Easy AI Daily News Digest | February 27, 2026

A curated roundup of AI industry news from the zhichai.net community, covering models, agents, infrastructure, research, products, and policy.

Models & Capabilities

Google launches Nano Banana 2 / Gemini 3.1 Flash Image

Google released Nano Banana 2 (Gemini 3.1 Flash Image preview), ranking #1 on Arena and ArtificialAnalysis image leaderboards at roughly $67 per 1,000 images—about half the price of Pro-tier rivals like GPT Image 1.5 and FLUX.2. It supports multi-character consistency (up to 5 people plus multiple objects), 4K upscaling, long-text and multilingual typography, and real-time image retrieval from Search, with rollout across Gemini App, Search AI Mode, Ads, and API/Vertex.

• Google official blog
• GoogleDeepMind release thread
• Arena image leaderboard

Community testing finds Nano Banana Pro still slightly better at curved text, non-human characters, and complex compositions, but Nano Banana 2 is close in quality, faster, and ~4x cheaper. Current API/Vertex speed is slow with occasional errors, and transparent PNG handling remains a weakness.

Perplexity open-sources pplx-embed

Perplexity released the pplx-embed and pplx-embed-context families (0.6B and 4B parameters) under MIT license for retrieval/RAG, claiming industry-leading results on an internal benchmark of 115M real queries and 30M documents. Available on HuggingFace and its own API.

Qwen3.5 local quantization: performance and pitfalls

Reddit testing shows Qwen3.5-122B running at 20–30 tok/s on three 3090s or 5090s, and 27B at 100 tok/s on a single 3090, offering better value than 35B-A3B. However, some Unsloth dynamic quantizations (e.g., 35B ud-q4_k_xl + MXFP4) show anomalous perplexity and coding performance; the community advises sticking to standard Q4_K_M quants for now.

Open models on real-repo coding benchmarks

APEX Testing evaluated 34 models on 70 real code repositories: Claude Opus 4.6 and GPT-5.2 Codex lead among closed models; among local models, quantized GLM-4.7 outperformed the Qwen3.5 series overall. The author stresses that agent framework choice can swing open-model scores by 50%+. (apexlab.org)

Self-hosted LLM leaderboard

Onyx published an S/A/B/C/D-graded self-hosted LLM leaderboard. Community members noted missing strong entries like Qwen 3.5 27B/122B and Qwen3-Next/Coder-Next.

Agents & Tooling

• Nous Hermes Agent (open source): multi-level memory, persistent machine access, file system/terminal/browser tool calls, and simultaneous Telegram/WhatsApp/Slack/Discord integration; also serves as a foundation for Atropos RL data generation. (GitHub)
• OpenClaw in production: one user automates rent reconciliation, repair coordination, and contract generation for property management; another rebuilt OpenClaw's scheduler with Mastra + Trigger.dev + Postgres to fix silent task failures, publishing a one-click deploy template.
• Cursor: Codex 5.3 Spark is now the default Cloud Agent with better speed and bug fixing; new PR Bugbot Autofix generates fixes from PR errors.
• LM Studio LM Link: securely connects remote machines via Tailscale so remote models behave like local ones, with end-to-end encryption and no port forwarding. (link.lmstudio.ai)
• Microsoft Copilot Tasks: generates a visible plan from a user request, then executes step by step with user review and mid-course intervention; currently in research preview.

Infrastructure & Hardware

• DeepSeek DualPath: places KV cache in decode-server memory and pushes it to prefill GPUs via GDRDMA, bypassing local PCIe bottlenecks for ~2x speedup on agent workloads.
• IterX (DeepReinforce): uses RL to auto-optimize FlashInfer MoE kernels, claiming 14.8x acceleration on B200.
• FlashInfer competition: a custom-DSL GDN decode kernel hit 2.56 microseconds—~760x faster than PyTorch eager and ~1.3x faster than the official Cute DSL kernel.
• GPU MODE community: chasing 4k×4k GEMM at ~90% of cuBLAS, uniform (once-per-warp) instructions via elect.sync, and warp-timeline profiling with tools like nanotrace.
• Cheap "scrap GPU" clusters: used mining cards (P104, CMP 100-210) running Qwen3.5 Q6 at 26 tok/s for ~$750 per node, with debate over PCIe 3.0/4.0 bottlenecks and riser splitters.

