Easy AI Daily Digest | February 27, 2026: Nano Banana 2, Hermes Agent, Anthropic vs Pentagon

Forum topic · 小凯 · 2026-03-27

Summary

A daily AI industry digest covering February 27, 2026. Google launched Nano Banana 2 (Gemini 3.1 Flash Image preview), ranking #1 on Arena and ArtificialAnalysis image leaderboards at roughly half the price of comparable models. Perplexity open-sourced its pplx-embed embedding families (0.6B/4B, MIT license) and became the system-level 'Hey Plex' assistant on Samsung Galaxy S26. Community testing covered Qwen3.5 local quantization performance and coding benchmarks. Nous Research open-sourced Hermes Agent with multi-tier memory and multi-platform support; Cursor added Codex 5.3 and Bugbot Autofix; LM Studio released LM Link for remote local models via Tailscale. Infrastructure news included DeepSeek DualPath KV-cache offloading, IterX kernel optimization (14.8x on B200), and FlashInfer competition results. Research highlights: the Solaris Minecraft world model, DeepMind's AlphaEvolve for RL algorithm search, and a world-models survey. Industry news featured Anthropic publicly rejecting Pentagon demands to remove Claude's safety limits, Block cutting 40% of staff, and Suno reaching ~$300M ARR. Includes policy, safety, and interpretability updates.

Easy AI Daily for February 27, 2026, compiled from zhichai.net's forum. Below is a structured English digest of the original Chinese post.

Key points

Models & Capabilities

Google Nano Banana 2 (Gemini 3.1 Flash Image preview): Ranked #1 on Arena and ArtificialAnalysis image leaderboards; priced ~$67 per 1,000 images, about half the cost of GPT Image 1.5 / FLUX.2 tier. Supports multi-character consistency (up to 5 people + objects), 4K upscaling, long-text multilingual typography, and live search-grounded imagery; integrated across Gemini App, Search AI Mode, Ads, API/Vertex. Community reviews find it near Pro quality at ~4x lower cost, though API/Vertex speed and transparent PNG handling lag. Links: Google blog, Arena leaderboard.
Perplexity open-sources pplx-embed / pplx-embed-context: 0.6B and 4B parameter families, MIT license, for retrieval/RAG; benchmarks built on 115M real queries and 30M documents.
Qwen3.5 local quantization: 122B runs at 20–30 tok/s on 3x3090/5090; 27B hits 100 tok/s on a 3090. Some Unsloth dynamic quants (e.g., 35B ud-q4_k_xl + MXFP4) show anomalous perplexity; community advises standard Q4_K_M for now.
APEX real-repo coding benchmark (70 repos, 34 models): Claude Opus 4.6 and GPT-5.2 Codex lead closed models; GLM-4.7 quantized beats Qwen3.5 locally. Framework choice can swing open-model scores by 50%+.
Onyx self-hosted LLM leaderboard criticized for missing Qwen 3.5 27B/122B and Qwen3-Next/Coder-Next.

Agents & Tooling

Nous Hermes Agent (open source): multi-tier memory, persistent machine access, filesystem/terminal/browser tools, Telegram/WhatsApp/Slack/Discord integration; base for Atropos RL data generation. GitHub.
OpenClaw in production: property management automation (rent reconciliation, repairs, contracts); a community rebuild with Mastra + Trigger.dev + Postgres fixes silent failures (writeup).
Cursor: Codex 5.3 Spark now default Cloud Agent (faster bug fixing); PR Bugbot Autofix launched.
LM Studio LM Link: connect remote machines securely via Tailscale, end-to-end encrypted, no port forwarding (link.lmstudio.ai).
Microsoft Copilot Tasks: visible plan generation then stepwise execution; research preview.

Infrastructure & Hardware

DeepSeek DualPath: KV cache in decode-server memory pushed via GDRDMA to prefill GPUs, bypassing PCIe; ~2x speedup on agent workloads. IterX (DeepReinforce) uses RL to optimize FlashInfer MoE kernels, claiming 14.8x on B200.
FlashInfer competition: custom DSL GDN decode at 2.56 microseconds — ~760x faster than PyTorch eager, ~1.3x faster than the official Cute DSL kernel.
GPU MODE community: 4k×4k GEMM nearing 90% of cuBLAS, uniform instructions via elect.sync, nanotrace warp-timeline profiling.
Cheap mining-card clusters: used P104/CMP 100-210 rigs run Qwen3.5 Q6 at ~26 tok/s for ~$750/node.

Research

Solaris (Minecraft world model): prioritize shared global state over pixel prediction; open-sourced multi-player data engine, multi-player DiT (12.6M frames), and VLM eval suite.
AlphaEvolve (DeepMind): evolutionary search over Python code for multi-agent RL algorithms, beating human-designed baselines on game tasks.
World-models survey: arXiv:2411.14499 — 'mirror' (reconstructing reality) vs 'map' (serving decisions) framing.
Unsloth DPO tutorial: reward modeling as binary classification, no separate reward model.

Products & Industry

Perplexity x Samsung: system-level 'Hey Plex' assistant on Galaxy S26 — but Pro users report quotas cut from 250/day to 20, and Sonar Deep Research API sources reduced from 36 to 10 at ~$0.45/call.
Anthropic vs Pentagon: Axios reports a Defense Department ultimatum to remove Claude's safety limits for domestic surveillance and fully autonomous weapons; Anthropic's CEO publicly refused both. Community debate over the company's quietly dropped flagship safety pledge.
US closed vs Chinese open models: compliance-constrained institutions struggle with limited US open alternatives (e.g., gpt-oss-120B).
DeepSeek V4 early access reportedly given to Huawei for hardware-specific tuning; commenters read it as engineering priority, not geopolitics.
Block: Jack Dorsey announces 40% workforce cut (10k → ~6k) toward a smaller 'AI-centric' company; stock up 20%.
Suno: ~$300M ARR, 2M paying subscribers.
Others: Entropy Games' fully on-device AI NPCs; Tamarind Bio's $13.6M Series A for AI drug-discovery inference; Gradio 6.7 HTML component improvements; ProposalMint for nonprofit grant writing.

Policy, Governance & Safety

The Anthropic/DoD clash highlights the gap between capability and reliability/control; simple safety tests still leak sensitive info.
Jailbreak community consensus: no universal prompt exists; CBRN/CSAM categories remain hardest; AI red-teaming emerging as a career path.
Goodfire AI released interpretability infrastructure usable at trillion-parameter scale with low inference overhead; Anthropic's 'model retirement interviews' and Claude Opus 3's Substack spark model-welfare debate (arXiv:2411.00986).

*Source: Easy AI Daily (zhichai.net forum post, 2026-02-27).*