Daily Brief Archive
Explore our complete collection of AI insights. Every edition, every breakthrough, all in one place.
ByteDance releases Doubao-Seed 2.0, claiming token prices are about an order of magnitude lower than top industry models and reporting HLE-text 54.2; Google and OpenAI warn of 'distillation attacks,' citing over 100K prompts probing Gemini and alleging DeepSeek bypassed restrictions to scrape outputs; Tencent HunYuan releases GradLoc, tracing RLVR training gradient anomalies to the token level and proposing mitigation strategies including TokenClip, SeqClip, and LayerClip
MiniMax M2.5 launched on February 12 and open-sourced on February 13, achieving 80.2% on SWE-Bench Verified with Lightning version output exceeding 100 TPS; Anthropic completed $30 billion in funding, valued at $380 billion, and disclosed annualized revenue of approximately $14 billion; Ant Group open-sourced Ring-2.5-1T, reducing memory access by 90% for long texts over 32K while tripling throughput, self-scoring 35/42 on IMO 2025
Anthropic completes $30 billion funding round, reaching a $380 billion valuation and reporting $14 billion in annualized revenue; Google upgrades Gemini 3 Deep Think, achieving 84.6% on ARC-AGI-2 and 3455 Elo on Codeforces; OpenAI launches GPT-5.3-Codex-Spark powered by Cerebras WSE-3 wafer-scale chip
Zhipu AI open-sources GLM-5: 744B MoE, 28.5T tokens, focused on long-horizon Agent tasks; DeepSeek upgrades web and app models to support 1M token context and reportedly prepares V4 flagship; Claude Desktop Extensions exposed to CVSS 10/10 zero-click RCE triggered via Google Calendar
Alibaba releases Qwen-Image-2.0 with an AI Arena text-to-image score of 1029, supporting 1K token input and 2K output; OpenAI tests ads on ChatGPT free and Go users in the U.S., stating it won't affect responses; Runway completes $315 million Series E funding round with a post-valuation of $5.3 billion to advance next-generation world models and video generation
Microsoft discloses GRP-Obliteration, a single prompt that can degrade safety alignment in 15 mainstream LLMs and may affect diffusion models; Databricks reportedly raises ~$5B, valued at ~$134B ahead of IPO; Hugging Face releases Transformers.js v4 preview, introducing a C++ WebGPU runtime that achieves ~4x speedup for BERT inference and reduces package size by 53%
AWS Bedrock silently launches Kimi K2.5 and GLM 4.7 with tool calling support; New York advances NY FAIR News Act mandating AI-generated news labeling with human review and proposes 3-year halt on new data center permits; OpenClaw discloses 341 malicious files in skill marketplace and integrates VirusTotal for pre-release full scanning
Databricks raises $4 billion at $134 billion valuation; SAMR announces 5 AI unfair competition cases with fines up to 360,000 RMB including DeepSeek/ChatGPT impersonation; Columbia and Microsoft RWML boosts ALFWorld success rate from 13.0% to 32.6%, reaching 87.9% with rewards
Four tech giants' AI capital expenditures in 2026 are projected to total nearly $700 billion, with Amazon expecting free cash flow to turn negative between -$17 billion and -$28 billion; Google reports GKE Inference Gateway reduces Vertex AI's first-token latency by 35% and improves P95 latency by 52% under burst traffic; Alibaba open-sources Zvec embedded vector database, achieving over 8,000 QPS retrieval throughput on the Cohere 10M dataset