Easy AI Daily | February 27, 2026
A translated digest of the Easy AI daily news roundup covering models, agents, infrastructure, research, products, industry moves, 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 alternatives like GPT Image 1.5 and FLUX.2. It supports multi-character consistency (up to 5 people plus objects), 4K upscaling, long-text and multilingual typography, and real-time web image search, integrated across the Gemini app, Search AI Mode, Ads, and API/Vertex.Links: Google blog | DeepMind announcement | Arena leaderboard | ArtificialAnalysis pricing
Nano Banana 2 vs Pro
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 downsides: slower API/Vertex speeds, occasional errors, and weak transparent PNG support.Links: Reddit impressions | Pricing discussion
Perplexity open-sources pplx-embed
Perplexity released the pplx-embed and pplx-embed-context families (0.6B and 4B parameters, MIT license) for retrieval/RAG, claiming industry-leading results on internal benchmarks built from 115M real queries and 30M documents. Available on HuggingFace and its API.Links: Announcement | Benchmarks
Qwen3.5 local quantization: performance and pitfalls
Reddit tests report Qwen3.5-122B running at 20–30 tok/s on three 3090/5090 cards, and 27B at 100 tok/s on a single 3090, with better value than 35B-A3B. Some Unsloth dynamic quants (e.g., 35B ud-q4_k_xl + MXFP4) show anomalous perplexity and coding results; the community advises sticking to standard Q4_K_M quants and avoiding MXFP4 post-quantizations for now.Links: Quant comparison | 122B multi-GPU | 27B vs 35B-A3B
Open models on real-repo coding benchmarks
APEX Testing evaluated 34 models across 70 real code repositories: Claude Opus 4.6 and GPT-5.2 Codex lead among closed models; quantized GLM-4.7 beats the Qwen3.5 family among local options. Notably, agent framework choice can swing open-model scores by 50%+.Links: Original post | APEX site
Self-hosted LLM leaderboard
Onyx published a self-hosted LLM leaderboard graded S/A/B/C/D. Community feedback notes missing entries like Qwen 3.5 27B/122B and Qwen3-Next/Coder-Next.Links: Leaderboard | Discussion
Agents & Tooling
Nous open-sources Hermes Agent
An open-source general agent with multi-level memory, persistent machine access, filesystem/terminal/browser tool use, and native Telegram, WhatsApp, Slack, and Discord integration. It also serves as a base for Atropos large-scale RL data generation. The first 750 registrants get a free month of Portal.Links: Product page | GitHub
OpenClaw in real businesses
Users report running OpenClaw for property management (rent reconciliation, repair coordination, contracts), while others rewrote its scheduling on Mastra + Trigger.dev + Postgres to fix silent task failures and gateway crashes.Links: Rewrite write-up
Cursor adds Codex 5.3 and Bugbot Autofix
Codex 5.3 Spark is now the default Cloud Agent, with community reports of faster speeds and better bug fixing. Bugbot Autofix generates fixes from PR errors automatically.Links: Cloud Agents docs | Bugbot Autofix
LM Studio launches LM Link
LM Link connects remote machines securely via Tailscale so remote models behave like local ones, with end-to-end encryption and no port forwarding. Some quip it's "Tailscale with a UI on top."Links: LM Link | Tailscale blog
Microsoft Copilot Tasks
Copilot Tasks turns requests into visible, step-by-step plans users can review and interrupt, currently in research preview.Links: Official blog
Infrastructure & Hardware
DeepSeek DualPath and IterX: inference as systems engineering
DeepSeek's DualPath paper 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. DeepReinforce's IterX uses RL to optimize FlashInfer MoE kernels, claiming 14.8x acceleration on B200.Links: DualPath commentary | IterX | IterX tutorials
FlashInfer competition: microsecond-level kernels
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.Links: Competition code
GPU MODE community
Work on 4k×4k GEMM approaching 90% of cuBLAS, uniform (elect.sync) warp instructions, and nanotrace-based warp timeline profiling.Links: nanotrace | GTC uniform-instruction talk
Cheap mining-GPU clusters
Hobbyists built clusters from used P104/CMP 100-210 cards running Qwen3.5 Q6 quant at 26 tok/s for ~$750 per node, debating PCIe 3.0/4.0 bottlenecks and riser splitters.Links: Example PCIe splitter
Research & Methods
Solaris: Minecraft world models with shared global state
