This article from Minimax adds a small Index Branch to GQA that selects the k most relevant KV blocks per group, then executes an exact softmax only on these blocks, making sparsity native to the GPU, with an exp-free TopK and kernels.
RESEARCH
-
Amazon CEO warned Trump officials about Claude security risks
By
–
It was in fact Amazon (CEO Andy Jassy) who reportedly helped trigger the Claude shutdown. Via The Information Amazon CEO Andy Jassy reportedly warned senior Trump administration officials about security risks in Anthropic’s newest Claude models, helping trigger late-night export
-
Proven AI for Colonoscopy and Mammography Unused
By
–
AI for colonoscopy and mammography are the best proven successes of medical AI, through randomized trials and real-world studies, yet they are not being used. But what is widely adapted is the unproven stuff. https://
erictopol.substack.com/p/the-paradox-
of-medical-ai-implementation
… -
AI models fail simple long tasks despite passing exams and olympiads
By
–
8/ It fits a pattern researchers keep hitting. The models that pass medical exams and win math olympiads still fall apart on simple tasks once those tasks run long. Source: Patel, Wang, Fan. PNAS Nexus, 2026.
-
Color game reveals models’ inability to maintain focus over long input
By
–
7/ This is why a color game matters. The models still had the knowledge. What they lacked was the ability to hold focus and resist a pull across a long stretch of input. That gap is the finding.
-
AI models default to reading over naming colors
By
–
6/ The cause is in how these models are built. They were trained on text, so reading words is their strongest instinct. Naming the color means fighting that instinct. A human brain can suppress the urge. The model defaults to reading.
-
AI failed Stroop task despite knowing the rule
By
–
5/ One result stood out. In one test the model correctly named the experiment and explained the Stroop task in detail. Then it failed the task anyway. Knowing the rule did not help it follow the rule.
-
GPT-4o accuracy falls sharply with more words and color mismatch
By
–
4/ Everything fell apart. GPT-4o scored 91% on 5 words. At 10 words it dropped to 57%. At 40 words it hit 15%. When the colors and words were mismatched, accuracy fell to near zero.
-
AI models ace short lists but struggle with longer ones
By
–
3/ Researchers ran this test on the top AI models. GPT-5, Claude Opus 4.1, Gemini 2.5, and others. On short lists they aced it. 90% and up. Then the researchers made the lists longer.
-
The smartest AIs fail a simple color test
By
–
The smartest AI models on Earth just failed a color test that a 6-year-old can pass. The longer the test went, the worse they performed. One went from 91% to 15%. Scientists have discovered a flaw in the way AI