Top now
6 storiesWhat is climbing fastest across every source
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. ·
AI existential risk probabilities are (still) too unreliable to inform policy
Researchers argue that current estimates of AI existential risk are too unreliable to guide policy because speculation is often presented as precise data. ·
AI companies in race to demonstrate their model most threatening to humanity
389 points and 310 comments on Hacker News, linking thecivilian.co.nz.
Prompting Claude Opus 5.5
174 points and 186 comments on Hacker News, linking platform.claude.com.
AI companies in race to demonstrate their model most threatening to humanity
389 points and 310 comments on Hacker News, linking thecivilian.co.nz.
Prompting Claude Opus 5.5
174 points and 186 comments on Hacker News, linking platform.claude.com.
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. ·
FuseReg: Regularizing Layer Fusion Mitigates the Reconstruction-Generation Gap in Representation Autoencoders
FuseReg improves representation autoencoders by training over random encoder layer subsets instead of fixed heuristics. This approach enhances both reconstruction quality and generation performance without requiring retraining. ·
RayOrch: Programming and Executing Lineage-Controlled Multi-Grain Dataflows for Foundation-Model Data Preparation
Researchers introduced RayOrch, a system for preparing foundation-model training data. It manages complex data pipelines by tracking relationships between parent items and their generated children. ·
XingChen-AGI/TeleOCR
A image text to text model from XingChen-AGI with 655 likes and 27.9k downloads on Hugging Face.
AI companies in race to demonstrate their model most threatening to humanity
389 points and 310 comments on Hacker News, linking thecivilian.co.nz.
Prompting Claude Opus 5.5
174 points and 186 comments on Hacker News, linking platform.claude.com.
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. ·
FuseReg: Regularizing Layer Fusion Mitigates the Reconstruction-Generation Gap in Representation Autoencoders
FuseReg improves representation autoencoders by training over random encoder layer subsets instead of fixed heuristics. This approach enhances both reconstruction quality and generation performance without requiring retraining. ·
RayOrch: Programming and Executing Lineage-Controlled Multi-Grain Dataflows for Foundation-Model Data Preparation
Researchers introduced RayOrch, a system for preparing foundation-model training data. It manages complex data pipelines by tracking relationships between parent items and their generated children. ·
XingChen-AGI/TeleOCR
A image text to text model from XingChen-AGI with 655 likes and 27.9k downloads on Hugging Face.
Trending topics today: heat
Table
| OpenAI | 1.48 |
| Image-2.1 | 0.84 |
| Microsoft | 0.8 |
| Qwen3.8 | 0.78 |
| Nvidia | 0.74 |
| TeleOCR | 0.71 |
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 | 4227 |
| Qwen/Qwen-Image-2.1 | 2558 |
| abenzerps/Qwen-Image-2.1-Uncensored-GGUF | 2197 |
| XingChen-AGI/Xing4.0-29B-A4B | 1790 |
| Edge0/Audio8-ASR-Infinite | 1360 |
| XiaomiMiMo/MiMo-V2.6-Pro-RL | 571 |
PirateBench: score out of 100
Table
| opus-5-5 | 91.15 |
| gpt-6-sol | 80.94 |
| deepseek-v4-flash-0731 | 62.82 |
Agents build our own projects from a brief; every run, failures included: https://cowerx.dev/daily/piratebench/
Holo4: powering generalist computer-use agents
From Hugging Face blog.
TabPFN and TabICL vs. tuned XGBoost: the model that doesn't train won 14/14
TabPFN and TabICL are models that predict on tabular data without training. They reportedly outperformed tuned XGBoost in all fourteen tested scenarios. ·
XingChen-AGI/TeleOCR
A image text to text model from XingChen-AGI with 655 likes and 27.9k downloads on Hugging Face.
CodeGraph: Open-Taxonomy Knowledge Graph for Source Code with Wikidata Grounding
Researchers built a pipeline using a code-focused language model to create a semantic knowledge graph from source code. They applied it to 167 million files in the Stack-Edu corpus. ·
SAGE: Mitigating Long-Horizon Reasoning Biases via Topological Guidance
Researchers introduced SAGE, a framework using algebraic and hyperbolic guidance to reduce exploration and compounding biases in long-horizon reasoning for large language models. ·
FuseReg: Regularizing Layer Fusion Mitigates the Reconstruction-Generation Gap in Representation Autoencoders
FuseReg improves representation autoencoders by training over random encoder layer subsets instead of fixed heuristics. This approach enhances both reconstruction quality and generation performance without requiring retraining. ·
110 upvotes on Hugging Face Daily Papers.
RayOrch: Programming and Executing Lineage-Controlled Multi-Grain Dataflows for Foundation-Model Data Preparation
Researchers introduced RayOrch, a system for preparing foundation-model training data. It manages complex data pipelines by tracking relationships between parent items and their generated children. ·
TabPFN and TabICL vs. tuned XGBoost: the model that doesn't train won 14/14
TabPFN and TabICL are models that predict on tabular data without training. They reportedly outperformed tuned XGBoost in all fourteen tested scenarios. ·
Tactile-JEPA: Topology-Aware Self-Supervised Representation Learning for Distributed Tactile Sensors
Researchers developed Tactile-JEPA, a self-supervised method that uses sensor connectivity graphs to learn representations for distributed tactile sensors. ·
InternW0-Δ: A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data
InternW0-Δ is a robot manipulation model that unifies visual dynamics and action generation. It uses a Mixture-of-Transformers architecture and a large open dataset to improve performance on simulation and real-world tasks. ·
FuseReg: Regularizing Layer Fusion Mitigates the Reconstruction-Generation Gap in Representation Autoencoders
FuseReg improves representation autoencoders by training over random encoder layer subsets instead of fixed heuristics. This approach enhances both reconstruction quality and generation performance without requiring retraining. ·
110 upvotes on Hugging Face Daily Papers.
