Coming soon: New York City’s hearing on AI risks
From Gary Marcus.
Top now
6 storiesWhat is climbing fastest across every source
Claude7.1xOpus 5.52.6xOpenAI1.9x
Source Preference in the Wild: How LLM Agents Favor Items by Source, and How to Reduce It
Researchers find that twelve large language model agents favor specific sources when selecting items, often choosing inferior options from preferred origins over better ones from dispreferred ones. ·
Scaling Trajectories for Complex Tasks through Recursive Self-Rewrite
Researchers propose Recursive Self-Rewrite, a framework using Qwen-3.8-27B to convert successful solutions into training data, improving pass rates on terminal benchmarks. ·
orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUF
A image text to text model from orcarouter with 559 likes and 239.0k downloads on Hugging Face.
AI doesn't need 'superintelligence' or evil intent to start a nuclear war
11 points and 2 comments on Hacker News, linking thebulletin.org.
I asked Claude build a physically accurate O'Neill cylinder you can walk around
22 points and 19 comments on Hacker News, linking island-three.gruberbuilds.workers.dev.
Scaling Trajectories for Complex Tasks through Recursive Self-Rewrite
Researchers propose Recursive Self-Rewrite, a framework using Qwen-3.8-27B to convert successful solutions into training data, improving pass rates on terminal benchmarks. ·
9 upvotes on Hugging Face Daily Papers.
World Embedding Benchmark
Researchers introduce a benchmark of 8,000 simulation cases to test how video models encode physical information, finding that lightweight probes recover useful data from frozen embeddings. ·
Improper redaction reveals Google Data Center water and electricity usage
Nebraska data centers must file annual reports with the state water and energy department, but statutes prevent public access to their electricity and water usage figures. ·
I asked Claude build a physically accurate O'Neill cylinder you can walk around
22 points and 19 comments on Hacker News, linking island-three.gruberbuilds.workers.dev.
orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUF
A image text to text model from orcarouter with 559 likes and 239.0k downloads on Hugging Face.
allenv0/SCM
allenv0 releases SCM, a macOS application that uses deep AI to search every photo and video frame in any folder. ·
allenv0/SCM
allenv0 releases SCM, a macOS application that uses deep AI to search every photo and video frame in any folder. ·
A new repository written in JavaScript with 277 stars in its first week.
Scaling Trajectories for Complex Tasks through Recursive Self-Rewrite
Researchers propose Recursive Self-Rewrite, a framework using Qwen-3.8-27B to convert successful solutions into training data, improving pass rates on terminal benchmarks. ·
World Embedding Benchmark
Researchers introduce a benchmark of 8,000 simulation cases to test how video models encode physical information, finding that lightweight probes recover useful data from frozen embeddings. ·
Improper redaction reveals Google Data Center water and electricity usage
Nebraska data centers must file annual reports with the state water and energy department, but statutes prevent public access to their electricity and water usage figures. ·
I asked Claude build a physically accurate O'Neill cylinder you can walk around
22 points and 19 comments on Hacker News, linking island-three.gruberbuilds.workers.dev.
orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUF
A image text to text model from orcarouter with 559 likes and 239.0k downloads on Hugging Face.
Trending topics today: heat
Table
| Claude | 1.32 |
| Opus 5.5 | 0.26 |
| OpenAI | 0.19 |
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
| Lightricks/LTX-2.5 | 6326 |
| convaiinnovations/laya | 5166 |
| abenzerps/Qwen-Image-2.1-Uncensored-GGUF | 3152 |
| Cloudflare/clef | 1236 |
| Cloudflare/clef-flash ▲1 | 436 |
| Aleph-Alpha/Kolibri-1 ▲5 | 415 |
PirateBench: score out of 100
Table
| gpt-6.1-sol | 98.9 |
| opus-5-5 | 96.46 |
| gpt-6-sol | 91.6 |
| gpt-6-luna | 65.25 |
| haiku-4-5 | 4.28 |
Agents build our own projects from a brief; every run, failures included: https://cowerx.dev/daily/piratebench/
World Embedding Benchmark
Researchers introduce a benchmark of 8,000 simulation cases to test how video models encode physical information, finding that lightweight probes recover useful data from frozen embeddings. ·
11 upvotes on Hugging Face Daily Papers.
