According to NVIDIA’s research, software frameworks rather than underlying models will be particularly important in solving artificial intelligence problems in the long run.
Photo credit: Steve A Johnson/unsplash.com
In this study, researchers used a special software shell for efficient memory management, supplemented by a “supervisor”-like component, to ensure that Claude Opus 5 achieved a 100% score on the interactive ARC-AGI-3 test, a set of 2D games without instructions in which the model must learn how to play and win like a human. Without the case, the Opus 5 scored 30%, the best result of any model tested.
The results confirm that while model choice is important, it still accounts for a much smaller proportion of agent systems than many imagine, especially for long-term problems. The control system makes the model an agent: it handles memory, context, and feedback. Figuring out how to make artificial intelligence perform tasks within a long-term planning horizon without getting distracted or getting stuck in one place is one of the main goals of artificial intelligence agent research.
“Generally speaking, the world thinks of proxies as API models,” Notify resources TechCrunch Adel El Hallak, vice president of product for Nvidia’s artificial intelligence division, noted that agents now mean more. “It’s a model. It’s the structure around the model, which we call enablers, which are a set of tools that you use. We give you access to a runtime and the skills and libraries associated with it.” ——Halak said. He noted that it is also important to bring in a supervisory agent in addition to the primary agent who performs the work. He said he “acts almost like a CEO who pushes agents when they go off track or start exploring a path that could lead to a dead end, or re-explore a path that has already been taken.”
OpenAI’s model scored less than 10% on the ARC-AGI-3 benchmark, and the company conducted its own research last month and found that just changing two parameters in the test environment could double the model’s performance.
If you find an error, select it with your mouse and press CTRL+ENTER.
