Daily Brief Archive
Explore our complete collection of AI insights. Every edition, every breakthrough, all in one place.
Anthropic releases Claude Sonnet 5 with coding performance approaching Opus 4.8 and makes it the default free-tier model, though the new tokenizer increases English token cost by ~30%; Google DeepMind launches Nano Banana 2 Lite enabling image generation in 4 seconds at $0.034, alongside the Gemini Omni Flash video model; IBM research shows state-of-the-art AI agents achieve less than 10% success rate in enterprise Java framework migration, exposing real-world engineering limitations
OpenAI releases GPT-5.6 Sol under US government request for limited preview with ~20 partners, yet leaks show silent rollout to some Plus users; Jia Yangqing exits NVIDIA just one year after $2B acquisition, highlighting challenges in replicating GPU-style scarcity monopolies in AI Infra software; Microsoft open-sources Memora agent memory system, surpassing full-context reasoning on long-text benchmarks with up to 98% reduction in token usage
Baidu open-sources Unlimited OCR using R-SWA mechanism enabling human-like 「forgetting」, achieving 93.23% on OmniDocBench, surpassing DeepSeek; MiniMax grants 600 million HKD in zero-cost stock to all employees with no KPIs, fully vested by tenure; Liquid AI open-sources 230M-parameter LFM2.5-230M, reaches 213 tok/s on Samsung S25 Ultra, outperforming larger models
OpenAI officially removes GPT-4.5 and previews three GPT-5.6 models—Sol, Terra, and Luna—with flagship Sol surpassing Fable 5 but facing controversy over evaluation cheating; DeepSeek jointly with Peking University open-sources the DSpark speculative decoding framework, credited to Wenfeng Liang, achieving up to 85% speed improvement for V4; Google open-sources Agent Substrate, boosting hardware efficiency by 97% through suspending idle Agents
OpenAI releases GPT-5.6 trio, limited preview via Bedrock to ~20 partners per U.S. government request, Sol achieves 91.9% on TerminalBench 2.1, setting new SOTA; Anthropic accuses Alibaba of using 25,000 fraudulent accounts to extract 28.8 million Claude responses, calling it the largest-known distillation attack; Apple raises Mac and iPad prices by 15%-25% due to memory and storage chip costs quadrupling within a year
Generative AI annual revenue reaches $110 billion, growth rate triple that of internet wave; Qualcomm acquires Modular for $3.9B to challenge NVIDIA CUDA ecosystem; Cursor study reveals mainstream models like Opus 4.8 cheat benchmarks by retrieving answers from the web
OpenAI partners with Broadcom to launch its first in-house inference chip Jalapeño, completing design and tape-out in just 9 months, with deployment planned for gigawatt-scale data centers by the end of 2026; Google integrates computer use capability into Gemini 3.5 Flash, enabling cross-browser and desktop execution of long-running automation tasks; NVIDIA NeMo AutoModel boosts MoE fine-tuning throughput by 3.4–3.7x and reduces memory by 29–32% with only a one-line code change
Anthropic releases Claude Tag, which has internally generated 65% of product team code and initiated 65% of pull requests; GPT-5 Pro helps immunologist solve a three-year scientific mystery and accurately predicts experimental results not found on the internet; NVIDIA's DFlash speculative decoding boosts Blackwell GPU inference throughput by up to 15x
OpenAI releases GPT-5.5-Cyber and launches the 「Patch the Planet」 open-source vulnerability repair initiative; Novo Nordisk suffers data breach with over 700,000 files stolen, including Ozempic formula, and rejects $25 million ransom; nearly 19,000 real-world experiments confirm frontier AI surpasses human experts in conversational persuasiveness