RayOrch: Programming and Executing Lineage-Controlled Multi-Grain Dataflows for Foundation-Model Data Preparation
Researchers introduced RayOrch, a system for preparing foundation-model training data. It manages complex data pipelines by tracking relationships between parent items and their generated children. ·
Tactile-JEPA: Topology-Aware Self-Supervised Representation Learning for Distributed Tactile Sensors
Researchers developed Tactile-JEPA, a self-supervised method that uses sensor connectivity graphs to learn representations for distributed tactile sensors. ·
InternW0-Δ: A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data
InternW0-Δ is a robot manipulation model that unifies visual dynamics and action generation. It uses a Mixture-of-Transformers architecture and a large open dataset to improve performance on simulation and real-world tasks. ·
Block Sparse Attention with Log-Linear Complexity
PISA is a block-sparse attention mechanism that uses a pyramid Top-K strategy to reduce self-attention complexity from quadratic to log-linear. It employs hierarchical key selection and hardware-aware kernels to improve efficiency for long-context language models. ·
X voices
5 storiesX posts other sources are linking to, then named voices on Bluesky and Mastodon
George Hotz’s opinion on AI coding
21 points and 10 comments on Hacker News, linking an X post.
Hospitals use AI to find more things to bill for. Insurers use AI to deny them.
2 points and 0 comments on Hacker News, linking an X post.
this statement:
The author argues that releasing cheaper models due to compute shortages contradicts the Jevons Paradox. ·
OpenAI’s dev day is on Tuesday. They have “too much to launch”, meaning they’ll likely spill over the launches into Monday & Wednesday.
OpenAI is holding a developer event on Tuesday, with announcements potentially extending to Monday and Wednesday. The company expects to release a new GPT-6 model and products utilizing persistent behavior. ·
I was reflecting on how many random contingencies went into creating the current accelerating AI moment and had Opus 5.5 put together a video, inspired by old science show Connections, to try and poin
A video created by Opus 5.5 illustrates how various unrelated ideas converged to produce the current rapid development of artificial intelligence. ·
George Hotz’s opinion on AI coding
21 points and 10 comments on Hacker News, linking an X post.
An agent in a loop optimizing a renderer with the goal to minimize frame times
2 points and 0 comments on Hacker News, linking an X post.
Font where each token is equal-width
26 points and 14 comments on Hacker News, linking an X post.
Hospitals use AI to find more things to bill for. Insurers use AI to deny them.
2 points and 0 comments on Hacker News, linking an X post.
George Hotz’s opinion on AI coding
21 points and 10 comments on Hacker News, linking an X post.
Font where each token is equal-width
26 points and 14 comments on Hacker News, linking an X post.
An agent in a loop optimizing a renderer with the goal to minimize frame times
2 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
| yetone/magpie | 1539 |
| dzhng/jevgrep | 1171 |
| mikehasa/golive-skill | 1032 |
| ollaya-dev/ollaya | 803 |
| yukitorido/short-video-generator-AI | 730 |
| feitangyuan/onetake | 662 |
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. ·
New release rust-v0.158.0 of openai/codex on GitHub.
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. ·
New release rust-v0.158.0 of openai/codex on GitHub.
zhuyansen/awesome-claude-video-skills
A developer published a curated list of 180 open-source repositories. These tools allow coding agents to generate video and include security grades. ·
admte/orc-claude-plugin
This is a Claude plugin designed for the ORC8R Cloud MCP server. ·
unslothai/unsloth Flash-Attention2, Causal-Conv1D, Mamba_SSM Binaries
Prebuilt Linux x86_64 wheels for flash-attn, causal-conv1d, and mamba-ssm are available. They support CUDA 13, PyTorch 2.13 and 2.14, and Python 3.13. ·
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. ·
New release rust-v0.158.0 of openai/codex on GitHub.
zhuyansen/awesome-claude-video-skills
A developer published a curated list of 180 open-source repositories. These tools allow coding agents to generate video and include security grades. ·
unslothai/unsloth Flash-Attention2, Causal-Conv1D, Mamba_SSM Binaries
Prebuilt Linux x86_64 wheels for flash-attn, causal-conv1d, and mamba-ssm are available. They support CUDA 13, PyTorch 2.13 and 2.14, and Python 3.13. ·
lemomo-ai/lemo-opuscar
A repository offers thirty-nine film styles, each with a reusable prompt and a short code-generated sample. Users can select a style and use an agent to direct their own story. ·
dzhng/jevgrep
This is a command-line tool for coding agents. It uses Jev to locate relevant files and source code by describing functionality. ·
Data centers & energy
5 storiesBuilds, power deals and the grid
Data-center and energy stories per day
Table
| Tue 22 | 3 |
| Wed 23 | 3 |
| Thu 24 | 3 |
| Fri 25 | 20 |
| Sat 26 | 1 |
| Sun 27 | 1 |
| Mon 28 | 15 |
Stories from DCD, Data Center Knowledge, EIA and Hacker News, by publication day.
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.
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.
Hardware prices
2 storiesGPUs, memory, and what tokens cost
Output price, $ per million tokens (newest models)
Table
| perceptron-mk1.5 | 1.5 |
| ember-1 | 15 |
| glm-5.3-prime | 8.8 |
| qwen3.8-max-prime | 12 |
| aion-3.5-mini | 1.4 |
| median, all paid models | 2 |
431 paid models listed; median of the 52 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. ·
