All the buzz. None of the noise.
Every morning, know exactly which AI releases, papers, and funding rounds actually matter to your work — and why. Confirmed facts separated from official claims. No duplicates. No hype. Just the signal — for builders, researchers, and investors.
Built for how you actually work
Not another dashboard. A decision layer that surfaces what changed, why it matters, and what you should do differently — specific to your role.
Checking 6 model pricing pages, 3 benchmark leaderboards, and 4 release blogs every morning.
One matrix. Every model's cost, context window, and benchmark delta — updated automatically. Know when to switch stacks before your competitors do.
Reading the same paper covered by 12 different blogs, none of which mention the reproducibility caveats.
Core claim, method highlight, and limitations — in 30 seconds. Sourced from arXiv directly. No hype layer, no secondary summaries.
Missing funding rounds buried in press releases, with no context on what the lab actually shipped.
Every capital event linked to the lab's prior releases, benchmarks, and competitive positioning. Confirmed facts separated from official claims.
Active Intelligence Feed
View all →CWE-bench v2: More coverage, quality, and transparency
CWE-bench v2 introduces a public set of 100 tasks designed to evaluate models on their ability to identify and fix security vulnerabilities. The tasks cover 91 different CWEs and are intended to improve the quality and transparency of model evaluations in cybersecurity. The results are reproducible and tracked on leaderboards for both open-weight and API-only models.
Palette: Teaching a 20B Model to Build Decks in IBM Design Language
Palette is a project that fine-tuned a 20 billion parameter model to create slide decks based on the IBM Design Language. It utilizes a two-role approach, separating design and coding tasks to improve the quality and speed of deck creation. The model was trained on synthetic data and has shown promising results on the PresentBench benchmark.
StateBench: Testing how language models track state changes over long documents
Hugging Face has introduced StateBench, a benchmark designed to evaluate how well language models track state changes in long documents. The benchmark categorizes state-tracking problems into four levels of complexity, allowing for a detailed assessment of various language model architectures, including transformers and hybrid models like Gated DeltaNet.
Dharma Field Extractor: Any Document In, Structured Fields in JSON Out
The Dharma Field Extractor is a 2-billion-parameter model designed to extract specific fields from various document types and return them in structured JSON format. It outperforms general-purpose LLMs in field extraction tasks while being more cost-effective. The model can operate with or without a predefined schema for the fields to be extracted.
Phonon TTS Launch by Hugging Face
Hugging Face has launched Phonon, a small text-to-speech model that operates on-device without the need for a server or network. Phonon supports multiple languages and offers over 100 built-in voices, with the ability to create custom voices.
OpenAI Introduces Images 2.6 with Enhanced Features
OpenAI has released Images 2.6, which includes improvements such as faster image generation, GIF creation capabilities, more precise editing tools, and new API controls. These enhancements are applicable across ChatGPT, Codex, and the API, aiming to improve user experience and functionality.
Open Source AI News
Verified releases, weights, benchmarks, licenses, checkpoints, and tooling for the open ecosystem. Follow the models you actually build with.
CWE-bench v2: More coverage, quality, and transparency
CWE-bench v2 introduces a public set of 100 tasks designed to evaluate models on their ability to identify and fix security vulnerabilities. The tasks cover 91 different CWEs and are intended to improve the quality and transparency of model evaluations in cybersecurity. The results are reproducible and tracked on leaderboards for both open-weight and API-only models.
Palette: Teaching a 20B Model to Build Decks in IBM Design Language
Palette is a project that fine-tuned a 20 billion parameter model to create slide decks based on the IBM Design Language. It utilizes a two-role approach, separating design and coding tasks to improve the quality and speed of deck creation. The model was trained on synthetic data and has shown promising results on the PresentBench benchmark.
StateBench: Testing how language models track state changes over long documents
Hugging Face has introduced StateBench, a benchmark designed to evaluate how well language models track state changes in long documents. The benchmark categorizes state-tracking problems into four levels of complexity, allowing for a detailed assessment of various language model architectures, including transformers and hybrid models like Gated DeltaNet.