Alibaba Reportedly Restricted Claude Code (1 minute read)
Alibaba reportedly planned to prohibit employee use of Claude Code from July 10 after classifying it as high-risk software. Employees were instead directed toward Alibaba's Qoder tool amid Anthropic's efforts to prevent unauthorized access and model distillation.
|
OpenAI might be preparing GPT-5.6 for next week's release (2 minute read)
OpenAI has moved GPT-5.6 into a narrow preview, splitting it into three tiers: Sol, Terra, and Luna. A notable change includes a new reasoning-effort control slider and an "ultra" mode for handling complex tasks. The release has implications for Codex users, amid Anthropic's Fable 5 changes, while broad access depends on US government review approvals.
|
|
Closing the Verification Loop (14 minute read)
Agents have made building cheap, but the cost just moved to verifying whether things actually work. The verification loop is the distance between a claim and its proof. Most teams close this loop with humans, but agents are starting to outpace humans. Compound Engineering is a plugin with skills for coding agents. This post discusses how /ce-dogfood, a skill within Compound Engineering, closes the loop and the persona strategy that gives it eyes.
|
Clouded Judgement 7.3.26 - The End of Compute Scarcity? Not So Fast (14 minute read)
Meta and SpaceX selling compute capacity could mean that there's excess compute capacity that two of the largest buyers no longer need. This may result in downward capex revisions for the hyperscalers. It's possible demand simply isn't there for what's being built with AI. However, this is unlikely, as anyone willing to sell capacity finds buyers immediately.
|
Understanding the Dynamics of the AI Ecosystem with Pace Layers (5 minute read)
Pace Layers is a framework for organizing fields and categories by how fast they change. Each layer is a differently paced component that is functionally different from the others and operates somewhat independently, but each layer influences and responds to the layers closest to it in a way that makes the whole system resilient. This post applies the Pace Layers framework to the AI ecosystem to help readers process its dynamics.
|
|
jamesob's guide to running SOTA LLMs locally (12 minute read)
This guide discusses the hardware used to run SOTA AI models locally and how to run speech-to-text (STT) locally. It also provides a ready-to-run configuration for running models within Docker containers. A budget of $2,000 can get you a setup that runs Qwen and good STT. $40,000 can get you a machine that can run almost-Opus-level models.
|
A Field Guide to Fable: Finding Your Unknowns (12 minute read)
The difference between the map and the territory is the unknown. When Claude runs into an unknown, it needs to make a decision based on its best guess. The more work that is done, the more unknowns Claude might run into. This post discusses how to use Claude to help discover unknowns. Discovering your unknowns before starting a project is a cheap way to find out what you didn't know before it gets expensive to fix.
|
Open Source AI Gap Map (Website)
The Gap Map maps out the open-source AI stack so people can collectively build what's missing. The current open source AI stack is seriously robust, but fragmented, duplicative, and hard to see as a coherent whole. The Gap Map helps people identify where to build new, where to invest in capability, and where to open up the tools. It is an invitation for greater collaboration and coordination across the space.
|
|
Own the Loop: A Field Guide to Agent Harnesses (5 minute read)
As coding models commoditize, the real differentiator is the harness: the control loop that manages tools, workflows, orchestration, and model routing. Vendor-native harnesses offer the best performance today, but owning a portable, model-agnostic loop may become the more durable advantage as models evolve and costs fall.
|
A brief history of distillation in AI (4 minute read)
Distillation started as a way to compress large models into smaller, cheaper ones, but has evolved into a core post-training technique for transferring instruction-following and reasoning abilities from frontier models. Methods used by DeepSeek, Qwen, and GLM now sit at the center of both open model development and ongoing disputes over whether training on proprietary model outputs constitutes unauthorized copying.
|
|
Leanstral (12 minute read)
Leanstral is an open-source 119B-parameter theorem-proving and code-verification agent built on Mistral's general-purpose coding framework.
|
|
|
|
|
