Recent performance data regarding the Humanities Last Exam (HLE) indicates a significant upward trend in model capabilities, according to discussions on the social media platform Reddit. Users have noted sharp increases in scores, specifically attributing this progress to recent releases from Google and the model known as Fable 5.2.

However, the HLE benchmark has faced criticism from some observers regarding its structural composition and grading methodology. Some contributors on Reddit have argued that the current HLE framework is “broken” due to its heavy weighting toward post-graduate mathematics. Critics suggest that because mathematics possesses a rigid structure, it may not be the ideal metric for assessing general intelligence in the same way other subjects are.

Specific concerns have been raised regarding the disproportionate impact of certain subjects on the final score. For instance, engineering accounts for only 4% of the grade, and chemistry accounts for 7%. Critics argue that a model could demonstrate a total inability to interpret circuit diagrams or chemical processes without significantly impacting its overall score. Furthermore, some users have suggested that the mathematics section should be replaced by FrontierMath Tier 4 to better assess mathematical ability, or that the industry should shift toward the HLE Diamond benchmark instead.

Google DeepMind Unveils AlphaGenome 2

Google DeepMind has announced the release of AlphaGenome 2, a model designed to predict how DNA changes affect various cell types. This follows a series of recent technical releases from the company, including AlphaProof Code, Gemini Robotics 2.5, and Gemini 4 Carbon.

According to the company’s announcement, AlphaGenome 2 was trained on more than 100 million individual cell profiles across 68 different tissues. In benchmarks involving 28 immune cell types, the model demonstrated a significant increase in accuracy compared to the original AlphaGenome released in June 2025. Specifically, AlphaGenome 2 identified nearly twice as many variants affecting gene activity—reporting a 30% success rate compared to the 16% achieved by the previous fine-tuned version.

A notable technical feature of the new model is its ability to make predictions for cell states it has not previously encountered by utilizing new cell data, performing these tasks without the requirement of retraining the model.

Regulatory and Industry Commentary

In separate discussions regarding the AI sector, Palantir CEO Alex Karp has commented on the issue of regulatory capture. Karp suggested that industry players may be underestimating the impact of regulatory shifts, stating that if certain entities do not adapt, they may face significant legal challenges from clients.