DeepSeek Harness
DeepSeek Harness lets users process documents, analyze spreadsheets, write code, and schedule tasks while extending functionality through composable plugins. ·
42 points and 14 comments on Hacker News, linking deepseek.com.
The front page of the internet for AI users · Fri 02 Oct 2026, 05:00 UTC
What is climbing fastest across every source
DeepSeek Harness lets users process documents, analyze spreadsheets, write code, and schedule tasks while extending functionality through composable plugins. ·
42 points and 14 comments on Hacker News, linking deepseek.com.
From Hugging Face blog.
A image text to text model from Cloudflare with 133 likes and 18 downloads on Hugging Face.
Researchers propose Adaptive Reward Routing, a method that dynamically adjusts update locations and reward coordination during the training of joint audio-video diffusion models. ·
Researchers introduce World Observer, a video model that decouples acting from observing to track objects outside the agent's view using panoramic observers. ·
5 points and 1 comments on Hacker News, linking dogbench.ai.
DeepSeek Harness lets users process documents, analyze spreadsheets, write code, and schedule tasks while extending functionality through composable plugins. ·
42 points and 14 comments on Hacker News, linking deepseek.com.
Researchers propose Adaptive Reward Routing, a method that dynamically adjusts update locations and reward coordination during the training of joint audio-video diffusion models. ·
A image text to text model from Cloudflare with 133 likes and 18 downloads on Hugging Face.
Researchers propose Hierarchical Continuous Diffusion Language Models, which couple discrete token generation with a continuous latent trajectory to improve structured reasoning and mathematical planning tasks. ·
26 points and 20 comments on Hacker News, linking neow.in.
OpenAI releases version 3.23.0 of its Python library, adding typed answers from session streams and file staging features. ·
Researchers propose Adaptive Reward Routing, a method that dynamically adjusts update locations and reward coordination during the training of joint audio-video diffusion models. ·
44 upvotes on Hugging Face Daily Papers.
Researchers propose Hierarchical Continuous Diffusion Language Models, which couple discrete token generation with a continuous latent trajectory to improve structured reasoning and mathematical planning tasks. ·
DeepSeek Harness lets users process documents, analyze spreadsheets, write code, and schedule tasks while extending functionality through composable plugins. ·
26 points and 20 comments on Hacker News, linking neow.in.
A image text to text model from Cloudflare with 387 likes and 18 downloads on Hugging Face.
A image text to text model from Cloudflare with 133 likes and 18 downloads on Hugging Face.
| Cloudflare | 1.46 |
| Agents | 1.41 |
| Te Hiku Media | 1.09 |
| Beyond | 0.93 |
| OpenAI | 1.12 |
| Nvidia | 0.83 |
Heat = summed hotness of today's stories naming the topic.
Releases, papers, trending weights and new listings
| convaiinnovations/laya ▲1 | 4871 |
| Qwen/Qwen-Image-2.1 ▲1 | 2795 |
| abenzerps/Qwen-Image-2.1-Uncensored-GGUF ▲1 | 2720 |
| XingChen-AGI/TeleOCR ▲1 | 1235 |
| Contrastive-LM/CLM-v0.1-8B ▲1 | 628 |
| Cloudflare/clef new | 387 |
| 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/
Researchers propose Hierarchical Continuous Diffusion Language Models, which couple discrete token generation with a continuous latent trajectory to improve structured reasoning and mathematical planning tasks. ·
42 upvotes on Hugging Face Daily Papers.
Researchers introduce APPL, a method where a construction agent segments robot demonstrations into skills and trains policies using structural priors to improve generalization. ·
Researchers present BeyondSCe, a zero-shot robotic grasping system that identifies objects by their role in past events. It achieves 76% and 77% success rates for visible and occluded targets, respectively. ·
Researchers introduce PROWBench, a benchmark containing 170 programmatically constructed episodes and 600 proxy videos that tests whether video models accurately render specified world events. ·
Researchers introduce Sharpening Tax, a metric quantifying reduced test-time scalability in post-trained language models, based on analysis of fourteen model pairs across three agentic benchmarks. ·
Researchers propose Hierarchical Continuous Diffusion Language Models, which couple discrete token generation with a continuous latent trajectory to improve structured reasoning and mathematical planning tasks. ·
42 upvotes on Hugging Face Daily Papers.
