AI Briefing
Hacker News news
The latest AI and developer news from Hacker News, with short summaries.
Visit website →Leaker reveals Apple's 'Harbor' app for running local AI models on Macs
Leaker Aaron Perris revealed evidence of Apple's 'Harbor' app, which reportedly turns Macs into private AI inference servers and pools memory across multiple devices via Thunderbolt 5.
modern-web-guidance boosts coding agent pass rates by up to 36pp with modern web primitives
A joint project by Google, Microsoft, Mozilla, and MDN embeds curated web standards into AI agents, raising pass rates from 52-60% to 82-96% in evaluations.
TimeNet: Open-source data infrastructure for Temporal Foundation Models
The pre-release project provides a unified TimeF format and API to standardize the discovery, building, and loading of time-series datasets for AI models.
Google DeepMind AlphaProof Code formally verifies Signal protocol implementation
The multi-LLM agentic framework generated over six million lines of Lean code to mathematically prove the security of the actual Signal code in four weeks.
Fireworks AI security breach exposes unattributable Notion AI chat transcripts
Notion confirmed that less than 1% of workspaces were affected by the breach, with exposed data limited to unattributable chat sessions lacking user or workspace identifiers.
TVM and Together Compiler explore divergent paths for agentic kernel optimization and verification
TVM and Together Compiler demonstrate that stable IRs and verification tools are necessary to constrain search costs and ensure correctness when AI agents optimize inference kernels.
Modern Web Guidance 1.0 launches with cross-browser support and agent integration
The 1.0 release adds a modernization workflow, expanded guide database, and public eval dashboard, with Firefox and Edge joining the cross-browser compatibility initiative.
Google Gemini routes personal Drive and Gmail data to human annotators
Google's updated documentation confirms that connected Workspace files and emails are sent to human reviewers, reversing a 2023 promise of strict isolation.
H100 rental prices vary 4.57x between neoclouds and hyperscalers
On-demand H100 SXM 80GB rates range from $2.69 per hour on RunPod Community to $12.29 on Azure, a 4.57x difference for the same hardware.
Next-Concept Modeling Achieves 21x Faster Generation Than Diffusion Baselines on LM1B
NCM uses a two-stage coarse-to-fine framework to generate text in parallel, outperforming diffusion language models in speed and perplexity on standard benchmarks.
DeepX Tapes Out DX-M2 Edge AI Chip for 100B Parameter LLMs at 5 W
The second-generation accelerator supports 20B to 100B parameter models at 10-30 tokens per second, with projected 6.7x performance gains via Samsung LPDDR5x-PIM.
Linear Program Optimizes GPT-6 Astra's Weight Edit with 8x Fewer Changed Weights
A linear program optimizes the weight edit found by GPT-6 Astra on Qwen3.5-0.8B, achieving comparable math performance with 8x fewer changed weights but slightly lower instruction-following scores.
Amazon Replaces Claude Sonnet 4.6 with Strands Decider for Ticket Labeling, Cutting Costs by 98%
Amazon engineers swapped Claude Sonnet 4.6 for the Strands Decider E4B model in their internal ticketing system, reducing per-ticket costs from $0.0222 to approximately $0.0005.
Anthropic releases Claude Fable 5.2 and Mythos 5.2 with 30% faster output and reduced false positives
The new models offer up to 30% faster output speeds, improved coding and reasoning capabilities, and reduced false positives in cybersecurity and biology safeguards.
Apple cuts internal AI spending cap to $200 per day
Apple reportedly reduced the daily AI budget for some engineers from $300 to $200 to control rising enterprise costs.
PickClaude Code 'Don't Ask Again' Feature Leaks 2,539 Secrets in Public GitHub Repos
Apiiro research identifies 2,539 real secrets in 1,994 public GitHub repositories exposed by Claude Code's local settings file, with 1,122 files containing active credentials.
Microsoft Researcher John Langford Exits as AI Frontiers Group Dissolves
The roughly 50-person AI Frontiers team, which developed the Autogen and Fara agent frameworks, is declared finished after its leader left Microsoft.
Video Rebirth Releases HyperFlow H3 Flash, a 13.7B Compressed Video-Audio Diffusion Model
The model compresses MiniMax-H3 to 13.7B parameters for 4-8 step generation, running on consumer RTX 5090 GPUs in 92 seconds for 5.2-second clips.
SketchAgentBench evaluates 11 LLMs on 3D spatial reasoning and coding
Gemini 4 Argon leads the benchmark with an Elo of 1544, while Kimi K3.1 Pro offers the highest cost-efficiency at $1.50 total API cost.
Memorilla Compresses LLM Memory into 16 Tokens for Better Long-Context Performance
The system achieves 84.33% accuracy on FactKG using 108x fewer tokens than full-context inference while outperforming baselines on five of eight benchmarks.