AI · Weekly · February 28, 2026

the distillation war

Anthropic says DeepSeek, MiniMax, and Moonshot spun up over 24,000 fake Claude accounts to siphon 16M chat exchanges for training, branding it 'industrial-scale distillation attacks.'[3] Those same Chinese labs are landing near‑frontier scores anyway—MiniMax M2.5 hits 80.2% on SWE‑bench, and GLM‑5 scores 81.8 on Extended NYT Connections and 77.8 on SWE‑bench Verified.[4]

DeepSeek reportedly trained on Nvidia’s top chips despite a U.S. ban and then gave early access to Huawei, turning export controls into a catalyst for domestic acceleration.[5]

Meanwhile, Qwen 3.5’s 400B‑parameter multimodal architecture and top‑of‑Hugging‑Face performance show that open‑weight Chinese stacks are no longer 'good enough' copies but genuine peers.[6]

Inside the labs, distillation from API outputs is framed as standard practice and accusations of theft as selective outrage, but outside, regulators and incumbents are already treating this as IP exfiltration and a national‑security problem.[7]

local-first stacks quietly catch up

Qwen3.5‑35B‑A3B GGUF variants hit 99.9% KL divergence across 9TB of GGUFs and run with roughly 22–32GB of RAM, while supporting context windows past 1M tokens on 32GB VRAM.[8]

On dual 3090s that same model processes prompts at ~2K tokens/sec and generates around 90 tokens/sec, giving local setups throughput that used to require mid‑tier cloud APIs.[9]

Llama 3.1 70B now runs on a single RTX 3090 via NVMe‑to‑GPU, Llama 3.2 1B hits 4.4 tok/sec on an AMD NPU, and Mistral 24B stays usable on a 16GB 5060 Ti.[10]

At the edge, TranslateGemma 4B translates 55 languages fully in‑browser via WebGPU, and LFM2.5‑1.2B‑Thinking pushes 200+ tokens/sec in the same environment.[11]

The flip side is brittleness: Qwen 3.5 122B is reported to hallucinate heavily, several Qwen3.5 quantizations are 'all broken,' LM Studio users see sluggish KV‑cache behavior, and vLLM has compatibility gaps with some Qwen variants.[12]

coding is solved, debugging is not

Claude Code already accounts for about 4% of public GitHub commits, with projections north of 20% by 2026, while some engineers report going through 2026 writing '0 lines of manually-written code.'[13] Codex 5.3 now beats Opus 4.6 on agentic coding benchmarks, and Andrej Karpathy says programming has changed more in the last two months than in years because of coding agents.[14]

Yet the hard data say the mess moved, not vanished: debugging AI‑generated code takes roughly 3x longer, AI‑driven production incidents average $40k each, and accumulated refactor costs per system can exceed $200k.[15]

Vibe‑coded apps have already leaked data from 18,000 users, 59% of developers admit shipping AI code they don’t fully understand, and Microsoft executives openly worry about wiping out entry‑level coding roles.[16]

Developers consistently report that AI assistants create denser, less readable code and lengthier debug sessions, so the new bottleneck is reasoning about what to build and how to untangle what the agents have generated.[17]

mcp + clis: the agent runtime solidifies

The Model Context Protocol is quietly becoming the default wiring layer: France now runs a national MCP server for all government data, and MCP standardizes tool access across LLM agents.[18]

OpenBrowser MCP is 3.2x more token‑efficient than Playwright MCP and 6x more than Chrome DevTools MCP, and auto‑generated CLIs from MCP servers can slash token use by 94%, so serious agent builders are converging on CLI‑first patterns.[19]

Zero‑copy vision transports read raw GPU frame buffers via shared memory instead of DOM scraping, TOON proxy shrinks JSON overhead by about 40%, and Memento/Sentry MCP servers add long‑term memory and automated on‑call triage.[20]

But the security surface is exploding: MCPwner found multiple 0‑days in OpenClaw, OpenClaw itself ships with 2,000+ known vulnerabilities including 10 critical ones, and 80% of AI agent repos show exploitable security issues.[21]

Latency from remote MCP servers and fragile tool schemas are already visible pain points, so teams chasing rich multi‑agent graphs are trading raw model tokens for orchestration complexity and new failure modes.[22]

video and images hit commodity speed, not commodity trust

Nano Banana 2 delivers pro‑grade images at about 4x the speed and roughly half the price of Nano Banana Pro—around $67 per 1,000 images—while supporting real‑world‑accurate renders and multilingual text.[23]

It’s now effectively uncensored for named people, and Google says journalists have used its SynthID watermarking more than 20M times for image verification, pushing identity and provenance questions into everyday workflows.[24]

On the video side, Seedance 2.0 can turn arbitrary media—including a child’s drawing—into cinematic clips or even a one‑shot 'film' from inside CapCut desktop, priced via $0.01 credits and editable immediately after generation.[25]

Yet users complain the feature often isn’t available despite the marketing, find the credit pricing aggressive for the actual output, and are uneasy about ownership of Seedance‑generated content, all while its global rollout is stalled under Hollywood copyright threats.[26]

Meanwhile Grok Imagine tops Arena.AI’s Image‑to‑Video leaderboard, WAN 2.2 plus LTX‑2 can upscale to 4K with 4x frame interpolation, but they demand 64GB‑class RAM, long render times, and steep ComfyUI‑style learning curves.[27]