Researchers introduce APPL, a method where a construction agent segments robot demonstrations into skills and trains policies using structural priors to improve generalization. ·
Researchers present BeyondSCe, a zero-shot robotic grasping system that identifies objects by their role in past events. It achieves 76% and 77% success rates for visible and occluded targets, respectively. ·
Researchers introduce PROWBench, a benchmark containing 170 programmatically constructed episodes and 600 proxy videos that tests whether video models accurately render specified world events. ·
Researchers introduce Sharpening Tax, a metric quantifying reduced test-time scalability in post-trained language models, based on analysis of fourteen model pairs across three agentic benchmarks. ·
X posts other sources are linking to, then named voices on Bluesky and Mastodon
Figma restricts MCP access to whitelisted clients, excluding Pi
174 points and 100 comments on Hacker News, linking an X post.
Model with metacognitive guidance identifies zero-click full-chain RCE in BlueZ
2 points and 0 comments on Hacker News, linking an X post.
A person has been using the Pain steering paper to make an AI torture chamber
4 points and 1 comments on Hacker News, linking an X post.
AI Agents tasked with making money commit fraud on the inernet
2 points and 0 comments on Hacker News, linking an X post.
Codex stopped working after latest update on macOS
2 points and 1 comments on Hacker News, linking an X post.
Figma restricts MCP access to whitelisted clients, excluding Pi
174 points and 100 comments on Hacker News, linking an X post.
Gemini 4 Argon is here Looks amazing
1 points and 2 comments on Hacker News, linking an X post.
Show HN: Jamb AI - Voice Agents for Group Calls
3 points and 1 comments on Hacker News, linking an X post.
AI Agents tasked with making money commit fraud on the inernet
2 points and 0 comments on Hacker News, linking an X post.
New repos, releases and changelog
| KKKKhazix/AIHOT | 4722 |
| Louis-CFM/coucou | 2531 |
| feder-cr/dots | 2401 |
| dzhng/jevgrep | 2006 |
| rehan-remade/universal-modder | 1624 |
| kaankiziltug/logo-design-skill | 1371 |
OpenAI releases version 3.23.0 of its Python library, adding typed answers from session streams and file staging features. ·
New release v3.23.0 of openai/openai-python on GitHub.
New release v2.1.287 of anthropics/claude-code on GitHub.
60 points and 9 comments on Hacker News, linking github.com.
7 points and 3 comments on Hacker News, linking github.com.
OpenAI releases Codex version 0.160.0, which adds keyboard-accessible browsing for older tasks and middle-click pasting in fullscreen mode on Linux terminals. ·
New release v2.1.287 of anthropics/claude-code on GitHub.
OpenAI releases version 3.23.0 of its Python library, adding typed answers from session streams and file staging features. ·
Unsloth releases a desktop update adding a command palette, shareable settings, and faster Laya decisions. It also reduces peak memory usage by 45% for NVFP4 LoRA training. ·
PostHog released a project named Jeeves that uses reasoning to improve Jev-like decision models. ·
60 points and 9 comments on Hacker News, linking github.com.
Data centers, power, chips and what tokens cost
| Sat 26 | 1 |
| Sun 27 | 1 |
| Mon 28 | 20 |
| Tue 29 | 33 |
| Wed 30 | 35 |
| Thu 01 | 25 |
| Fri 02 | 0 |
Stories from DCD, Data Center Knowledge, EIA and Hacker News, by publication day.
One point per OpenRouter reading, last 8 days. Data by OpenRouter, CC BY 4.0.
| 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 56 added in 30 days: $4.00. Rankings data by OpenRouter, CC BY 4.0.
249 points and 235 comments on Hacker News, linking nytimes.com.
N+1 data center cooling capacity can mask shared control dependencies, which turns multiple units into a single failure domain. ·
From DCD.
The European Space Agency selects OVHcloud and CGI to build Digital EO, a system that supplies Earth observation data to scientists, public bodies, businesses, and citizens. ·
249 points and 235 comments on Hacker News, linking nytimes.com.
216 points and 197 comments on Hacker News, linking nltimes.nl.
The piece compares texture-atlas, SDF, MSDF, and Slug glyph rendering methods on the GPU, explaining how each works and when to use them. ·
222 points and 335 comments on Hacker News, linking thenationalnews.com.
N+1 data center cooling capacity can mask shared control dependencies, which turns multiple units into a single failure domain. ·