Nathaniel Whittemore reports on Google's Gemini 4 Carbon release and its reported benchmark gains, while noting that limited independent access makes practical assessment premature. He also discusses cloud earnings, infrastructure spending and reported antitrust scrutiny of Nvidia.
Whittemore explains that Artificial Analysis revised its Intelligence Index to include more private datasetsAn evaluation set is a collection of examples kept for measuring an AI system rather than training it. and tasks that reflect agentic coding and scientific work. The changes are intended in part to make it harder for training on known benchmark material to inflate scores.
The new mix changes the rankings and the reported gap between proprietary and open-weight modelsAn open-weight model makes its trained parameter values available so others can run or adapt it under the release's licence.. Whittemore discusses the differences between overall scores and individual task results, as well as the wide variation in reported costs per taskCost per completed task measures the total AI, tool, infrastructure, retry, and repair expense for each verified useful outcome..
Whittemore cautions that the weighting of science and coding reflects choices about which uses matter. He suggests examining the benchmarksA benchmark is a defined test or collection of tasks used to compare a system's performance under specified conditions. relevant to a particular task and considering model effort settings and costs rather than treating one aggregate score as definitive.
Watch on YouTube
