We found 24 Android vulnerabilities using our open source AI security agent
Researchers used an open-source AI security agent to identify 24 vulnerabilities in Android applications. ·
6 points and 2 comments on Hacker News, linking github.blog.
Top now
6 storiesWhat is climbing fastest across every source
OpenAI6.0xWorld Labs7.7xAnthropic6.9xClaude Sonnet 5.57.0xNvidia7.0xOpus 5.57.0xAMD7.7xLearning6.9xLLM4.7xAustralia5.6x
OpenAI Says It Will Not Release Newest A.I. Model Over Safety Concerns
33 points and 31 comments on Hacker News, linking nytimes.com.
EmbodiedMemory-Bench: Benchmarking Embodied Memory for Long-Horizon Embodied Tasks
Researchers introduced EmbodiedMemory-Bench, a dataset of 2,554 interactive episodes, to evaluate how agents retain and update environmental memory during long-horizon tasks. ·
AdaTutoRank: Learning to Rerank Document Sets via Adaptive Tutoring Optimization for RAG and Deep Research
Researchers introduced AdaTutoRank, a reranker for RAG systems that uses adaptive tutoring optimization to improve document set composition. ·
KKKKhazix/AIHOT
A website framework automatically identifies trending topics and generates daily reports. Users can customize its sources and selection criteria to create industry-specific news sites. ·
akatz-ai/MiniMax-H3-Character-Swap-LoRA
A video to video model from akatz-ai with 128 likes and 3,436 downloads on Hugging Face.
Claude Sonnet 5.5
191 points and 138 comments on Hacker News, linking anthropic.com.
KKKKhazix/AIHOT
A website framework automatically identifies trending topics and generates daily reports. Users can customize its sources and selection criteria to create industry-specific news sites. ·
OpenAI Says It Will Not Release Newest A.I. Model Over Safety Concerns
33 points and 31 comments on Hacker News, linking nytimes.com.
EmbodiedMemory-Bench: Benchmarking Embodied Memory for Long-Horizon Embodied Tasks
Researchers introduced EmbodiedMemory-Bench, a dataset of 2,554 interactive episodes, to evaluate how agents retain and update environmental memory during long-horizon tasks. ·
AdaTutoRank: Learning to Rerank Document Sets via Adaptive Tutoring Optimization for RAG and Deep Research
Researchers introduced AdaTutoRank, a reranker for RAG systems that uses adaptive tutoring optimization to improve document set composition. ·
anthropics/claude-code v2.1.284
Anthropic released Claude Code v2.1.284, making Claude Sonnet 5.5 the default model with a 1M token context window. ·
Claude Sonnet 5.5
191 points and 138 comments on Hacker News, linking anthropic.com.
KKKKhazix/AIHOT
A website framework automatically identifies trending topics and generates daily reports. Users can customize its sources and selection criteria to create industry-specific news sites. ·
EmbodiedMemory-Bench: Benchmarking Embodied Memory for Long-Horizon Embodied Tasks
Researchers introduced EmbodiedMemory-Bench, a dataset of 2,554 interactive episodes, to evaluate how agents retain and update environmental memory during long-horizon tasks. ·
Anthropic's IPO prospectus shows AI vision, surging costs
43 points and 36 comments on Hacker News, linking reuters.com.
AdaTutoRank: Learning to Rerank Document Sets via Adaptive Tutoring Optimization for RAG and Deep Research
Researchers introduced AdaTutoRank, a reranker for RAG systems that uses adaptive tutoring optimization to improve document set composition. ·
anthropics/claude-code v2.1.284
Anthropic released Claude Code v2.1.284, making Claude Sonnet 5.5 the default model with a 1M token context window. ·
Trending topics today: heat
Table
| OpenAI | 2.5 |
| World Labs | 1.67 |
| Anthropic | 1.5 |
| Claude Sonnet 5.5 | 1.48 |
| Nvidia | 1.07 |
| Opus 5.5 | 1.06 |
Heat = summed hotness of today's stories naming the topic.
