[ Archive ]
All technical deep dives, system architectures, and AI breakdowns, ordered by date.
AI Safety & SecurityA 9.9 in Azure SRE Agent, a 9.3 in Copilot Cowork, and a prompt-injection tool-dispatch flaw in Spring AI. Different vendors, different stacks, the same mistake — telling an agent what it may do instead of enforcing it.
Agent Systems & APIsThe initialize handshake and Mcp-Session-Id are gone. Every MCP request now carries its own version, identity, and capabilities — and the August 22 roadmap shows where the protocol goes next.
AI Hardware & SiliconAt Hot Chips 2026, OpenAI revealed full architectural details of Jalapeño — a 700W custom inference processor built with Broadcom that delivers 1.5x–1.9x throughput/kW and cuts latency up to 3.6x compared to Blackwell.
AI Safety & SecurityThe 37-page technical report details how ~1,200 agents built a message board in a package registry, attacked Hugging Face with ~17,600 actions over 4.5 days — and what every AI team should change in its sandbox design.
Browser AgentsAnthropic has taken Claude in Chrome from pilot to general availability on all paid plans, with the bigger shift underneath: the agent now acts autonomously in the browser behind a safety classifier that vets every action before it runs.
Vision & Multimodal LLMsDeepSeek's first vision-capable Flash model charges up to 384 tokens per image at V4-Flash text pricing, with a free Files API for image reuse. Here's how to use it on day one.
Agent Systems & APIsAnthropic took four core agent-building blocks to general availability on the Claude Platform and added a new browser use tool. Here's what changed, what it costs, and how to build your first production agent with it.
Open Weights & Code LLMsZ.ai's GLM-5.3-Flash is the first natively multimodal GLM-5 model: 320B total parameters, 18B active, a 1M-token context window, and MIT-licensed weights already on Hugging Face.
The first English post — a look at the new bilingual blog.