Research & Methods

• Solaris world model: argues world models should model shared global state rather than only pixel prediction; open-sources a multi-player Minecraft data engine, a multi-player DiT trained on 12.6M frames, and a VLM evaluation suite.
• AlphaEvolve (DeepMind): automatically mutates Python code to search for multi-agent RL algorithms, beating human-designed baselines on game tasks.
• World model survey: "Understanding World or Predicting Future?" (arXiv:2411.14499) surveys JEPA/V-JEPA, Dreamer, Genie, and Sora, debating whether world models are "mirrors" (reconstructing reality) or "maps" (serving decisions).
• Unsloth DPO tutorial: frames reward modeling as binary classification, eliminating the separate reward model, with a simplified end-to-end pipeline.
• Eleuther discussions: skepticism about Bezier Flow's practicality on ImageNet (32 parameters, 5 epochs), plus a neuron-pruning approach using "activation momentum" for sparser, more diverse neuron patterns.

Products & Applications

• Perplexity on Galaxy S26: built into Samsung's system with "Hey Plex" wake word, handling web search, research, and generative requests at the OS level, while Bixby handles on-device actions.
• Suno: 2M paying subscribers and ~$300M ARR, positioning AI music as a consumer creation-entertainment platform.
• Entropy Games: fully on-device AI NPCs using self-trained language and speech models for real-time dialogue and branching storylines.
• Tamarind Bio: raised $13.6M Series A for an open inference stack unifying molecular models for AI drug discovery.
• Gradio 6.7: improved custom HTML components with frontend-triggered backend functions, one-click Hub publishing, and faster Tabs/collapse components.
• ProposalMint: an AI-assisted web app helping ~50 Florida NGOs write grant applications.

Industry & Business

• Perplexity backlash: Pro users report quotas cut from 250/day to 20, and Deep Research API source counts reduced from 36 to 10 at ~$0.45 per call, pushing some users back to GPT/Claude.
• Anthropic vs. Pentagon: per Axios, the US Defense Secretary issued an ultimatum demanding removal of Claude's safety limits for mass domestic surveillance and fully autonomous weapons, threatening Defense Production Act pressure. Anthropic's CEO publicly refused on both counts, citing insufficient reliability. Reddit users also noted Anthropic quietly removed its flagship safety pledge, fueling debate; industry figures published open letters supporting clear "red lines."
• US closed vs. Chinese open models: compliance-constrained organizations struggle—US open models lag, leaving gpt-oss-120B as the main option; some suggest Mistral/Cohere or local retraining/audits of Chinese models.
• DeepSeek V4 & Huawei: reports that Huawei received early V4 access for hardware optimization before NVIDIA/AMD; commenters attribute this to engineering priorities rather than alliance politics.
• Block layoffs: Jack Dorsey announced cutting headcount from 10,000 to ~6,000 to become a "smaller, more AI-centric company"; shares rose 20%, though even AI teams were affected.

Policy, Governance & Safety

• The Anthropic-DoD clash highlights a core governance tension: capabilities are strong, but reliability and control lag; models still leak sensitive information under simple safety tests.
• Jailbreak communities agree no universal prompt works across all models and topics; CBRN, explosives, and CSAM categories remain hardest. AI red-teaming is emerging as a formal career path.
• Interpretability & model welfare: Goodfire AI demonstrated interpretability infrastructure scalable to trillion-parameter models with low inference overhead; Anthropic's "model retirement interviews" and Claude Opus 3's Substack reignited debate on anthropomorphizing models and model welfare research.

---

📌 Source: Easy AI Daily