Models
5 storiesReleases, papers, trending weights and new listings
Top 6 trending on Hugging Face: likes
Table
| convaiinnovations/laya | 4323 |
| Qwen/Qwen-Image-2.1 | 2595 |
| abenzerps/Qwen-Image-2.1-Uncensored-GGUF | 2287 |
| XingChen-AGI/Xing4.0-29B-A4B | 1801 |
| Edge0/Audio8-ASR-Infinite | 1410 |
| XingChen-AGI/TeleOCR | 803 |
PirateBench: score out of 100
Table
| opus-5-5 | 91.15 |
| gpt-6-sol | 81.78 |
| deepseek-v4-flash-0731 | 62.82 |
Agents build our own projects from a brief; every run, failures included: https://cowerx.dev/daily/piratebench/
Program-Verified Self-Evolution for Vision-Language Models
Researchers introduced VQS, a method that uses fixed programs to generate verifiable questions for vision-language models. This approach raised the accuracy of generated answers to 94% in human evaluations. ·
5 upvotes on Hugging Face Daily Papers.
Rethinking Training-Inference Mismatch in LLM Reinforcement Learning: Where It Arises and How to Correct It
Researchers introduced calibrated importance sampling to correct probability discrepancies between inference and training engines in large language model reinforcement learning. The method caps importance ratios based on token confidence to bound error. ·
YuE2: Unifying Symbolic and Audio Music Generation at Frontier Quality
YuE2 unifies symbolic and audio music generation by first writing a readable score before producing full-song audio. Experts preferred its output over versions without symbolic planning. ·
Diffusion Reward Models
Researchers introduced DRM, a diffusion-based reward model that estimates reward distributions without fixed parametric assumptions. It matches or surpasses baselines across five benchmarks despite modest training scale. ·
Fewer Tokens, More Self-Teaching: On-Policy Self-Distillation for Extreme Visual Token Reduction
Researchers introduced LT-OPD, a training framework that uses on-policy self-distillation to help multimodal models retain performance when using very few visual tokens. ·
Claude Sonnet 5.5
191 points and 138 comments on Hacker News, linking anthropic.com.
Program-Verified Self-Evolution for Vision-Language Models
Researchers introduced VQS, a method that uses fixed programs to generate verifiable questions for vision-language models. This approach raised the accuracy of generated answers to 94% in human evaluations. ·
Rethinking Training-Inference Mismatch in LLM Reinforcement Learning: Where It Arises and How to Correct It
Researchers introduced calibrated importance sampling to correct probability discrepancies between inference and training engines in large language model reinforcement learning. The method caps importance ratios based on token confidence to bound error. ·
YuE2: Unifying Symbolic and Audio Music Generation at Frontier Quality
YuE2 unifies symbolic and audio music generation by first writing a readable score before producing full-song audio. Experts preferred its output over versions without symbolic planning. ·
Diffusion Reward Models
Researchers introduced DRM, a diffusion-based reward model that estimates reward distributions without fixed parametric assumptions. It matches or surpasses baselines across five benchmarks despite modest training scale. ·
Claude Sonnet 5.5
191 points and 138 comments on Hacker News, linking anthropic.com.
Program-Verified Self-Evolution for Vision-Language Models
Researchers introduced VQS, a method that uses fixed programs to generate verifiable questions for vision-language models. This approach raised the accuracy of generated answers to 94% in human evaluations. ·
Rethinking Training-Inference Mismatch in LLM Reinforcement Learning: Where It Arises and How to Correct It
Researchers introduced calibrated importance sampling to correct probability discrepancies between inference and training engines in large language model reinforcement learning. The method caps importance ratios based on token confidence to bound error. ·
Jeff – Jev-compatible 0.8B decision models, trained at home, ~30 ms
313 points and 126 comments on Hacker News, linking github.com.
YuE2: Unifying Symbolic and Audio Music Generation at Frontier Quality
YuE2 unifies symbolic and audio music generation by first writing a readable score before producing full-song audio. Experts preferred its output over versions without symbolic planning. ·
X voices
5 storiesX posts other sources are linking to, then named voices on Bluesky and Mastodon
@HedgieMarkets
Hospitals use AI to find more things to bill for. Insurers use AI to deny them.
14 points and 1 comments on Hacker News, linking an X post.
@drfeifei
AMD Acquires Dr. Fei Fei Li's World Labs
11 points and 1 comments on Hacker News, linking an X post.
@EdLudlow
AMD is acquiring Fei-Fei Li's World Labs in an $8.2B, all-stock deal
2 points and 1 comments on Hacker News, linking an X post.