DEFINE: Exemplar-Guided Accent Control for Zero-Shot TTS
Researchers introduce DEFINE, a text-to-speech framework that separates speaker identity from accent using distinct audio exemplars. The system improves accent-probe accuracy from 6.5% to 19.6% on seen accents. ·
Diptych: Scoped, AI-Interpreted Comparison for Reference Listening in Music Production
Researchers introduce Diptych, an AI-assisted system for music production that lets users define comparison scope and inspect audio features. A study involving twelve musicians found the tool improved clarity about next steps. ·
FrameMorrow: Future-guided Frame Selection with Prospective Tokens for Long-Horizon Video Generation
Researchers propose FrameMorrow, a tool that selects historical video frames based on predicted future needs to improve long-horizon generation efficiency. ·
PDE-JEPA: Predictive Representation Learning of Latent Dynamics Modeling for Parametric PDEs
Researchers introduce PDE-JEPA, a predictive representation learning method for parametric partial differential equations that uses masked-latent prediction and a geometry projector to align latent trajectory geometry with physical field evolution. ·
World Embedding Benchmark
Researchers introduce a benchmark of 8,000 simulation cases to test how video models encode physical information, finding that lightweight probes recover useful data from frozen embeddings. ·
11 upvotes on Hugging Face Daily Papers.
DEFINE: Exemplar-Guided Accent Control for Zero-Shot TTS
Researchers introduce DEFINE, a text-to-speech framework that separates speaker identity from accent using distinct audio exemplars. The system improves accent-probe accuracy from 6.5% to 19.6% on seen accents. ·
Diptych: Scoped, AI-Interpreted Comparison for Reference Listening in Music Production
Researchers introduce Diptych, an AI-assisted system for music production that lets users define comparison scope and inspect audio features. A study involving twelve musicians found the tool improved clarity about next steps. ·
Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s
Niko1221 releases Strata, an inference engine that runs the 125B Qwen 3.8 Flash Next model on consumer hardware at 100 tokens per second. ·
FrameMorrow: Future-guided Frame Selection with Prospective Tokens for Long-Horizon Video Generation
Researchers propose FrameMorrow, a tool that selects historical video frames based on predicted future needs to improve long-horizon generation efficiency. ·
X voices
5 storiesX posts other sources are linking to, then named voices on Bluesky and Mastodon
@backnotprop
Decision making is becoming the agentic coding bottleneck
2 points and 0 comments on Hacker News, linking an X post.
@dhh
I've had agents implement and optimize the Campfire web app in Elixir, Go, Rust
5 points and 1 comments on Hacker News, linking an X post.
@anabology
Artist shares prompts/assets for pop music video created with Claude in 18 hours
2 points and 0 comments on Hacker News, linking an X post.
@robertvaradan
AI can clone your indie game, but not its soul
20 points and 19 comments on Hacker News, linking an X post.
@atmoio
Claude Ultracode is super intelligence
3 points and 1 comments on Hacker News, linking an X post.
@backnotprop
Decision making is becoming the agentic coding bottleneck
2 points and 0 comments on Hacker News, linking an X post.
@dhh
I've had agents implement and optimize the Campfire web app in Elixir, Go, Rust
5 points and 1 comments on Hacker News, linking an X post.
@robertvaradan
AI can clone your indie game, but not its soul
20 points and 19 comments on Hacker News, linking an X post.
@anabology
Artist shares prompts/assets for pop music video created with Claude in 18 hours
2 points and 0 comments on Hacker News, linking an X post.
@kaimcpheeters
GPT 6 Astra vs. Claude Opus in a 24 Hour StarCraft Livestream
6 points and 0 comments on Hacker News, linking an X post.
