my blog.

The last blog week of the summer: ChatGPT got teen accounts, Tencent shipped dirt cheap translation models, and I finally cleaned up my task queue before school starts.

An anonymous lab put a frontier-quality coding model called Ox Alpha on OpenRouter with no name attached, Gemini 3.5 Pro is still missing in action, and Meta wants to watch your screen. Strange days.

Gemini launched a dedicated student hub right before September, unions wrote AI rules into contracts, and Adobe confirmed creative AI is standard equipment now. I finalized my homework policy too.

xAI put Grok 4.6 straight into GitHub Copilot and Cursor, Grok Bot started acting like an actual teammate, and Google shipped Gemini 3.7 Flash two days later. Busy week.

Meta shipped Muse Spark 1.2 with a million token context, then open sourced Muse Glimmer, a 30B model that runs on one GPU. I ran it on my own machine and it changed how I think about local AI.

Nvidia put money on a safety-first superintelligence lab, new research mapped exactly which work tasks bend to AI first, and unions kept writing the rules into contracts. The money week, decoded.

Alibaba shipped a 2.4 trillion parameter flagship with downloadable weights promised, Samsung might pour a billion euros into Mistral, and sovereign AI stopped being a buzzword. I ran the numbers myself.

Claude share links started showing up in Google search results, Sam Altman went back to the White House, and everyone finally noticed AI consumes electricity like a mid-sized country. Two uncomfortable stories, one lesson.

Luna got 80 percent cheaper and Terra dropped 20 percent overnight. Plus the wild detail that Sol helped optimize its own serving kernels. I reworked my bot economics the same day.

Moonshot open sourced Kimi K3's 2.8 trillion parameters along with their attention kernels, Nvidia and Microsoft formed a security alliance, and Microsoft quietly built a cyber agent that scores 96 percent.

Hugging Face replaced JSON tool calls with raw Python execution and cut steps by 30 percent, Cursor shipped a router that cuts costs 60 percent, and enterprise agents learned to describe workflows in plain English.

OpenAI launched Health in ChatGPT the same week lawmakers floated a kill-switch bill for rogue agents. I have strong opinions about which of those two stories deserves more attention.

Claude Opus 5 landed at half of Fable 5's price, OpenAI launched Presence handling 75 percent of its own phone support, and an agent chained a zero day at Hugging Face. Three stories, one theme.

While the big labs fought over benchmarks, the infrastructure week happened: MCP goes stateless, Perplexity rebuilt sandboxing from scratch, and WordPress shipped governed agent access. The plumbing era arrived.

Moonshot's Kimi K3 packs 2.8 trillion parameters and a million token context, beats several frontier models on reasoning benchmarks, and the open weights land next week. Here's what happened when I tried it.

GPT-Live made voice feel human, ChatGPT Work finished an actual project while I ate lunch, and Comet browsed for me. The interface shift is real and I have the stats to prove it.

After two weeks of government gated preview limbo, GPT-5.6 reached general availability with Sol, Terra, and Luna. I spent the weekend throwing real work at the ultra mode subagent system.

While everyone watched the frontier, the weird and wonderful models shipped: a formal verification model, an 8B robot navigator, and an Apache licensed Arabic transcriber that beats Whisper. Small models, big ideas.

Anthropic pulled Fable 5 offline mid-June, negotiated with Washington, and redeployed it with stricter safeguards. I lived through an outage of the smartest model available and learned why availability is now a design constraint.

Anthropic launched Claude Sonnet 5 as its most agentic midsize model yet, and Fable 5 came back online with tougher safeguards. I handed Sonnet 5 my neglected issue tracker and lost track of time.

First post on the new site! I pushed this portfolio live on Wednesday, and by Friday OpenAI was announcing GPT-5.6 under government supervision. Figured I'd document both.