@danielhayesmith
I built a native Apple app for 4 platforms in one afternoon with my AI workflow
10 points and 2 comments on Hacker News, linking an X post.
@drfeifei
AMD Acquires Dr. Fei Fei Li's World Labs
11 points and 1 comments on Hacker News, linking an X post.
@thesupermannx
Researchers proved AI has deleted every reason universities exist
2 points and 3 comments on Hacker News, linking an X post.
@EdLudlow
AMD is acquiring Fei-Fei Li's World Labs in an $8.2B, all-stock deal
2 points and 1 comments on Hacker News, linking an X post.
@danielhayesmith
I built a native Apple app for 4 platforms in one afternoon with my AI workflow
10 points and 2 comments on Hacker News, linking an X post.
@drfeifei
AMD Acquires Dr. Fei Fei Li's World Labs
11 points and 1 comments on Hacker News, linking an X post.
@thesupermannx
Researchers proved AI has deleted every reason universities exist
2 points and 3 comments on Hacker News, linking an X post.
@danielhayesmith
I built a native Apple app for 4 platforms in one afternoon with my AI workflow
10 points and 2 comments on Hacker News, linking an X post.
@EdLudlow
AMD is acquiring Fei-Fei Li's World Labs in an $8.2B, all-stock deal
2 points and 1 comments on Hacker News, linking an X post.
GitHub
5 storiesNew repos, releases and changelog
New AI repos this week: stars
Table
| yetone/magpie | 1630 |
| dzhng/jevgrep | 1464 |
| mikehasa/golive-skill | 1049 |
| ollaya-dev/ollaya | 888 |
| feitangyuan/onetake | 858 |
| KKKKhazix/AIHOT | 815 |
ESP32S3 cluster running 1.58-bit (BitNet) Language model
A seven-node ESP32-S3 cluster runs a 0.4B language model using 1.58-bit ternary quantization over an SPI daisy-chain. ·
37 points and 3 comments on Hacker News, linking github.com.
Jeff – Jev-compatible 0.8B decision models, trained at home, ~30 ms
313 points and 126 comments on Hacker News, linking github.com.
langchain-fireworks==1.7.0
LangChain released version 1.7.0 of its Fireworks integration. The update adds prompt caching middleware and fixes mid-stream read timeout classification. ·
anthropics/claude-code v2.1.284
Anthropic released Claude Code v2.1.284, making Claude Sonnet 5.5 the default model with a 1M token context window. ·
langchain==1.4.3
LangChain released version 1.4.3, adding support for Bedrock Mantle chat models and fixing cache settings for fallback models. ·
Jeff – Jev-compatible 0.8B decision models, trained at home, ~30 ms
313 points and 126 comments on Hacker News, linking github.com.
anthropics/claude-code v2.1.284
Anthropic released Claude Code v2.1.284, making Claude Sonnet 5.5 the default model with a 1M token context window. ·
ESP32S3 cluster running 1.58-bit (BitNet) Language model
A seven-node ESP32-S3 cluster runs a 0.4B language model using 1.58-bit ternary quantization over an SPI daisy-chain. ·
unslothai/unsloth Laya Decision Models + Library
Unsloth added decision model support, a document library, and Apple Silicon optimizations. The update also enables local Laya execution and accelerates image and video generation by up to 4.5 times. ·
openai/codex 0.158.0
OpenAI released version 0.158.0 of Codex, adding features for terminal interaction, OAuth connections, and image handling. The update also includes fixes for sandbox issues on Windows, Linux, and macOS. ·
anthropics/claude-code v2.1.284
Anthropic released Claude Code v2.1.284, making Claude Sonnet 5.5 the default model with a 1M token context window. ·
New release v2.1.284 of anthropics/claude-code on GitHub.
