Agent & Model Releases
Nous Arete-27B. Nous Research released Arete-27B, an SFT of Qwen3.8-27B that cuts Chinese state framing from 22% to 0.9% on their 803-prompt eval; Apache 2.0 weights and GGUF quants are available.
Kimi K3.1 quantized. An experimental 775GB quant of Kimi-K3.1-Pro-Coder runs on a single GPU with system RAM/NVMe via Strata and tiered MoE quantization, reaching 36.8 batch TPS and 96.7% on AIME/LiveCodeBench.
Terminal-RSI-27B agent. Terminal-RSI project released a Qwen3.8-based terminal-agent model for repository inspection, code editing, and multi-step tasks, using recursive self-rewrite and critic-inheriting policy refinement.
Drex Max decision model. Nace released multimodal decision model Drex Max, ranking #1 on Artificial Analysis and achieving Pareto frontier; weights are open with Open RAIL-M and v1.1 planned for local inference.
Tencent DSpine 3x. Tencent posted Qwen3.8-27B-DSpine-b16 on Hugging Face, a speculative decoding method that claims up to 3x faster decoding for 27B models.
Tools & Frameworks
Spec-first coding plugin. Google open-sourced Elephant-Goldfish Model plugin, which makes coding agents write a spec first and then has fresh subagents review it before code; thousands of Google engineers use it, though token usage is heavy.
Torchlens inspection tool. Torchlens is a GitHub repo for inspecting, visualizing, and intervening on any operation in any neural network, aiding interpretability.
C vs Rust SQLite. Simon Willison used Claude to compare C and Rust SQLite extensions for parse_number; Rust was initially faster due to strtod, but a custom C parser eventually matched or beat it, while the C extension was much smaller.
Taming PR descriptions. Simon Willison suggests hiding AI-written PR descriptions behind <details><summary> because they are overly comprehensive and read like slop; a human-approved intro should summarize the change.
Research Highlights
Open-HumanScale robotics. Researchers released Open-HumanScale, an open recipe and 50k-hour egocentric video corpus for pretraining robot policies; scaling human video yields log-linear gains, and hybrid human-robot training reaches 92.08% on RoboTwin 2.0 Clean2Clean.
Coding agents cleanup. Alibaba and partners built Co-Design Bench from real user change requests; the best model GPT-6 Astra fully completes only 26.4% of tasks, meaning human cleanup is needed on about three out of four tasks.
Parsimony code complexity. New Parsimony benchmark measures syntax units in patches; median model adds 1.9x the human fix on SWE-bench Verified, and harness choice alone can move code overhead from 0.80x to 0.99x.
MoE routing compared. Two guides compare routing strategies across open MoE models: expert counts range 8 to 384, top-k 1 to 8, and loss-free bias balancing is the dominant recipe; expert choice can leak future tokens in causal models.
Sutton on scaling limits. Richard Sutton argues raw compute alone won't produce intelligence; AI needs clever new methods, agents with world models that learn continuously, and he calls synthetic data a big mistake.
Industry & Business
Nvidia spy probe. Nvidia's board is launching an independent investigation into CEO Jensen Huang's relationship with Alice Huang, an AI executive reportedly recruited by China's MSS; Huang apologized for an affair in a memo, while the board says no evidence secrets were compromised.
Apple Siri agent push. Apple is accelerating Siri's roadmap, with fine-grained control over first-party apps expected in iOS 27.3 and third-party app integration later in spring, aiming to compete with OpenAI Dots and Meta Muse.
Suno music lawsuits. Elvis Presley and Eminem rights holders are suing Suno, alleging illegal voice and lyric cloning under Tennessee's ELVIS Act and common law.
Karpathy's weekend build. Andrej Karpathy used Claude Opus 5.5 and Claude Code to build HackerTube in a weekend, filtering 43 million HN comments into 6,454 curated videos with nearly all work automated.
Stanford AI-native course. Mihail Eric shares how Stanford CS146S adapted to AI coding agents by having students learn unfamiliar repos, propose changes, and explain patches to open-source maintainers; his Monaco team uses agent workspaces and on-call agents.
Local AI & Hardware
Intel Arc 32GB value. Intel Arc Pro B70 32GB at $1,299 delivers about half the generation speed and one-third the prefill of an RTX 5090 ($6,500+) on Qwen3.8-27B, making it a strong price/performance choice.
4x R9700 benchmarks. A ~$16K CAD build with four Radeon R9700 GPUs reaches 190 tok/s single-user, 6,800 tok/s prefill, and 80-225 tok/s per user for 8 users at 128K context; agent follow-ups start in 0.2-2.5s.
Assembly Qwen inference. Qwen3.5-0.8B runs in pure Assembly on one CPU thread at 6.51 tok/s, faster than PyTorch reference, with no libc or Python runtime.
Legacy GPU CUDA. Developer ported modern LLM training stack to legacy CUDA GPUs (SM 3.5-7.5) on Windows, enabling Mamba-1 on old NVIDIA cards with plans to backport Mamba-2/3.
Local bash helper. A fine-tuned Qwen3.5-0.8B converts plain English to bash offline in 1.7s on CPU; author shares hitting a 74% plateau and the lesson that training/test distribution mismatch matters.
Local model use cases. A user runs DeepSeek v4.1 Flash locally for scouts that offload web research from frontier models, masons that build code from frontier specs, and content-filtering agents.
Local motion graphics. Qwen3.8 Flash Next IQ3_XXS on 16GB VRAM produced a motion graphics video via Three.js/Playwright/ffmpeg in about 2 hours, showing local models can handle creative coding tasks.
That's everything for this day – about a 5-minute read.