история о том, как мы исследуем разные AI модели в UC Berkeley чтобы они рекомендовали ваши продукты

история о том, как мы исследуем разные AI модели в UC Berkeley чтобы они рекомендовали ваши продукты все началось в конце августа 2025. мы (я и мой ко фаундер) спросили вопрос, как AI модели отвечают в реальном времени, если информация на которой они ищут информацию уже устарела? OpenAI Summer Meetu это было после chatgpt 5 launch мы написали первую research paper на ArXiv Computer Science > Artificial Intelligence arXiv:2509.10762 (cs) [Submitted on 13 Sep 2025] AI Answer Engine Citation Behavior An Empirical Analysis of the GEO16 Framework Arlen Kumar, Leanid Palkhouski View PDF HTML (experimental) AI answer engines increasingly mediate access to domain knowledge by generating responses and citing web sources. We introduce GEO-16, a 16 pillar auditing framework that converts on page quality signals into banded pillar scores and a normalized GEO score G that ranges from 0 to 1. Using 70 product intent prompts, we collected 1.700 citations across three engines (Provo мы ее презентовали на самой большой GEO (generative engine optimization) конференции в Сан Франциско в декабре мы наняли больше research людей чтобы помогли делать эксперименты 1 of 9 мы опубликовали уже новую research статью, которая уже говорит про влияние неправильной информации для AI Search Freshness as a First-Class Reliability Constraint in AI Search: Evidence from Controlled Staleness in Retrieval-Augmented Generation Wrodium March 2026 Abstract Retrieval-augmented AI search systems are increasingly used in enterprise settings, yet con- tent freshness is often under-specified in both evaluation and operations. This paper investigates how stale corpus content degrades end-to-end pipeline behavior and argues that freshness infras- tructure should be treated as a first-class systems requirement. Using FreshRAG, a controlled study spanning five enterprise domains and four staleness conditions (0%, 10%, 30%, 50%), we identify a temporal-semantic failure mode in which conventional metrics remain comparatively stable while temporally valid evidence declines. At 50% staleness, fresh answer-bearing retrieval declines by 24.2% despite minimal change in Precision[95] and Recall@5. Subsequent pipeline stages (reranking, context assembly, and verification-regeneration) only partially mitigate this degradation and can increase computational and economic overhead. We conclude that robust AI search requires explicit freshness infrastructure: measurable freshness objectives, temporal observability, and policy-governed freshness controls. 1 Introduction Retrieval-Augmented Generation (RAG) has become the default architecture for enterprise AI 3 of 9 Fig. 3: Limited Temporal Correction Reranking does not reduce stale intrusion in our experiments. In high-staleness settings, stale passages are frequently promoted to top ranks, suggesting rerankers inherit the same temporal blind spot as hot retrievers (Figure 3). The time-aware breakdown in Figure 3 indicates this distribution is not uniform: time-sensitive intents lose more fresh supporting evidence than time-insensitive ones. 4 немного инсайдов Retrieval Precision Retrieval Recall Stale Document Intrusion Figure 2: Retrieval degradation under increasing corpus staleness. Headline metrics (Precision@5, Recall@5) remain comparatively stable while stale intrusion rises and fresh answer-bearing retrieval declines. Precision@5 Recall@5 Stale Intrusion Rate Figure 3: Time-sensitive queries are disproportionately affected by staleness, showing larger stale intrusion and steeper drops in fresh supporting evidence. 4.3 Context Assembly: Freshness Dilution Context freshness falls to 83.0% at 50% staleness, with stale token ratio increasing to 17.0% (Fig- ure 5). Notably, contradiction density decreases rather than increases, indicating stale content is often compatible but temporally outdated, making contradiction filters insufficient. 4.4 Generation and Verification: High Cost, Modest Recovery Hallucination-related failure rates remain high (roughly 93.8-95.8%). Verification triggers gener- ation for more responses, but only 3.9-10.0% are successfully recovered. End-to-end cost per query increases by approximately 3x, indicating poor economic efficiency of post hoc repair under stale retrieval. YOKUBO мы нашли определение паттерны и они оправдывали наш hypothesis мы нанимаем больше людей, расширяем наш продукт и находим новых клиентов, кто хочет чаще показываться в ChatGPT и других AI моделях для своих клиентов такая вот история. сохраняйте