Jeff – Jev-compatible 0.8B decision models, trained at home, ~30 ms
313 points and 126 comments on Hacker News, linking github.com.
unslothai/unsloth Laya Decision Models + Library
Unsloth added decision model support, a document library, and Apple Silicon optimizations. The update also enables local Laya execution and accelerates image and video generation by up to 4.5 times. ·
ESP32S3 cluster running 1.58-bit (BitNet) Language model
A seven-node ESP32-S3 cluster runs a 0.4B language model using 1.58-bit ternary quantization over an SPI daisy-chain. ·
openai/codex 0.158.0
OpenAI released version 0.158.0 of Codex, adding features for terminal interaction, OAuth connections, and image handling. The update also includes fixes for sandbox issues on Windows, Linux, and macOS. ·
Data centers & energy
5 storiesBuilds, power deals and the grid
Data-center and energy stories per day
Table
| Wed 23 | 3 |
| Thu 24 | 3 |
| Fri 25 | 20 |
| Sat 26 | 1 |
| Sun 27 | 1 |
| Mon 28 | 20 |
| Tue 29 | 0 |
Stories from DCD, Data Center Knowledge, EIA and Hacker News, by publication day.
The Next Generation of AI Infrastructure Design Starts with Efficiency
AI infrastructure design should integrate power, cooling, and compute to optimize efficiency, resilience, and adaptability rather than focusing solely on capacity. ·
From Data Center Knowledge.
Goodman withdraws application for 90MW data center campus in Sydney, Australia
Goodman withdrew its application for a 90-megawatt data center campus in Sydney, citing significant changes to the local policy and regulatory environment. ·
Re-engineering silicon and system design for AI-era workloads
The piece examines the design of silicon and systems required to support current and future data center workloads for artificial intelligence. ·
Can Supercapacitors Smooth AI’s Power Swings?
Supercapacitors offer millisecond response to AI power fluctuations. Experts view them as complements to existing battery and UPS systems rather than replacements. ·
The Next Generation of AI Infrastructure Design Starts with Efficiency
AI infrastructure design should integrate power, cooling, and compute to optimize efficiency, resilience, and adaptability rather than focusing solely on capacity. ·
From Data Center Knowledge.
Goodman withdraws application for 90MW data center campus in Sydney, Australia
Goodman withdrew its application for a 90-megawatt data center campus in Sydney, citing significant changes to the local policy and regulatory environment. ·
Re-engineering silicon and system design for AI-era workloads
The piece examines the design of silicon and systems required to support current and future data center workloads for artificial intelligence. ·
Can Supercapacitors Smooth AI’s Power Swings?
Supercapacitors offer millisecond response to AI power fluctuations. Experts view them as complements to existing battery and UPS systems rather than replacements. ·
Every Household in This Rural Town Receives $10k If a Data Center Gets Built
10 points and 5 comments on Hacker News, linking wsj.com.
The Next Generation of AI Infrastructure Design Starts with Efficiency
AI infrastructure design should integrate power, cooling, and compute to optimize efficiency, resilience, and adaptability rather than focusing solely on capacity. ·
Goodman withdraws application for 90MW data center campus in Sydney, Australia
Goodman withdrew its application for a 90-megawatt data center campus in Sydney, citing significant changes to the local policy and regulatory environment. ·
Re-engineering silicon and system design for AI-era workloads
The piece examines the design of silicon and systems required to support current and future data center workloads for artificial intelligence. ·
Hardware prices
2 storiesGPUs, memory, and what tokens cost
Output price, $ per million tokens (newest models)
Table
| claude-sonnet-5.5 | 10 |
| claude-sonnet-5.5:batch | 5 |
| perceptron-mk1.5 | 1.5 |
| ember-1 | 15 |
| glm-5.3-prime | 8.8 |
| median, all paid models | 2 |
434 paid models listed; median of the 56 added in 30 days: $4.00. Rankings data by OpenRouter, CC BY 4.0.
Import AI 474: Platonic mindspace; TPUs in space; Zhipu starts an outer RSI loop
Import AI issue 474 covers Platonic mindspace, space-based TPUs, and Zhipu's outer RSI loop. ·
From Import AI.
Import AI 474: Platonic mindspace; TPUs in space; Zhipu starts an outer RSI loop
Import AI issue 474 covers Platonic mindspace, space-based TPUs, and Zhipu's outer RSI loop. ·
From Import AI.
Import AI 474: Platonic mindspace; TPUs in space; Zhipu starts an outer RSI loop
Import AI issue 474 covers Platonic mindspace, space-based TPUs, and Zhipu's outer RSI loop. ·
From Import AI.
Virtio-nvgpu: Near-native Nvidia GPU access inside a KVM guest
An experimental project provides a virtio device that allows KVM virtual machines to access NVIDIA GPUs with near-native performance. ·
