Most teams evaluating agent memory compare storage options: vector database, graph, hosted service. The choice matters less than what you learn after wiring anything up, which is that storing is easy
I write this as someone who does not write code. A friend runs a pet supplies shop. Their support agents call customers 铲屎官 and themselves 小毛, and they never say "dear." Customers get answers in under
I came into this test with a lazy assumption: for translation, use the model that was built for translation. Twenty Japanese and Korean slang terms later, that assumption was gone. qwen-mt-uni got two
The more useful question is structural: how does the model divide the work, where does the output become unusable, and what does a report actually cost in tokens and minutes? I ran Bailian's qwen-deep
Real-time voice is usually sold as a demo: a person talks, a model talks back, everyone claps. The engineering question is less glamorous. Where does the latency come from, what does an interrupt actu
World models were the demo of the year. Google's Genie 3 looked incredible and callable by exactly nobody. Then I noticed Alibaba's Bailian had quietly listed HappyOyster, the model behind China's vir
Quick field notes on decision-model-preview, the model Alibaba's Bailian platform listed on September 24 with a spec-sheet oddity: maximum output length 0. It's a decision model built on the TypeSafe
Most model comparisons measure one build of a model against another. This post is about a rarer and more practical phenomenon: the same model behaving differently across deployment forms. The open-wei
On 2026-09-30 OpenAI announced Dot, an always-on agent: GPT-6 Astra, its own cloud computer, a dedicated browser, more than 4000 app integrations through plugins, sold as a virtual employee that repor