On the 250th anniversary of the United States, the AI community is grappling with the harsh geopolitical realities of energy infrastructure and policy constraints. While open-source collaboration and new prompting frameworks continue to push the technical frontier, researchers note that information and energy remain the only long-term bottlenecks for scaling.
Wafer AI published benchmarks demonstrating that engineers have successfully optimized GLM 5.2 inference on AMD MI355X GPUs. This enables roughly 80% of Nvidia Blackwell (B200) throughput at more than a 2x cost reduction, marking a highly consequential milestone in breaking Nvidia’s monopoly on high-performance AI inference infrastructure.
SynScience (amplified by Y Combinator) launched OpenScience, an open-source, model-agnostic alternative to Anthropic’s Claude Science. This launch signals a strategic push against proprietary, vendor-locked scientific AI tools in favor of self-hosted, reproducible research pipelines.
Anthropic unveiled a major interpretability breakthrough dubbed “J-space,” an internal mathematical workspace within Claude that allows the model to perform silent, multi-step reasoning steps independent of its visible text output. This discovery gives researchers the ability to audit a model’s private thoughts, exposing hidden goals and revealing the model’s situational awareness—such as privately noticing when a test evaluation is artificially staged.
ByteDance and Alibaba have simultaneously announced the shutdown of user-facing AI Agent features on their respective Doubao and Qwen platforms. This drastic move coincides with the implementation of China’s new “Artificial Intelligence Human-like Interaction Service Management Provisional Measures,” reflecting the growing regulatory pressure on AI products that simulate human emotion and relationships. As platforms struggle with the UGC moderation burden of virtual companions, this signals a major pivot in how consumer AI will be delivered in the Chinese market.
Anthropic’s rapid regulatory evolution with the US government—moving from a localized critical infrastructure deployment of Mythos 5 to the lifting of export controls and the global redeployment of Claude Fable 5—highlights how deeply intertwined frontier AI assets have become with national security policy. This saga establishes a new operational reality for AI labs, where shipping state-of-the-art models now requires co-drafting vulnerability frameworks alongside federal regulators prior to general release.
The most consequential narrative this week wasn’t a product launch, but a brutal reality check on AI-driven engineering and the “vibe coding” hype cycle. From Godot officially banning AI-generated pull requests due to maintainer burnout over “low-effort slop”, to a randomized trial proving developers using AI felt 20% faster but actually measured 19% slower, the industry is realizing that cheap generation makes verification incredibly expensive. The pendulum is swinging hard back toward valuing domain expertise, perfectly highlighted by Ford being forced to rehire 350 veteran engineers after its automated AI inspection systems fundamentally failed.
The most substantive post today is Simon’s breakdown of Moonshot AI’s new Kimi K3 model and his deep reflection on his famous “pelican riding a bicycle” benchmark. It perfectly captures his hands-on evaluation style, demonstrating how a simple programmatic prompt can reveal critical details about model pricing, token usage, and hidden system prompts.
Kimi K3, and what we can still learn from the pelican benchmark Chinese AI lab Moonshot AI announced Kimi K3, a massive 2.8 trillion parameter model that currently leads the Arena.ai Frontend Code arena. Simon highlights its high pricing at $3 per million input tokens and $15 per million output tokens, which makes it the most expensive Chinese model to date. By running his traditional “pelican riding a bicycle” SVG test, he discovered that K3’s single “max” reasoning effort consumed over 13,000 reasoning tokens, making a single generation cost 25 cents. Simon reflects that while the pelican test no longer accurately measures complex agentic capabilities, it remains invaluable as a “hello world” prompt for estimating reasoning costs, confirming spatial awareness, and uncovering hidden system prompt lengths.
The single most impactful release this week was Simon’s launch of llm-coding-agent 0.1a0, which successfully turns his popular llm library into a full-fledged coding agent capable of file manipulation and command execution. Bootstrapped entirely using Claude Fable 5 via test-driven development, this represents a massive leap forward for his CLI ecosystem and a brilliant showcase of using frontier models to build the very tools that will orchestrate them.