 89% of tasks.
- Tomasz Tunguz, Routing everyday queries to local models cuts energy 80%, compute 77%, and cost 74% while matching frontier quality on 89% of tasks. Stanford research shows local models match frontier quality on 89% of everyday queries while routing to local cuts energy 80% and cost 74% versus an all-cloud baseline.