Lesen

More than 90% of the 430,000 microdramas launched online in China in the first eight months of 2026 were AI-generated, according to the National Radio and Television Administration.

Microdramas run a few minutes per episode and rely on fast pacing and twists. AI has cut production costs so far that supply is saturated; creators say the challenge has shifted from making videos cheaply to standing out among lookalike productions.

The South China Morning Post reported on October 4 that NetEase this summer used AI to reconstruct actress Joey Wong's classic roles, a signal of the industry pivot. The report does not specify how the 90% share was measured, and the regulator's original release is not linked.

Tavus released Griffin, a full-duplex video interaction model, on October 1. In the company's own test, 48% of participants believed they were talking to a human; its previous system scored 2.4% on the same test.

Griffin abandons the cascaded pipeline of speech recognition, language model, speech synthesis and avatar rendering. A single model handles listening, speaking, expressions and pixels at once, generating 720p video in real time with an average response latency of 0.43 seconds. On NVIDIA's VideoFDB full-duplex benchmark, Griffin-Lite scored 3.83 on generation, close to the human reference of 3.92.

The 48% figure comes from a test Tavus designed itself: participants were led to believe they were on a call with a human, the sample was 54 people over one-minute calls, and the protocol was not a standard Turing test — community notes on X flag it as independently unverified. Those who grew suspicious mostly saw through it within 20 seconds, and Griffin-Lite is open only to a small set of trusted testers, not general users.

A Reddit user used an open-source tool to move part of a large model's layers onto an iPhone 17 Pro Max, cutting prefill time for Qwen3.8-27B on a 24GB MacBook Pro: at 16K context, speed rose from 109 to 157 tokens per second, a 44% gain.

According to IT之家's October 4 report citing Wccftech, the user kept the first 40 of every 256-token batch's layers on the Mac's M4 Pro and streamed activations to the iPhone's A19 Pro GPU for layers 41 to 64; older context is compiled onto the phone's Neural Engine, cutting single-token write time from 279ms to 176ms at 140K context. Gains were 35% at 8K and 29% at 32K.

The tool, called backburner, is open source on GitHub. The limits are clear: within 64K context the phone does not speed up text generation, which stays entirely on the Mac, so the benefit applies only to long-context preprocessing.

When training an AI model, you repeatedly fine-tune thousands of parameters along the gradient; how far each step goes is determined by the learning rate — too large and it oscillates, too small and it's too slow. Adam (proposed in 2015) is a method that sets the step size automatically: it records the magnitude and variability of each parameter's gradient, giving stable parameters large steps and jittery parameters small steps, and it is insensitive to overall gradient scaling.

Today, open any deep learning framework or tutorial and the default optimizer is most likely this one; new methods still use it as the comparison baseline when publishing papers. The rule it established — "hyperparameters barely need tuning" — has not been replaced to this day.

If you want theoretical convergence guarantees beyond convex problems, or want to use it to judge models and data the paper didn't test, don't use it to draw conclusions; evidence on long-term performance in non-convex deep learning settings is limited.

Adam: A Method for Stochastic Optimization (2015) | Next review 2027-09-20