Solaris argues world models should prioritize modeling shared global state over pure pixel prediction. Released: a multi-player Minecraft data collection engine, a multi-player DiT trained on 12.6M frames, and a VLM evaluation suite for multi-agent consistency.Links: Paper intro
AlphaEvolve: evolving multi-agent RL algorithms
DeepMind's AlphaEvolve mutates Python code to search for multi-agent RL algorithms, with new algorithms beating human-designed baselines on game tasks.Links: Overview
World models survey: "mirror vs map"
The survey Understanding World or Predicting Future? reviews JEPA/V-JEPA, Dreamer, Genie, and Sora, debating whether world models should reconstruct reality or merely serve decision-making.Unsloth DPO tutorial
A practical breakdown of Direct Preference Optimization as a binary classification problem, avoiding a separate reward model.Links: Blog post
Eleuther discussions
Skepticism about Bezier Flow's practicality on ImageNet (32 parameters, 5 epochs), plus neuron-pruning work using "activation momentum" for sparser, more diverse neuron patterns.Links: Neuron pruning paper
Products & Applications
Perplexity becomes Galaxy S26 system assistant "Hey Plex"
Deep Samsung integration puts Perplexity at OS level for search, research, and generation on Galaxy S26, with Bixby handling on-device actions.Links: Announcement
Suno: $300M ARR
Suno reports 2M paying subscribers and ~$300M annual recurring revenue, positioning AI music as a consumer creation-entertainment platform rather than a demo.Links: Two-year data
Entropy Games: on-device AI NPCs
Fully local language and voice models generate NPC dialogue and branching stories in real time on player hardware, emphasizing privacy and low latency.Links: Technical report
Tamarind Bio raises $13.6M
Series A for a unified open inference layer for molecular models serving pharma and labs.Links: Announcement
Gradio 6.7
Improved custom HTML components that can trigger backend functions directly, plus performance gains for Tabs/collapsible components.Links: Release notes
ProposalMint
An AI-assisted grant-writing web app for ~50 Florida NGOs — a showcase of niche vertical AI tools.Links: Site
Industry & Company News
Perplexity under fire from power users
Pro users report daily quotas cut from 250 to 20, and Deep Research API sources reduced from 36 to 10 at ~$0.45 per call, pushing some back to GPT/Claude.Anthropic vs US Department of Defense
Per Axios, the Defense Secretary issued an ultimatum demanding Claude's safety restrictions be removed for mass domestic surveillance and fully autonomous weapons, threatening Defense Production Act and supply-chain measures. Anthropic's CEO publicly refused on both points, citing insufficient reliability. Reddit users noted Anthropic also quietly removed a flagship safety pledge, drawing mixed reactions; industry peers published open letters of support.Links: Axios report | Anthropic statement | Safety pledge discussion
US closed models vs Chinese open models
Compliance-constrained organizations struggle: they avoid Chinese models on national-security grounds, but US open-weight options lag — with gpt-oss-120B a rare candidate. Some suggest Mistral, Cohere, or local fine-tuning/auditing of Chinese models.Links: Original thread
DeepSeek gives Huawei early V4 access
Reports that Huawei received early DeepSeek V4 access for hardware optimization sparked alliance speculation, though commenters see it as engineering priority rather than geopolitics.Links: Discussion
Block cuts 40% of staff
Jack Dorsey announced shrinking headcount from 10,000 to ~6,000 to become "smaller but more AI-centric"; the stock rose 20%, though internal AI teams were also affected.Links: Jack's post
Policy, Governance & Safety
The Anthropic incident as a governance flashpoint
The conflict reflects a core tension: capabilities are strong, but reliability and control mechanisms lag. Related research shows models still leak sensitive information under simple safety tests.Links: Reliability concerns
Jailbreak and red-team community
Practitioners agree no universal jailbreak prompt exists across models and topics; CBRN and CSAM categories remain especially resistant. AI red teaming is increasingly discussed as a career path.Interpretability and model welfare
Goodfire AI announced interpretability infrastructure that scales to trillion-parameter models with low inference overhead. Anthropic's "model retirement interviews" and Claude Opus 3's Substack sparked debate about anthropomorphizing models and model welfare research.Links: Goodfire | Claude's Corner | Model welfare paper
---
📌 Source: Easy AI Daily