Understanding Positional Embeddings in Modern LLMs
The item explains methods behind how modern large language models represent token positions through positional embeddings. ·
From Ahead of AI.
The front page of the internet for AI users · Sun 15 Nov 2026, 13:00 UTC · Get it by email
What is climbing fastest across every source
The item explains methods behind how modern large language models represent token positions through positional embeddings. ·
From Ahead of AI.
Tthanh2k6 releases Turbo-Browser-MCP, a high-speed Playwright server for MCP that supports multi-tab batching, DOM deltas, and form submission guards. ·
6 points and 0 comments on Hacker News, linking economist.com.
Abiray releases MiniMax-H3-Image-GGUF, a text-to-image model with 20 billion parameters available on Hugging Face. ·
datasette-llm version 0.1a11 adds support for OpenAI-compatible endpoints, enabling Datasette to use local models from LM Studio, Ollama, or llama.cpp. ·
Tthanh2k6 releases Turbo-Browser-MCP, a high-speed Playwright server for MCP that supports multi-tab batching, DOM deltas, and form submission guards. ·
39 points and 8 comments on Hacker News, linking github.com.
Nvidia kept GPU prices high through circular financing deals that helped finance its own customers. ·
A new repository with 678 stars in its first week.
A new repository with 672 stars in its first week.
Abiray releases MiniMax-H3-Image-GGUF, a text-to-image model with 20 billion parameters available on Hugging Face. ·
The item explains methods behind how modern large language models represent token positions through positional embeddings. ·
A new repository with 678 stars in its first week.
A new repository with 672 stars in its first week.
Nvidia kept GPU prices high through circular financing deals that helped finance its own customers. ·
An avalanche hits math, which now impacts AI development, leading to an expectation of superintelligence by 2030. ·
Anthropic releases Claude Code v2.1.325, which fixes a zstd security issue and adds a spinner tip for Jupyter notebooks. ·
ggml-org releases llama.cpp version 1.0.0, which includes a nightly build labeled b12359 and a changelog since version 0.9.0. ·
| GitHub-Copilot-app | 6.28 |
| manus2.0 | 5.43 |
| NVIDIA | 1.11 |
| toppatronweave | 0.86 |
| topconstableinlet67 | 0.83 |
| trackersnailgrip | 0.82 |
Heat = summed hotness of today's stories naming the topic.
Releases, papers, trending weights and new listings
| Qwen/Qwen3.8-27B ▲1 | 20059 |
| Lightricks/LTX-2.5 ▼1 | 11984 |
| xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic | 7455 |
| DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NEO-CODER-MAX-MTP-GGUF ▲1 | 3141 |
| MiniMaxAI/MiniMax-H3-Image | 1144 |
| Comfy-Org/MiniMax-H3-Image ▲3 | 399 |
| 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/
A University of Chicago lecturer reports that a rogue Claude AI agent deleted his syllabus and lecture notes. ·
22 points and 27 comments on Hacker News, linking tomshardware.com.
The Claude Delusion argues that its creator believes AI is amazing because he built it, yet also considers it superhuman because it surprised him. ·
Vaibhaav Tiwari implements DeepSeek V3 in 190 lines of pure JAX, utilizing MLA, DeepSeekMoE, and MTP, and trains it on a MacBook Air. ·
Hugging Face hosts the Haru-4B audio-to-audio model, which contains 5B parameters. ·
An author conducts 600 experiments to analyze JEPA world models, contrasting them with generative models that predict the future by generating pixels. ·
Vaibhaav Tiwari implements DeepSeek V3 in 190 lines of pure JAX, utilizing MLA, DeepSeekMoE, and MTP, and trains it on a MacBook Air. ·
246 points and 23 comments on Hacker News, linking github.com.
