AI Daily — 2026-03-20

English 中文

Anthropic hires editor to lead economics and policy · Kimi-k2.5 Powers Composer-2 Frontier with R...

Covering 24 AI news items

🔥 Top Stories

1. Anthropic hires editor to lead economics and policy

An author announces they are joining Anthropic’s editorial team to lead economics and policy, collaborating with the Anthropic Institute. They emphasize the need for high-quality information as AI progress accelerates and note ongoing roles with the Institute for Foreign Policy (IFP) and Recoding America, starting in a few weeks. The move marks a transition from a frontier think tank to a frontier lab, described as bittersweet after three formative years at IFP. Source-twitter

2. Kimi-k2.5 Powers Composer-2 Frontier with RL

Kimi-k2.5 is reported as the strongest base model in perplexity-based evaluations. The team then conducts continued pre-training and high-compute RL (a 4x scale-up) to reach frontier-level performance for Composer-2, leveraging Fireworks’ inference and RL samplers. Cursor AI notes the collaboration and confirms access to Kimi-k2.5 via FireworksAI HQ. Source-twitter

3. Karpathy on AI phase shift, AutoResearch, and model landscape

Karpathy discusses the ongoing phase shift in AI-enabled engineering, AI psychosis, and the potential for AutoResearch. The conversation covers the model landscape, second-order effects, and a SETI-at-Home-like collaboration in AI, plus skills and collaboration surfaces for humans and AI. Source-twitter

📰 Featured

Open Source

OpenDataLoader PDF: AI-Ready PDF Parser with OCR — OpenDataLoader PDF is an open-source PDF parser that exports Markdown, JSON (with bounding boxes), and HTML for AI-ready data. It includes built-in OCR (80+ languages) and a deterministic local mode plus an AI hybrid mode for complex pages, achieving top benchmarks (0.90 overall, 0.93 table accuracy on 200 real-world PDFs). It’s designed for RAG pipelines and is installable with a three-line pip command. Source-github
Local AI renders interactive charts and forms offline — Anthropic’s Claude showcased interactive artifacts, and a developer released Inline Visualizer to bring similar capabilities to any model with tool calling. The BSD-3 licensed tool renders inline HTML/SVG visualizations within chat, wrapped in a themed shell with dark mode, and includes a bidirectional JavaScript bridge for model-driven interactions, all offline and without external services. Source-reddit

AI

Qwen3 30B Runs at 7–8 t/s on Raspberry Pi 5 8GB — An edge AI demo shows Qwen3-30B-A3B-Instruct-2507-GGUF running locally on Raspberry Pi 5 (8GB) with SSD, achieving 7–8 t/s at a 16,384-token context. With ByteShape quantization (2.66 BPW) and a 4-bit variant delivering 4–5 t/s, the setup is packaged as Potato OS, a flashable headless Debian image that auto-downloads Qwen3.5 2B with vision encoder after boot. All processing is local (no API calls, no eGPU), illustrating strong edge-inference progress. Source-reddit
BioReason-Pro: First reasoning model for protein function — BioReason-Pro is introduced as the first reasoning model for protein function, signaling AI-driven analysis of proteins. The announcement appears as a Twitter/X thread outlining the model’s capabilities and potential impact on protein biology. This represents a notable AI biology development at the intersection of AI and life sciences. Source-twitter
TinyLlama 1.1B Runs Locally on a 2002 PowerBook G4 — MacinAI Local is a native local AI inference platform for classic Macintosh hardware that operates without internet access. It uses a ground-up C89 engine, supports multiple models (GPT-2 124M, TinyLlama, Qwen 0.5B, SmolLM) and any HuggingFace/LLaMA-architecture model via a Python export script, and includes a Macintosh-trained 100M parameter transformer with AltiVec optimization delivering 7.3x speedups on PowerPC. The project demonstrates offline AI on Mac OS 9 from a CD. Source-reddit

AI Policy

White House Unveils National AI Policy Framework — The White House released a National AI Policy Framework aimed at ensuring all Americans benefit from AI. It champions a single federal AI policy rather than a patchwork of state rules and signals readiness to work with Congress on legislation. Source-twitter

