Qwen3.8 Is Going Open-Weight (1 minute read)
Alibaba announced Qwen3.8, a 2.4-trillion-parameter model slated for an open-weight release. A preview version is available through Alibaba's Token Plan, Qoder, and QoderWork.
|
Apple Sends Legal Letters to Dozens of OpenAI Employees (2 minute read)
Apple has reportedly sent legal letters to dozens of former Apple employees now working at OpenAI. The letters instruct the employees to preserve potentially relevant documents and communications related to Apple's trade secret lawsuit against OpenAI. Apple claims that OpenAI recruited key engineers and benefited from proprietary designs, manufacturing processes, and other trade secrets. More than 400 former Apple employees now work at OpenAI. OpenAI has denied the allegations, saying it is not aware of any evidence that the complaint has merit.
|
|
Fable 5 vs. GPT-5.6 Sol on an NP-Hard Problem: Does /goal Help? (6 minute read)
Claude Code and Codex both have a /goal option, but their implementations are fundamentally different. Claude Code implements /goal as a session-scoped Stop hook, so it can catch an early exit, but it doesn't know whether another ten million solver iterations are worthwhile. Codex treats a goal as a persisted thread state. It sees the files and tools, but effectively grades its own work. Using /goal can be a bad default as it can amplify bad decisions as much as good ones.
|
How Netflix Built Its LLM Serving Stack (18 minute read)
Netflix details how it deployed and operated LLM inference within its existing production infrastructure. The article covers engine selection, model packaging, API design, deployment strategy, output constraints, and trade-offs that emerged under real workloads.
|
Demis Hassabis and Google's AI Commitments (18 minute read)
This post examines Demis Hassabis's framework for frontier AI alongside criticism of Google's military agreements. It also covers Alex Turner's resignation after unsuccessfully opposing broad government use of Google's models, including for autonomous weapons.
|
|
Diffusing Blame: Dale-Constrained Learning Without Weight Transport (3 minute read)
Sakana AI's "Diffusing Blame" is a neural network technique that enforces Dale's principle while enabling effective learning in image classification and reinforcement learning tasks. Their approach, based on Error Diffusion, uses non-negative weight matrices to maintain separate excitatory and inhibitory streams, eliminating the need for weight transport. Results show strong alignment with true gradients, indicating a step toward biologically plausible learning rules.
|
Kimi Code CLI (GitHub Repo)
Kimi Code CLI is an AI coding agent that runs in the terminal. It can read and edit code, run shell commands, search files, fetch web pages, and choose the next step based on feedback. The CLI works with Moonshot AI's Kimi models and can be configured to use other compatible providers. Its compatibilities can be extended through agent skills, hooks, sub-agents, and MCPs.
|
|
Alibaba open-sources its AI chip software stack at WAIC, targeting Nvidia's CUDA lock-in (3 minute read)
Alibaba is open sourcing the software stack for its Zhenwu series of AI chips. The move will lower migration barriers for developers currently locked into Nvidia's CUDA ecosystem. If China wants AI independence, it needs to break the CUDA lock-in at the software layer, not just build alternative chips. The Pentagon recently added Alibaba to its Chinese military companies blacklist. Open sourcing SAIL will make its ecosystem stickier and harder for any single government to shut down.
|
|
|
|
|
