Anthropic обвинила DeepSeek, Moonlight и Alibaba в дистиляции

Статья о том, как китайские компании массово выкачивают ответы западных нейросеток, чтобы дообучать свои модели. Anthropic говорит о 24 000 фейковых аккаунтов и 16 млн запросов к Claude. Вывод простой: рынок ИИ входит в фазу жёсткого копирования, а китайские модели Qwen, Kimi и DeepSeek могут быстро догонять лидеров дешевле.

➡️ Читайте на сайте: https://aff.top/blog/anthropic-obvinila-deepseek-moonlight-i-alibaba-v-distiliacii

🧠 Ещё больше инсайтов → в канале AFF.top

Geo holdout tests measure true incremental lift; standard A/B can't when treatment leaks across users.
Use geo when the change isn't user-randomizable — pricing, brand campaigns, shared-cart features.
Cost: geo tests have tiny sample (you have ~50 regions, not 50k users), so MDE is large and you need big effects to detect anything.
— A/B ▮▮▮▮▮▮ precision, leaks on spillover
— Geo ▮▮ precision, clean on spillover
Reserve geo for non-splittable changes; default to user-level A/B otherwise.
Read the number, not the story. [design choice · spillover-driven]

Security trust badges lift checkout ~1.5%; review-count social proof lifts ~2.8% — and they don't stack linearly.
From tests run both isolated and combined; combined lift was ~3.4%, not the 4.3% sum.
Both reduce perceived risk, so they compete for the same uncertainty budget. Adding the second gives diminishing returns once the first has reassured the user.
— Badge alone ▮▮▮
— Reviews alone ▮▮▮▮▮
— Both ▮▮▮▮▮▮ (sub-additive)
Test the combination, never assume effects add. Pick the stronger single signal first.
Read the number, not the story. [n=combined · sub-additive]