According to @DeepLearningAI, Google researchers introduced ATLAS, a transformer-like language model that replaces attention with a trainable memory module and processes inputs up to 10 million tokens; source: @DeepLearningAI. According to @DeepLearningAI, the team trained a 1.3 billion-parameter model on FineWeb and updates only the memory module at inference; source: @DeepLearningAI. According to @DeepLearningAI, ATLAS achieved 80 percent on BABILong with 10 million-token inputs and averaged 57.62 percent across eight QA benchmarks, outperforming Titans and Transformer++; source: @DeepLearningAI. According to @DeepLearningAI, the source does not mention cryptocurrencies, but the reported long-context benchmarks and memory-augmented inference provide concrete performance data that traders can track when assessing AI-related market narratives; source: @DeepLearningAI.

According to @DeepLearningAI, researchers found a method to estimate how many bits a model memorizes from its training data. Source: DeepLearning.AI on Twitter, Aug 28, 2025. In tests on hundreds of GPT-2-style models trained on synthetic data and FineWeb subsets, the approach compares the negative log likelihood of a trained model to a stronger model. Source: DeepLearning.AI on Twitter, Aug 28, 2025. The post did not provide performance numbers, release details, or market implications, so no direct crypto trading signal is indicated. Source: DeepLearning.AI on Twitter, Aug 28, 2025.