[ Archive ]
All technical deep dives, system architectures, and AI breakdowns, ordered by date.
Self-Hosting & Open SourcePangolin 1.22 puts an identity-aware proxy in front of OpenAI, Anthropic and Gemini as well as your own Ollama and vLLM servers, with virtual API keys and budget enforcement. It is also a clean case study in open-core: the gateway itself is AGPL-3, but the session logging is a 727-byte stub whose real 10 KB implementation sits under a proprietary licence header.
Open Weights & Code LLMsTencent's Hy team released Hy4 preview — 770B total, 49B active, 1M-token context, full weights under Apache 2.0 on Hugging Face with an FP8 variant and official vLLM/SGLang recipes. A preview with real limitations, but the most permissive flagship license in the open-field race so far.
Open Weights & Code LLMsTwo weeks after launch, Z.ai published the full GLM-5.3 weights on Hugging Face: 753B parameters in FP8, a 1M-token context, and a custom license that requires a security review before giant MaaS providers can use it commercially. Here is what is actually downloadable, and what it means.
Agent Systems & APIsMistral AI has released Agentic Search and the Mistral Search Toolkit, introducing an active five-tool inspection loop—search, open, navigate, read, grep—that boosts FinanceBench accuracy from 26.7% to 86.0% while slashing token usage by 33.7%.
Vision & Multimodal LLMsGoogle has released Gemini 3.5 Transcribe and Transcribe Live in the Gemini API, bringing 2.6% batch WER, full-duplex WebSocket streaming, speaker diarization up to three voices, and a built-in semantic engine that strips disfluencies and resolves self-corrections on the fly.
AI Economics & PricingThe Information reports Nvidia has agreed to acquire Hugging Face for $12.9 billion — a 3x jump from its 2023 valuation. Neither company has confirmed. What the deal would mean for the platform every open-weights developer depends on.
Vision & Multimodal LLMsGoogle moved its conversational video model from preview to general availability on August 27, adding clip extension, first/last-frame interpolation, and a resolution parameter — and put the preview endpoint on a September 30 deprecation clock.
Open Weights & Code LLMsAlibaba's Qwen team released Qwen3.8-Flash-Next as open weights on Hugging Face — a 125B-parameter MoE model with only 6B active per token, redesigned across attention, residuals, embeddings, and optimizer to preview the architecture behind the upcoming Qwen4 family.
Agent Systems & APIsThree OpenAI platform products hard-stop on November 30, 2026, and Evals goes read-only a month earlier. Here is what each one maps to, the code to write, and the order to do it in.
AI Economics & PricingHeadlines say Claude Sonnet 5 jumps 50% on September 1. Anthropic's own pricing page says the increase will not occur. Here is what OpenAI, Anthropic, and DeepSeek actually changed — and the tokenizer detail that breaks most price comparisons.
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.