AI posts claim big model wins
- Social posts this weekend bundled three separate AI claims into one hype cycle: Emory physicists’ dusty-plasma paper, GAIR’s ASI-Arch repository with 106 model designs, and a 2024 Alzheimer’s CNN-GCN study. - The headline numbers were real but narrower than the posts suggested: Emory reported force descriptions above 99% accuracy, ASI-Arch logged 1,773 experiments over 20,000 GPU hours, and the Alzheimer paper’s 100% came from one best fold. - None of the three claims was a brand-new weekend paper release; they came from a 2025 PNAS paper, a 2025 arXiv preprint and GitHub repo, and a February 2024 journal article. (pnas.org) (arxiv.org) (mdpi.com)
A burst of weekend AI posts stitched together three older research claims and presented them like fresh wins. (pnas.org) (arxiv.org) (mdpi.com) The first claim came from Emory University physicists studying dusty plasma, an ionized gas with tiny particles suspended inside it. Their paper says a machine-learning system inferred interaction laws in that system and described the forces with more than 99% accuracy. (news.emory.edu) (pnas.org) That work was not a brand-new weekend drop. Emory says PNAS published the findings in 2025, and the paper frames the result as a physics-specific method built from lab data on particle motion, not a general proof that AI is discovering universal laws on its own. (news.emory.edu) (pnas.org) The second claim centered on ASI-Arch, a system from the GAIR-NLP group that searches for new neural-network designs. In plain terms, it is software that proposes model blueprints, turns them into code, trains them, and keeps the ones that score best. (arxiv.org) (github.com) The 106 number is in the authors’ materials, but it comes with boundaries. The arXiv paper says ASI-Arch ran 1,773 autonomous experiments over 20,000 GPU hours and discovered 106 state-of-the-art linear-attention architectures, which is a narrower category than “106 novel neural architectures” in the broad sense. (arxiv.org) (github.com) That project also appears as a July 24, 2025 arXiv submission and a public GitHub repository, not as a peer-reviewed journal paper released this weekend. The repository says the 106 architectures are open-sourced, while the paper argues the system can scale architecture discovery with compute. (arxiv.org) (github.com) The third claim involved Alzheimer’s disease classification, where models sort brain scans into disease stages. Here the viral number came from a February 1, 2024 MDPI paper by University of Texas at El Paso authors describing a combined convolutional neural network and graph convolutional network, or CNN-GCN. (mdpi.com) The paper used 6,400 magnetic resonance imaging scans from the Alzheimer’s Disease Neuroimaging Initiative and reported 100% overall accuracy for the CNN-GCN model on the best fold of a five-fold cross-validation setup. The same results table listed 99.06% for the standalone graph model, 71.17% for VGG16-based transfer learning, and 43.83% for a scratch CNN. (mdpi.com) That is a more limited claim than “100% accuracy on a test set” suggests in a casual post. The article says the authors highlighted the best fold rather than an average across all five folds, which makes the result harder to compare with papers that report mean performance on held-out data. (mdpi.com) Taken together, the weekend posts pointed to real documents, real repositories, and real numbers. What they did not show was a single new event on April 26, 2026, or a new round of peer-reviewed confirmations that would turn three separate claims into settled science. (pnas.org) (arxiv.org) (mdpi.com)