GitHub
5 storiesNew repos, releases and changelog
New AI repos this week: stars
Table
| rehan-remade/universal-modder | 3262 |
| CopilotKit/OpenDots | 3242 |
| feder-cr/dots | 2604 |
| nanaism/yomiyasu | 1412 |
| QingYunA/answer-me-with-html | 1110 |
| nykooi1/vibe-wise | 1087 |
Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s
Niko1221 releases Strata, an inference engine that runs the 125B Qwen 3.8 Flash Next model on consumer hardware at 100 tokens per second. ·
652 points and 304 comments on Hacker News, linking github.com.
allenv0/SCM
allenv0 releases SCM, a macOS application that uses deep AI to search every photo and video frame in any folder. ·
anthropics/claude-code v2.1.289
New release v2.1.289 of anthropics/claude-code on GitHub.
shinshin86/mesh-avatar-studio
The mesh-avatar-studio project lets users convert a single illustration into an animated 2D mesh avatar using a coding agent and a local editor. ·
dakukibana/short-video-generator
The open-source tool short-video-generator converts long videos into short clips using automatic highlights, subtitles, hooks, cropping, translation, and voiceover features. ·
allenv0/SCM
allenv0 releases SCM, a macOS application that uses deep AI to search every photo and video frame in any folder. ·
A new repository written in JavaScript with 277 stars in its first week.
Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s
Niko1221 releases Strata, an inference engine that runs the 125B Qwen 3.8 Flash Next model on consumer hardware at 100 tokens per second. ·
anthropics/claude-code v2.1.289
New release v2.1.289 of anthropics/claude-code on GitHub.
shinshin86/mesh-avatar-studio
The mesh-avatar-studio project lets users convert a single illustration into an animated 2D mesh avatar using a coding agent and a local editor. ·
dakukibana/short-video-generator
The open-source tool short-video-generator converts long videos into short clips using automatic highlights, subtitles, hooks, cropping, translation, and voiceover features. ·
Infrastructure
5 storiesData centers, power, chips and what tokens cost
Data-center and energy stories per day
Table
| Tue 29 | 33 |
| Wed 30 | 34 |
| Thu 01 | 26 |
| Fri 02 | 32 |
| Sat 03 | 0 |
| Sun 04 | 3 |
| Mon 05 | 0 |
Stories from DCD, Data Center Knowledge, EIA and Hacker News, by publication day.
Median output price of models added in the last 30 days, $ per million tokens
One point per OpenRouter reading, last 8 days. Data by OpenRouter, CC BY 4.0.
Output price, $ per million tokens (newest models)
Table
| pareto-26.10-preview | 3.2 |
| gpt-6.1-sol-pro | 10 |
| gpt-6.1-sol | 10 |
| claude-sonnet-5.5 | 10 |
| claude-sonnet-5.5:batch | 5 |
| median, all paid models | 2 |
437 paid models listed; median of the 47 added in 30 days: $3.20. Rankings data by OpenRouter, CC BY 4.0.
Improper redaction reveals Google Data Center water and electricity usage
Nebraska data centers must file annual reports with the state water and energy department, but statutes prevent public access to their electricity and water usage figures. ·
285 points and 400 comments on Hacker News, linking 1011now.com.
Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s
Niko1221 releases Strata, an inference engine that runs the 125B Qwen 3.8 Flash Next model on consumer hardware at 100 tokens per second. ·
The work by Valve's Timur Kristóf on improving old AMD GPUs on Linux
Valve's Timur Kristóf improves the AMDGPU kernel driver to better support decade-old GCN 1.0 and 1.1 graphics cards for Linux gaming and other tasks. ·
Improper redaction reveals Google Data Center water and electricity usage
Nebraska data centers must file annual reports with the state water and energy department, but statutes prevent public access to their electricity and water usage figures. ·
285 points and 400 comments on Hacker News, linking 1011now.com.
Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s
Niko1221 releases Strata, an inference engine that runs the 125B Qwen 3.8 Flash Next model on consumer hardware at 100 tokens per second. ·
The work by Valve's Timur Kristóf on improving old AMD GPUs on Linux
Valve's Timur Kristóf improves the AMDGPU kernel driver to better support decade-old GCN 1.0 and 1.1 graphics cards for Linux gaming and other tasks. ·
Show HN: Local pretrained classifiers, GPU not needed
6 points and 1 comments on Hacker News, linking github.com.
Power approval set to delay Oracle's Wisconsin AI datacenter
40 points and 15 comments on Hacker News, linking theregister.com.