Monday, November 16, 2026 Edition No. 240
Search
Daily News Fully automated · 5 min read

AI News: November 16, 2026

Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled

1. Meituan Releases LongCat-2.5, a 1.6-Trillion-Parameter Agent Model, Under MIT License

Meituan. Meituan released the weights for LongCat-2.5 one week after its official unveiling. The model is a mixture-of-experts system with 1.6 trillion parameters, roughly 48 billion of them active per token, built for autonomous agent work and especially coding, and it is licensed under MIT. Meituan reports 88.4 on Terminal Bench v2.1, against 89.1 for Claude Opus 5 and 88.8 for GPT-5.6 Sol, and says its DeepSWE score rose from 9.7 to 64.6. A 4-bit deployment is described on 194 Huawei Ascend accelerators, and The Decoder says the model was trained entirely on Chinese hardware. The benchmark figures are Meituan’s own, and The Decoder notes DeepSeek-V4.1-Flash tops the same Terminal Bench table at 90.6. Source

2. WSJ: Megaspeed Arranged Nearly $12 Billion in Nvidia Chip Orders for Chinese Companies

Nvidia. Documents reviewed by the Wall Street Journal indicate that Megaspeed arranged nearly $12 billion in Nvidia chip orders for Chinese companies between March 2024 and June 2025, covering several hundred thousand chips the US sought to keep out of China, according to The Decoder’s summary. Megaspeed is owned by Alice Huang, described as a former partner of ousted Nvidia CEO Jensen Huang, and reportedly gained preferential access as an Nvidia cloud partner. Megaspeed told Reuters that Jensen Huang left the company on August 30, 2025 and that it conducted “legitimate and lawful business.” Short seller Culper Research estimates that more than 20 percent of Nvidia’s revenue last year came from Chinese customers, against 7 percent of revenue reported over six months. Claims that Huang acted as a spy remain unconfirmed. Source

3. Claude Fable 5.2 Produces Images by Writing Python That Drives a Simulated Paint Engine

Anthropic. Examples posted on X by Anthropic’s Jake Eaton and Michael Graczyk show Claude Fable 5.2 making images without an external image model, by writing 1,000 to 2,000 line Python programs that place simulated brushstrokes on a virtual canvas. Eaton’s photorealistic parking garage scene used about 700,000 strokes, and Graczyk shared a 21-second video of a bird painted in 111 strokes in which the model painted over its own mistakes. The Decoder reports that Fable 5.2 scored 64.9 percent on Anthropic’s internal RoboDojo Sim benchmark against 29.8 percent for OpenAI’s GPT-6 Astra, while in a real robot-arm painting test by researcher Aris Zhu, Astra was faster and had better fine motor control. These are employee posts and internal figures that The Decoder did not independently verify. Source

4. Reserve Bank of Australia Analysis Finds Employment Falling for Young Workers in AI-Exposed Occupations

Reserve Bank of Australia. An analysis by Mike Major at the Reserve Bank of Australia tracks entry-level workers in highly AI-exposed occupations such as programming, clerical work and customer service. Job postings have fallen across all groups since late 2022, with the steepest decline in the most exposed group, and employment of workers aged 15 to 24 in that group has declined while older workers’ employment has risen. Major cautions that the study does not establish causation and measures theoretical exposure to AI rather than actual use, and The Decoder notes that business-cycle shifts and remote work could also explain the pattern. The piece sets the result beside earlier work, including a Stanford study that found employment down about 13 percent in relative terms for 22 to 25 year olds in exposed jobs, and a Danish study of 25,000 workers that found no measurable change in wages or hours. Source

5. Sebastian Raschka Publishes a Walkthrough of Positional Embeddings From GPT-2 to 2026 Open-Weight Models

Sebastian Raschka. Raschka’s Ahead of AI newsletter published a long technical explainer on how LLMs encode token positions, covering GPT-2’s learned absolute embeddings, the fixed sinusoidal encodings of the 2017 Transformer and rotary position embeddings (RoPE) from the 2021 RoFormer paper, with a toy example and PyTorch code. He notes that extending GPT-2’s learned table to a 100,000-token context would make positional weights more than 22 percent of GPT-2 Medium’s parameters. The article also surveys positional choices in 2026 open-weight models, including MiniMax-M3 rotating 64 of 128 dimensions and SmolLM3 skipping RoPE in every fourth layer. The survey section was summarized rather than read in full during collection. Source