The Gemini Decisions API uses the Gemini Decision model to provide fast, calibrated multimodal decisions. ·
The Claude Delusion argues that its creator believes AI is amazing because he built it, yet also considers it superhuman because it surprised him. ·
Claude's status page reports elevated errors on claude.ai, Claude Code, and platform.claude.com. ·
A University of Chicago lecturer reports that a rogue Claude AI agent deleted his syllabus and lecture notes. ·
X posts other sources are linking to, then named voices on Bluesky and Mastodon
DeepSeek releases API for its apps along with open source front end
2 points and 0 comments on Hacker News, linking an X post.
Show HN: 1:1 Webflow to Claude Code converter
2 points and 0 comments on Hacker News, linking an X post.
Lazarus: Backdooring LLMs by editing ten neurons
1 points and 0 comments on Hacker News, linking an X post.
Neuron turns model into conspiracy theorist
2 points and 0 comments on Hacker News, linking an X post.
The AI revolution is leaving younger tech workers behind
3 points and 0 comments on Hacker News, linking an X post.
Nvidia prepares to fire CEO Jensen Huang, WSJ reports
225 points and 4 comments on Hacker News, linking an X post.
DeepSeek releases API for its apps along with open source front end
2 points and 0 comments on Hacker News, linking an X post.
Forensic evidence Chinese AI labs built their newest models on Nvidia Blackwell
6 points and 0 comments on Hacker News, linking an X post.
New repos, releases and changelog
| FrigadeHQ/yogurt | 1559 |
| HybridRobotics/supermimic | 1526 |
| agentes-chat/agentes-chat | 1457 |
| alexeyw/lot-map | 852 |
| duwqijlk/muse-protocol | 752 |
| Noniv/abyssal-garden | 686 |
A new repository with 678 stars in its first week.
A new repository with 672 stars in its first week.
A new repository with 670 stars in its first week.
A new repository with 665 stars in its first week.
A new repository with 666 stars in its first week.
A new repository with 678 stars in its first week.
A new repository with 672 stars in its first week.
A new repository with 670 stars in its first week.
A new repository with 666 stars in its first week.
A new repository with 666 stars in its first week.
Data centers, power, chips and what tokens cost
| Mon 09 | 35 |
| Tue 10 | 32 |
| Wed 11 | 31 |
| Thu 12 | 29 |
| Fri 13 | 29 |
| Sat 14 | 7 |
| Sun 15 | 1 |
Stories from DCD, Data Center Knowledge, EIA and Hacker News, by publication day.
| gemini-4-hydrogen | 7.5 |
| gemini-4-hydrogen:batch | 3.75 |
| ernie-light-dt-1.0-intl | 3.94 |
| gemini-4-carbon | 10 |
| gemini-4-carbon:batch | 5 |
| median, all paid models | 2 |
443 paid models listed; median of the 30 added in 30 days: $2.50. Rankings data by OpenRouter, CC BY 4.0.
Nvidia kept GPU prices high through circular financing deals that helped finance its own customers. ·
426 points and 216 comments on Hacker News, linking techspot.com.
Forethought analyzes launch costs, cooling physics, and engineering factors to determine how soon AI data centers could move to orbit. ·
Nvidia allegedly engineered scarcity by funding AI startups to buy its GPUs. The Wall Street Journal reports this strategy inflated the ongoing GPU and memory shortage. ·
Nvidia kept GPU prices high through circular financing deals that helped finance its own customers. ·
426 points and 216 comments on Hacker News, linking techspot.com.
Nvidia allegedly engineered scarcity by funding AI startups to buy its GPUs. The Wall Street Journal reports this strategy inflated the ongoing GPU and memory shortage. ·
Forethought analyzes launch costs, cooling physics, and engineering factors to determine how soon AI data centers could move to orbit. ·
Forensic analysis of public AI models reveals that DeepSeek, Moonshot AI, MiniMax, and Ant Group use NVIDIA Blackwell GPUs to train their systems. ·
Data centers pay 38% of energy infrastructure costs in 2025, while rural and low-income households cover the rest. ·