Claude Introduces Message Threads (2 minute read)
Anthropic released message threads in public beta, allowing agents to send only new messages instead of resending their entire conversation history. The feature reduced median time to first token by approximately 35% for requests with 500Kโ1M tokens, without changing token usage.
|
|
The Hitchhiker's Guide to Training on Compute Islands (18 minute read)
Being able to train reliably on multiple smaller clusters widens the usable training pool for frontier training. A single run could mix accelerator generations and vendors as each island runs on its own schedule. Older generations or less common chips could also participate in a frontier run. One risk with this model is that integrity and safety may become harder when models are trained across providers, as one faulty or compromised site can quietly degrade or tamper with models in ways that are harder to trace.
|
How to Train a Swarm (19 minute read)
Swarms are AI systems where multiple agents cooperate towards a shared goal. This post discusses a technique called collective inference-time scaling, where agents scale the amount of useful work performed in parallel and the communication needed to achieve it. It also explains how to design incentives for parallel work, and covers the general topic of designing incentives for reinforcement learning.
|
Lessons from building Copilot Cowork (21 minute read)
Microsoft's Copilot Cowork, an AI-driven agentic execution layer, achieved global availability in four and a half months, switching its runtime from Python to Rust while live. Key practices included writing clear specs before coding, running narrow agents with clear boundaries, and ensuring automated checks genuinely verify code changes. The transition to Rust improved efficiency and highlighted the importance of clear testing, boundary setting, and human oversight in the development process.
|
|
GPU-accelerated Weight Teleportation (17 minute read)
GPU-accelerated Weight Teleportation optimizes RL training by improving weight syncing between trainers and samplers, reducing pause times to under 0.23 seconds. This is achieved using a co-designed compression/decompression pipeline with GPU-accelerated parallel processing, utilizing efficient codecs like LZ4 and Snappy, combined with a Zstd layer.
|
Harness 2: executing agent code as it streams (6 minute read)
Framer's Harness 2 significantly upgrades the agent's execution process, allowing entire workflows in one program, achieving up to 33% cost reduction and 20% faster execution. This is done by executing JavaScript as it streams in, avoiding multiple model calls and using eager execution and early cancellation to enhance efficiency. The redesign also replaces DSL with method calls for real-time updates, making task execution streamlined for both in-app and external agents, reducing tool steps by 43% and cached input tokens by 39% per message.
|
NVIDIA's Cosmos Framework (6 minute read)
NVIDIA introduced new Cosmos Framework capabilities for training, fine-tuning, reinforcement learning, and optimizing world models for physical AI. Developers can generate synchronized video and LiDAR simulations for robotics and autonomous vehicles, then deploy models through GPU-accelerated Cosmos NIM microservices.
|
|
OpenAI is finally merging ChatGPT's Chat and Work tabs (2 minute read)
OpenAI plans to merge the Chat and Work tabs within the ChatGPT experience after an initial testing period with select users. Codex will remain a separate experience in the ChatGPT desktop app. The change will enable users to move from exploring ideas to having the chatbot complete work within the same space. Users can import files and plugins. Both ChatGPT and Codex will soon draw from the same allowance pool.
|
LLMs are not superhuman at reading and writing neural network weights (4 minute read)
A recent paper demonstrates GPT-6 Astra's ability to interact with neural network weights directly, surpassing human or conventional algorithm capabilities. Astra significantly improved a 0.8B Qwen LLM's mathematical skills by editing weights natively. While impressive, these capabilities reflect advanced problem-solving akin to skilled human researchers, not radically new AI capabilities.
|
|
Teaching Qwen Why (47 minute read)
The research extends Anthropic's "Teaching Claude Why" methods to new models, focusing on reducing agentic misalignment through synthetic document finetuning (SDF) and supervised finetuning (SFT) with fictional stories and ethical dilemmas.
|
|
|
|
|
