The thread critiques ZeroBench, a benchmark for evaluating visual reasoning in models. The author expresses dissatisfaction with the examples provided, arguing that they don't effectively measure a model's visual reasoning capabilities. The main complaint is that the tasks are overly simplistic, primarily involving basic counting of objects and performing simple arithmetic. For instance, one task requires counting the number of pens with caps, which the author feels doesn't challenge the models adequately. On a different note, there's a section predicting features for GPT-4. Expectations include a larger context window, ranging from 16,000 to 32,000 tokens, and the ability to use tools like web browsing and coding. The author anticipates a stronger emphasis on human feedback and user-generated data, improved data curation, and scaling laws. The prediction caps the model's parameters at 200 to 400 billion, indicating a focus on efficiency rather than sheer size. Overall, these insights highlight ongoing concerns about model evaluation methods and the evolution of AI capabilities.