Aleph Alpha has launched Kolibri-1, a 78-billion parameter model featuring 3.46 billion active parameters and an expansive 1-million token context window. Released under the permissive Apache 2.0 license, the model allows businesses and the open-source community to run it on private infrastructure.
Sources: HuggingNews, Reddit AI
North Korean official Kim Yo Jong announced that Pyongyang successfully integrated autonomous guidance systems into the flight test of an intermediate-range strategic missile. This marks a significant and concerning development in the military application of artificial intelligence for weapons systems.
Sources: HuggingNews
Microsoft has introduced FrogNano-4B-2609, a compact 4B parameter model derived from Qwen3.5-4B, specialized via reinforcement learning for repository-level software engineering tasks. Designed for the 'GPU poor', it uses the Leaf harness for structured tool calls and multi-turn coding trajectories.
Sources: Reddit AI
Major retailer Micro Center has started requiring customers purchasing prebuilt PCs with NVIDIA's flagship RTX 5090 GPU to sign no-export declarations and provide identification to curb unauthorized exports to restricted regions like China.
Sources: Reddit AI
YouTuber PewDiePie has released Ajax, a fine-tuned open-weight AI model designed to run locally as part of his Odysseus workspace. The model was created using model distillation and automatic abliteration to remove safety refusals.
Sources: Reddit AI
Cua has released Cua Spaces, a free and source-available virtualization platform designed for AI agents to operate real desktop applications across macOS, Linux, and cloud environments.
Sources: Vault Bookmarks
Amazon has introduced Bedrock AgentCore, featuring an AgentCore Gateway that allows Claude Desktop to connect to a managed Model Context Protocol (MCP) compatible Web Search tool. This bridges the model's knowledge cutoff gap while keeping query traffic securely within AWS infrastructure.
Sources: AWS ML Blog
Developers on the llama.cpp project optimized memory usage for the qwen4exp indexer by computing scores one head at a time and adopting the lightning indexer, cutting compute buffer requirements by half.
Sources: Reddit AI
Keerthana Gopalakrishnan, Research Lead for Gemini Robotics at Google DeepMind, returns to discuss the release of Gemini Robotics 2, whole-body intelligence, and the pursuit of generalist humanoid…
After leading Meta’s Llama models, Ahmad Al-Dahle is now transforming Airbnb with AI — from how its teams develop products to how it serves guests.
Founder and CEO of full-stack AI chip company Fractile, Walter Goodwin, joins Sarah Guo to discuss the bets he’s made on the future of the chip market as other major players like Broadcom, NVIDIA,…
We catch up with RLM first author Alex Zhang, MIT PhD, on Jev, PhD masxing, and the future of harnesses.
Tsung-Hsien (Shawn) Wen, CTO of PolyAI, tells Tim Scarfe why voice agents are harder than text agents. Voice adds time, and a good conversation depends on adapting to the person on the line, not just…
Fall of the Incas & Aztecs.
Nathan Labenz and Prakash Narayanan examine rapid shifts in technology, diplomacy, and medicine with guests Jeremie and Edouard Harris, Steve Hou, Joel Borgen, and Daniel McKinnon. The discussions…
Our DevDay coverage - the first pod on the DevDay lineup - dives in with the leaders of OpenAI’s CUA team and API platform.
Leonardo de Moura created Lean and co-created Z3. ---This episode is sponsored by Parallel.Parallel, where agents find answers: web search, extraction and deep research APIs built for AI agents.Start…
AI systems have gone from struggling with grade-school math to helping solve research problems that have resisted mathematicians for decades, including Navier-Stokes. In this episode, Greg Burnham,…