LLM

Mistral Small 4 Opens Weights, Scores 27 on AI Index — Mistral released Small 4, an open-weights model with a 119B mixture-of-experts architecture that supports both reasoning and non-reasoning modes. In reasoning mode, it scores 27 on the Artificial Analysis Intelligence Index, a 12-point gain from Small 3.2 and placing it among Mistral’s more capable models, yet it still trails open-weights peers with similar total parameters such as gpt-oss-120B, NVIDIA Nemotron 3 Super 120B A12B, and Qwen3.5 122B A10B. Source-twitter
Generation Models Use Implicit 3D Priors for Scene Understanding — Multimodal LLMs show strong semantic abilities but struggle with spatial and geometric reasoning. Existing methods rely on explicit 3D modalities, facing data scarcity and generalization challenges. The work proposes leveraging implicit spatial priors embedded in large-scale video generation models to enhance scene understanding. Source-huggingface
GLM-5.1 Will Be Open Source — According to a post on X (Twitter), GLM-5.1 is slated to be released as open source. The announcement emphasizes openness with no immediate details on licensing or timeline provided. Source-twitter
Merged Qwen 3.5-35B with Claude Opus 4.6 for Small GPUs — A community-driven merge combines Qwen 3.5-35B-A3B-Uncensored-Claude-Opus-4.6-Affine with HauhauCS’s and Jackrong’s variants to add enhanced reasoning capabilities. The workflow included a KL-divergence cleanup and integrated “thinking skills” from Jackrong’s model. It was executed in Google Colab Free Tier without unpacking the model (IQ4_XS format), targeting low-end GPUs like the RTX 3060 12GB. Source-reddit

Multimodal

SAMA: Factorized Semantic Anchoring and Motion Alignment for Instruction-Guided Video Editing — Current instruction-guided video editing models struggle to modify content precisely without disturbing motion, often relying on external priors like vision-language features. The paper introduces SAMA (Factorized Semantic Anchoring and Motion Alignment), a framework that decomposes editing into semantic anchoring and motion alignment to improve robustness and generalization without heavy priors. This approach aims to balance semantic precision with faithful motion preservation in video edits. Source-huggingface
FASTER: Rethinking Real-Time Flow VLAs — Real-time execution is essential for deploying Vision-Language-Action models in the physical world. The work critiques asynchronous inference for prioritizing trajectory smoothness while overlooking latency in reacting to environmental changes, proposing a redefinition of reaction within action chunking. It offers a systematic analysis of factors shaping reaction time and reports that reaction time follows a uniform distribution driven by timing parameters. Source-huggingface

AI Tools

gsd-build launches Get Shit Done meta-prompting tool for Claude Code — gsd-build releases ‘Get Shit Done’, a lightweight meta-prompting, context engineering, and spec-driven development system. It supports Claude Code, OpenCode, Gemini CLI, Codex, Copilot, and Antigravity, and helps mitigate context rot by preserving relevant context within the model window. The project is cross-platform via npx and open-source on GitHub, with adoption cited by engineers at Amazon, Google, Shopify, and Webflow. Source-github

⚡ Quick Bites

Codex for Students: $100 Credits in US and Canada — OpenAI’s Codex for Students gives U.S. and Canadian college students $100 in Codex credits. The initiative aims to boost hands-on learning by building, breaking, and fixing with AI-powered coding tools. It promotes practical coding education through access to Codex credits. Source-twitter
Gemini Android now pauses; iOS version arriving soon — Gemini’s Android app update lets users pause mid-speech without being cut off by the assistant. This fixes a behavior where the AI would jump in too quickly, improving speaking flow for users. An iOS version is expected to roll out in a few weeks. Source-twitter
T3 Code Adds Claude Support for AI Coding — T3 Code now supports Claude. If you have the Claude Code CLI installed and signed in, you can use Claude directly within T3 Code. The post closes with a tongue-in-cheek note about potential legal pushback. Source-twitter
Interactive 3D/2D GPT-2 Activation Visualization — A new online visualization at llm-visualized.com animates real activations and attention scores from GPT-2 Small (124M) during a forward pass. Built with Three.js for 3D and plain HTML/CSS/JS for 2D, it aims to help users learn how LLMs work by exposing internal computations, with developer feedback invited. Source-reddit
Minimax 2.7 to Ship Closed Weights — A Reddit post claims Minimax 2.7 will use closed weights, diverging from open-sourced policy, impacting the LocalLLaMA community. The post notes no official confirmation yet. If true, this marks a notable shift in access to AI weights within the open-source ecosystem. Source-reddit
Efficient Reasoning in LRMs via Balanced Thinking — Large Reasoning Models show strong reasoning but risk overthinking (unnecessary steps) and underthinking (insufficient exploration), causing inefficiency and errors. This limits deployment in resource-constrained settings. The piece surveys mitigations like suppressing reflective keywords or adjusting reasoning, and advocates a balanced thinking approach to improve efficiency and reliability. Source-huggingface
Trained 0.8M Model for Business Email Generation — Someone trained a 0.8-million-parameter model for business email generation using a different architecture, following a prior ‘28m model email experiment’ post. They tested it with a prompt to write a polite refusal email, and the provided example shows a templated email containing placeholders and some odd phrasing. Source-reddit
Openclaw: What are its practical use cases? — A Reddit post questions the hype around Openclaw, asking for concrete use cases and practical value. The author expresses skepticism about why it is useful and seeks explanations on its applications, in a thread on /r/LocalLLaMA. Source-reddit

Generated by AI News Agent | 2026-03-20