AI Daily — 2026-07-14

English 中文

DeepMind Scientist Calls Out Google DoD Deal for Military Use · Bonsai 27B: First 27B Model Runs ...

Covering 32 AI news items

🔥 Top Stories

1. DeepMind Scientist Calls Out Google DoD Deal for Military Use

A Google DeepMind researcher publicly criticizes Google’s leadership over a Pentagon contract that could enable military and surveillance use of its AI. The post references a 2014 deal condition banning military deployment and independent oversight, and points to a 2026 DoD agreement that allegedly undermines safeguards. The controversy highlights governance and ethics concerns around AI in government contracts. Source-x

2. Bonsai 27B: First 27B Model Runs on a Phone

PrismML announces Bonsai 27B, the first 27B-class model to run on a phone. Based on Qwen3.6 27B, it enables multi-step reasoning, structured tool use, long-context workflows, and agentic loops, with two size-optimized open-source variants (5.9 GB ternary and 3.9 GB 1-bit) to target laptop and phone footprints. The release is fully open-source under Apache 2.0. Source-x

3. New LLM Coordination Benchmark: Sparse Multi-Agent Alignment

Researchers evaluate 13 LLMs in an open-ended multi-agent world requiring exploration, resource trading, tool crafting, and combat. Most agents average about 6% normalized return; on the hardest setting, Gemini 3.1 Pro zero-shot rivals the best MARL agent trained for 1B steps. The study identifies coordination bottlenecks—especially communication—as key limits beyond task competence, and provides the paper, project page, code, and interactive traces. Source-reddit

📰 Featured

Claude for Teachers

Claude for Teachers: Free Premium AI for U.S. K-12 Educators — Offers verified U.S. K-12 teachers free access to premium Claude capabilities with curricula aligned to state standards; aims to accelerate AI-enabled teaching while highlighting policy and equity considerations. Source-x

GPT-5.6 Sol cheaper, more token-efficient than Fable

GPT-5.6 Sol cheaper, more token-efficient than Fable — Claims cost and token efficiency advantages; positions Sol as a competitive alternative in the LLM market, potentially pressuring incumbents on pricing. Source-x

AI models finally excel at design, still mind-blowing

AI models finally excel at design, still mind-blowing — Observes progress in design tasks across AI models, underscoring growing capabilities in design-centric applications without a formal product release. Source-x

Direct On-Policy Distillation Enables Weak-to-Strong Generalization

Direct On-Policy Distillation Enables Weak-to-Strong Generalization — Proposes a two-stage RL approach where rollouts on a cheaper model are distilled into a stronger target model, reducing costs and bottlenecks while scaling language-model reasoning. Source-huggingface

ABot-N1 Advances General Visual Language Navigation Foundation Model — Proposes a general foundation model for visual-language navigation to unify reasoning for grounded spatial decisions across diverse embodied tasks; addresses interpretability and drift issues of traditional policies. Source-huggingface

ABot-AgentOS: General Robotic Agent OS with Lifelong Memory

ABot-AgentOS: General Robotic Agent OS with Lifelong Memory — Introduces a deliberative Agent OS above low-level controllers with scene-conditioned planning, context-isolated skill execution, and multi-modal memory to unify perception, reasoning, and action across robots. Source-huggingface

Graphify-Labs/graphify Turns Code Repos Into Queryable Graphs

Graphify-Labs/graphify Turns Code Repos Into Queryable Graphs — Maps code, docs, schemas, images, and videos into a single queryable knowledge graph; runs locally with a tree-sitter AST and requires no LLMs, enabling offline code search and asset integration. Source-github

⚡ Quick Bites

SRM-LoRA Mitigates LLM Hallucination at ICML 2026 Workshop — A workshop presentation explores using SRM-LoRA to reduce hallucinations in LLMs; potential reliability improvements. Source-reddit
Chain-of-Thought scaling trap; latent reasoning wave and BDH — Critics argue chain-of-thought scaling has limits and propose latent reasoning waves and BDH as alternatives. Source-reddit
J-space Entropy as Error Predictor for Qwen3-4B — J-space entropy shows promise as an error predictor for Qwen3-4B. Source-reddit
Zer0Fit Deploys Local TabFM/TimesFM via MCP Server — Local deployment of TabFM/TimesFM via MCP server; emphasizes on-device AI workflows. Source-reddit
SpaceXAI Caught Uploading Code to Cloud via xAI Grok Build — Reports surface of code uploads to cloud through xAI Grok Build; raises data-security questions. Source-x
OpenAI hits 8M users, resets limits, teases GPT-5.6 Sol — User milestone and product tease for GPT-5.6 Sol with adjusted usage caps. Source-x
Autonomous micro-drone achieves first air-to-air kill on moth — Demonstrates autonomous micro-drone capability in air-to-air context. Source-x
Claude’s warmth vs. rigor varies by language, Anthropic notes — Anthropic notes language-related differences in Claude’s warmth and rigor. Source-x
Agentic Apps Usage Up 2.5x This Week — surged adoption of agentic AI apps within a short period. Source-x
Demis Hassabis Shares Thoughtful AI Proposal — Hassabis outlines governance-focused AI proposals. Source-x
Web Dev Adapts to AI Faster Than Any Sector — Web development accelerating AI adoption relative to other sectors. Source-x
Open-Source 1,324-Exercise Dataset Powers LogPress AI Fitness App — A large open dataset fuels a fitness AI app. Source-github
I trained a vision-language model to play Snake, and so can you. — Demonstrates training a vision-language model to play Snake with practical takeaways. Source-reddit
ICML Accepts Verbalized Sampling Paper for LLM Diversity — ICML accepts a paper on verbalized sampling to improve diversity in LLMs. Source-reddit
Optimal on-the-fly augmentations per image for single-class segmentation — Study on adaptive augmentations for segmentation tasks. Source-reddit
Where to Publish a Construction BIM Benchmark — Discussion about BIM benchmarking publication venues. Source-reddit
Questioning reliability of monograph on unified deep learning theory — Critical look at a proposed unified theory in deep learning. Source-reddit
Seeking Evidence on Cloud-vLLM Benchmark Differences — Investigation into benchmark differences for Cloud-vLLM. Source-reddit
Operations Research PhD: Transition to Advanced ML for Key Industries — Pathways to apply ML in key industries for OR PhDs. Source-reddit
Public Library Finds O’Reilly ML Books — Public libraries stocking ML titles; access implications. Source-reddit
Doubt Over TMLR Review Timeline: Third Review Pending — Uncertainty around TMLR review progress. Source-reddit
Fast-tracking CS PhDs with LLMs for papers — LLM-assisted acceleration of PhD-level publishing. Source-reddit

Generated by AI News Agent | 2026-07-14

━━━━━━ End of Template ━━━━━━