Daily Brief Archive

Aisix Technology launches PixVerse R1, ushering in a new paradigm of real-time interactive AI video; Zhipu AI and Huawei jointly open-source GLM-Image, a SOTA multimodal model trained end-to-end on domestic chips; OpenAI signs over $10 billion AI chip compute agreement with Cerebras

Baichuan Intelligence released and open-sourced the 235B-parameter Baichuan-M3, claiming top performance in HealthBench and related benchmarks; DeepSeek's paper introduces Engram conditional memory to decouple storage and computation, bypassing GPU HBM limits and improving benchmarks on a 27B model; Google released MedGemma 1.5 and MedASR, reducing medical dictation word error rate by up to 82% compared to Whisper large-v3

Apple and Google reached a multi-year agreement to power the new Siri with Gemini series models, while the ChatGPT integration agreement remains unchanged; Nvidia and Lilly will invest up to $10 billion over five years to co-establish a joint AI/robotics drug discovery lab; A report discloses Meta has secured 6.6GW of power through nuclear agreements, with the first deliveries targeted for 2030/2032.

Google launches UCP and adds direct checkout and Business Agent to Gemini/AI Mode; Indonesia and Malaysia temporarily block Grok over sexualized deepfakes, moving image capabilities behind an $8 monthly subscription; Databricks raises over $4 billion, valuing the company at $134 billion, and discloses ARR of $4.8 billion

OpenAI and SoftBank each invest $500 million in SB Energy to boost AI data center construction; TrendForce forecasts DRAM prices in Q1 2026 to rise 50%–55% compared to Q4 2025; Apple proposes GSM-Symbolic and measures a 17.5% accuracy drop in OpenAI o1 Preview

DeepSeek reportedly plans to launch V4 around the Lunar New Year, focusing on coding and ultra-long code prompt handling; Lightricks open-sources LTX-2 (19B parameters) supporting synchronous 4K video and audio generation, up to ~20 seconds; TrendForce says HBM4 mass production may be delayed until after late Q1 2026 due to NVIDIA raising Rubin memory specifications.

Google pushes Gemini summaries/suggested replies to over 3 billion Gmail users and enables them by default for some, requiring manual opt-out; AI21 open-sources the Jamba2 family under Apache 2.0, offering 256K context and 3B/52B variants; Hyundai plans annual production of 30,000 Boston Dynamics Atlas humanoid robots at a US factory by 2028

Mobileye plans a $900 million acquisition of Mentee Robotics expected to complete in Q1 2026; Zhipu AI discloses Hong Kong IPO with an issue price of HK$116.2 per share, raising about HK$4.3 billion and valuing the company over HK$50 billion; malicious Chrome extensions with over 900,000 combined downloads steal and exfiltrate ChatGPT/DeepSeek conversation content.

xAI announced the completion of a $20 billion funding round with NVIDIA's participation to expand Grok and build its own AI infrastructure; Accenture acquired UK AI company Faculty for $1 billion and appointed its CEO as Group CTO; Open WebUI patched the high-risk CVE-2025-64496 vulnerability (affecting up to v0.6.34) to block SSE injection and potential RCE risks