Daily Brief Archive

Multiple sources indicate DeepSeek plans to release V4 in early March: trillion-parameter MoE, ~32B activated parameters, 1M token context, with priority optimization for Ascend/Cambricon; OpenAI discloses Pentagon contract red lines: bans on domestic mass surveillance and autonomous weapons, using cloud deployment to control risks; China releases national standard system for embodied AI and humanoid robots, covering full industrial chain and safety/ethics

AWS launches OpenAI-compatible Projects API on Bedrock Mantle with IAM isolation and tag-based cost management; OpenAI reaches agreement with the U.S. Department of Defense allowing model deployment on classified networks while prohibiting domestic surveillance and fully autonomous weapons; DeepMind introduces Unified Latents, reporting ImageNet-512 FID 1.4 and Kinetics-600 FVD 1.3

OpenAI completes $110 billion fundraising and partners with AWS to build Stateful Runtime, committing to ~2GW of Trainium compute; U.S. government lists Anthropic as 'supply chain risk' and requires federal agencies to phase out its technology within 6 months; Google Cloud previews Nano Banana 2 on Vertex AI with integrated SynthID and C2PA content credentials

Reuters reports Amazon plans up to $50B investment in OpenAI, with $35B tied to AGI milestones or IPO; Meta's Manus agent exposes SilentBridge zero-click indirect prompt injection vulnerability, CVSS 9.8, risking Gmail data leaks and code execution; MIT introduces TLT method to speed up inference-oriented LLM training by 70%–210% without additional compute

Nvidia Q4 revenue reached $68.13 billion with guidance for next quarter around $78 billion; AMD reportedly signed a more than $100 billion agreement with Meta to deploy up to 6GW of Instinct GPUs, including warrants for 160 million shares; Ant Group open-sourced two 1T models, Ling/Ring-2.5, claiming memory access reduced to 1/10 and throughput tripled with support for 1M context

Inception Mercury 2 achieves approximately 1009 tokens/s and about 1.7 seconds of end-to-end latency using diffusion-based generation, with pricing at $0.25/$0.75 per million tokens; the U.S. Department of Defense demands Anthropic accept 'any lawful use' under a $200M contract, threatening to label it a supply chain risk and invoke the Defense Production Act; Bridgewater forecasts combined 2026 AI infrastructure spending of approximately $650 billion by Alphabet, Amazon, Meta, and Microsoft, up from $410 billion in 2025

Reuters claims DeepSeek may have used embargoed Nvidia Blackwell chips to train its model, potentially violating U.S. export controls; an OpenClaw agent mistakenly published internal threat intelligence marked TLP:RED to the public site ClawdINT.com; reports indicate Anthropic's Claude Code Security research preview detected over 500 long-standing vulnerabilities in open-source projects and provided fix recommendations

Databricks launches Lakebase, a serverless PostgreSQL with pgvector support for RAG and agent memory; 86 countries sign the New Delhi AI Impact Declaration, proposing seven pillars and encouraging open-source collaboration; Google and the University of Virginia introduce DTR and Think@n, achieving 94.7% accuracy at 155.4k tokens on AIME 2025, cutting inference costs by nearly 50%

OpenAI forecasts over $280 billion revenue by 2030 and plans to invest approximately $600 billion in compute infrastructure by 2030; Taalas launches HC1 custom ASIC claiming ~17,000 tokens/s on Llama 3.1 8B at ~250W; OpenAI proposes Harness Engineering, reporting Codex agents delivered a ~1 million line code beta product with almost no human-written source code