In a series of significant developments, the AI industry is seeing major strides in large-scale language model (LLM) training and interactive AI models. These advancements, announced in early July, are set to transform how researchers and developers approach AI, making it more efficient, accessible, and interactive.
Miles, a new framework, is making frontier-scale LLM reinforcement learning (RL) post-training more manageable. This tool simplifies the process, making it easier to build, reproduce, and operate. As models grow larger and more complex, Miles ensures that they remain composable, reproducible, and scalable, even as they run across distributed and specialized hardware.
A recent article addresses the issue of GPU bubbles, where the GPU sits idle waiting for the CPU to complete its tasks. The solution, known as pipeline decoding, allows the GPU to start working on the next token while the CPU finishes the previous one. This technique significantly reduces idle time, improving the efficiency of AI model computations.
Thinking Machines, an AI research lab, is pioneering new interaction models that facilitate continuous collaboration between humans and AI. These models integrate interactivity directly into the AI, allowing for real-time feedback and redirection. The company plans to offer a limited research preview soon, with a wider release expected later this year.
Anthropic has launched Claude Science, an AI workbench app now available in beta for Pro, Max, Team, and Enterprise users on macOS and Linux. This integrated environment consolidates various scientific tools, enabling natively rendered 3D protein structures, genome browser tracks, and chemical structures, streamlining the research process.
The Department of Commerce has lifted export controls on Anthropic's Claude Fable 5 and Mythos 5. Access to these models will be restored starting tomorrow, with further updates anticipated soon.
Anthropic has also introduced Claude Sonnet 5, a lower-cost model with enhanced agentic performance in planning, tool use, coding, and knowledge work. Its capabilities are said to approach those of Opus 4.8, while significantly improving over Sonnet 4.6.
Ai2 has released MolmoMotion, an open vision-language model that predicts the 3D paths of objects in video sequences. Given a few seconds of video and a plain-language instruction, MolmoMotion forecasts the movement of marked points on an object. The model is open source, including its weights, training code, and a large dataset of action-described 3D point trajectories.
Subscribe to our newsletter for the latest AI news, tutorials, and expert insights delivered directly to your inbox.
We respect your privacy. Unsubscribe at any time.
Comments (0)
Add a Comment