Cur8 — Sunday, September 20, 2026

Alibaba’s Qwen-Image-2.1 generates images with comparable quality to DALL-E 3 while using a smaller 7B parameter model and requiring less compute—demonstrating efficiency gains in diffusion models. It's available now via API.

Cache-to-Cache (C2C) enables direct semantic communication between LLMs via KV-cache, boosting accuracy by 6.4-14.2% and accelerating latency 2.5x compared to text-based methods. The technique avoids token generation overhead while fusing specialized semantics; code is available at [https URL].

Researchers exploited Open AI vulnerabilities via Anthropic's Claude, accessing employee accounts and data in July; Open AI has since implemented security changes. Anthropic’s self-improvement using Claude now leads over 25% of its R&D, raising control concerns.

NYT lawsuit reveals Microsoft directors view AI scraping as "largest theft of labor," while OpenAI calls ChatGPT an "existential threat" to publishers, potentially undermining fair use defenses. Microsoft now concedes paywalled content requires licensing for training.

TypeSafe AI released Jev, a new "machine-native" AI model, claiming it’s 40-200x faster and 238x cheaper than existing LLMs like Fable 5.1 while eliminating hallucinations. It outputs calibrated probabilities for structured decisions, enabling automation beyond human language processing—a key limitation of current LLMs.

AI threat mitigation is likely too late; leaked models and autonomous learning pose risks requiring US-China cyber collaboration. Experts advocate independent AI verification & a pause on superintelligence development until safety measures are established. Claude now handles 26% of Anthropic’s R&D, raising control concerns.