Skip to content

Technology

Data & training Latest news

Datasets, synthetic data, pretraining, post-training, compute and training costs.

4 selectionsPast 30 days 3items Total indexed: 30items

Updated

Data & training selections

TodayNov 17, 2026TueIssue 1–4 items
  1. ChinaTalk76

    BIS issues 132-page rule to curb China's AI compute access

    The US Commerce Department's Bureau of Industry and Security released a 132-page export control rule on Monday targeting three channels for China's AI compute: license requirements for chip exports to Singapore, Malaysia, Thailand, Indonesia and Brazil; caps on compute rented to Chinese firms; and a renewed block on Chinese chip designers using TSMC. ChinaTalk estimates remote access supplies about two-fifths of China's AI compute.

    Why it matters: The rule's remote-access cap and TSMC restriction could meaningfully slow Chinese AI training, but an exemption for compute inside the US leaves a major loophole that enforcement will determine.

  2. SiliconANGLE · AI78

    Alibaba launches Qwen4-Max with training and attention optimizations

    Alibaba released Qwen4-Max, its newest flagship LLM, built on a Qwen4 architecture that made training more efficient. Its backbone has 2.4 trillion parameters but activates only 71 billion at a time, supported by 328 billion N-gram parameters acting as a cache. Alibaba says it used up to 35% fewer tokens than Qwen3.8-Max on an internal benchmark and 49% fewer on Terminal-Bench 4.0.

    Why it matters: Token savings of this size on coding and knowledge-work tasks bear directly on inference economics, and Alibaba's plan to open-source smaller Qwen4 models could widen adoption beyond its cloud.

Nov 10, 2026Tue
  1. Amazon Science60

    Amazon releases Mitra-v2 tabular foundation model, open-sourced on Hugging Face

    Amazon Science introduced Mitra-v2, a second-generation tabular foundation model for classification and regression, pretrained entirely on synthetic tables with 10 times as many rows as Mitra-v1. It leads the single-model field on TabArena's 51 real-world datasets and ranks first on ScoringBench for regression. Fine-tuned models and inference code are released under Apache-2.0 on Hugging Face.

    Why it matters: Open, synthetic-only pretraining that tops public tabular benchmarks could reduce per-table feature engineering and tuning, a common cost in enterprise ML pipelines.

Oct 13, 2026Tue
  1. Amazon Science75

    AWS priming cuts hybrid model pretraining cost by over 99%

    Amazon Web Services researchers introduced priming, a method that reuses pretrained Transformer weights to build hybrid state space models, cutting pretraining cost by more than 99% and using under 0.5% of the source model's token budget. A 32-billion-parameter Gated KalmaNet hybrid matched its source Transformer on reasoning while decoding up to twice as fast at 128,000-token context. Code and models were released under Apache 2.0.

    Why it matters: Cheap, controlled architecture experiments could accelerate the shift away from pure Transformers for long-context agents, where KV-cache memory costs dominate serving economics.