Easy AI Daily News Digest - February 28, 2026

Key points

Industry & Companies

• OpenAI raises $110B at ~$840B valuation: SoftBank invested $30B; NVIDIA invested $30B plus 3GW inference and 2GW training compute; Amazon invested a total of $50B with expanded cloud ties, and AWS will be the exclusive third-party cloud for OpenAI Frontier. Microsoft did not participate but maintains its streamlined partnership.
• Links: OpenAI announcement | Scaling AI for everyone | Amazon partnership | Microsoft joint statement
• Anthropic vs. Pentagon: Anthropic refused to provide Claude versions for mass domestic surveillance or fully autonomous weapons and threatened litigation if labeled a supply-chain risk. The DoD is reportedly considering requiring contractor audits and discontinuation of Anthropic services.
• Links: Anthropic statement | Axios report | NPR report
• Sam Altman says OpenAI shares Anthropic's red lines on mass surveillance and autonomous lethal weapons, while negotiating cloud-hosted, firewalled deployments with the DoD.
• DeepSeek V4 favors domestic hardware: Huawei and other Chinese chipmakers reportedly received early V4 access for optimization; NVIDIA/AMD did not. Analysts note training still relies heavily on NVIDIA — more hardware enablement than a ban.
• Link: Reuters
• Burger King pilots voice AI "Patty" (OpenAI-based, BK Assistant) in ~500 US stores, scoring employees' politeness in real time — raising workplace surveillance concerns.

Models & Capabilities

• OpenAI user data: ChatGPT weekly actives over 900M; Codex weekly actives 1.6M (3x since start of year); 50M+ personal subscriptions; 9M+ paying business customers (up from 3M mid-2025).
• Alibaba Qwen3.5 expansion: new 27B dense, 122B-A10B MoE, and 35B-A3B MoE models under Apache 2.0, 260K context (extendable to 1M); Intelligence Index up to 42 per Artificial Analysis.
• Links: Artificial Analysis | Qwen blog
• LMArena February rankings: open text top three: GLM-5, Qwen3.5-397B-A17B, Kimi-K2.5 Thinking; code: GLM-5 first, Kimi-K2.5 and MiniMax-M2.5 tied second. Arena-Rank tool open-sourced.
• Google Nano Banana 2 (Gemini 3.1 Flash Image): better spatial layout and proportion handling, still hallucinates text; pricing $0.50 input / $3.00 output, roughly half of Pro.
• Link: Official blog
• Qwen3.5-35B-A3B quantization study (Unsloth): hundreds of GGUF experiments; KV q8_0 nearly free quality-wise; the 35B MoE runs ~10x faster than the 27B dense, 60+ tok/s on an RTX 4070 Super.
• Link: Hugging Face experiments
• Sakana AI Doc-to-LoRA / Text-to-LoRA: a hypernetwork generates LoRA weights directly from long documents or text descriptions, "compiling" knowledge into adapters and outperforming raw context windows on needle-in-a-haystack tasks.
• Real-world throughput: Qwen3.5-35B MoE Q4_K_M ~62 tok/s on 4070 Super; GPT-OSS 20B ~100 tok/s on MacBook (1M tokens in 3 hours), fueling local-plus-API hybrid setups.

Infrastructure & Hardware

• vLLM on AMD ROCm: 7 new attention backends, up to 4.4x decode throughput on MI300X via VLLM_ROCM_USE_AITER=1; MLA KV compression squeezes ~8K dims to 576.
• DeepSeek DualPath (arXiv 2602.21548): RDMA-based inference system coordinating prefill and decode nodes to use idle storage/network bandwidth, claiming ~1.9x speedup on a 660B model.
• Link: Paper
• Google Colab adds RTX PRO 6000 at ~$0.81/hour — an order of magnitude cheaper than A100 high-memory instances.
• GPU MODE discussions: PTX acquire-release memory model semantics; cuTile/CuTeDSL multimem-based reduce-scatter for fused operator+communication kernels.

Research & Methods

• Logit Fusion: fusing logits from multiple models/checkpoints during training as ensemble+curriculum without inference overhead.
• Link: Notes
• NNsight 0.6: 2.4–3.9x faster intervention tracing, multi-GPU/multi-node vLLM support, vision-language and diffusion model support.
• CoDA-GQA-L attention: landmark-bank + grouped-query design with two custom Triton kernels cuts KV cache memory, but Eleuther analysis shows PPL rising from 4.81 to 5.75 after replacing 32 layers with only 18.6% of params finetuned.
• World models salon: paper clinics on JEPA/V-JEPA, Dreamer, Genie, Sora, World Labs — "mirror vs. map" (generation vs. representation).
• Link: Paper
• Benchmark methodology debate: whether multi-turn CoT few-shot counts as "template bias."

Agents & Tooling

• LLM Connection Strings: URIs like llm://provider/model?... to configure providers, models, and params in one string.
• Link: Design writeup
• MCP ping semantics: Python SDK requires initialize before ping, complicating container health checks (e.g., Bedrock AgentCore creating throwaway sessions).
• Multi-agent coding workflows: one engineer shipped 118 commits/day over 72 days with 5–10 agents, but context loss and file conflicts persist; communities advise treating agents as assistants, not outsourcing.

Products & Applications

• LLmFit recommends runnable models from your hardware specs (accuracy questioned); Onyx launched a self-hosted LLM leaderboard scoring coding, math, reasoning, and efficiency.
• Codeflash's Claude Code experiment: 76K lines of code across two features; benchmarking found 118 functions with up to 446x speedup potential — LLMs write correct code but rarely profile or optimize.
• Link: SWE-fficiency Benchmark
• Nano Banana 2 in production: clear gains in spatial understanding; text hallucination persists; Banana Pro still preferred for character consistency.

Policy, Governance & Safety

• DoD vs. Anthropic: legal experts note the DoD can restrict contractor usage in military projects but can't easily bar commercial use; seen as a landmark precedent for who draws the "acceptable use" line.
• Jailbreak arms race: communities continue probing Gemini Pro 3/3.1, Grok, Claude, and ChatGPT, with harder alignments but ongoing attack attempts.

📌 Source: Easy AI Daily (zhichai.net)