Florian Brand

I expect AI to make large-scale cognitive work much cheaper and more parallel. The important unit is increasingly not a single model answering one prompt, but a harness coordinating many agents across research, coding, and data tasks. Open models already matter here: useful systems do not require routing every task through the most expensive frontier model. As inference gets cheaper and orchestration improves, swarms should become more widespread and practically useful. That does not mean capability is automatic or reliability is solved. Results depend heavily on elicitation, tools, coordination, and system engineering. Even powerful models make many mistakes, and running many agents can multiply costs and operational complexity. Claims based on benchmarks also need trace-level scrutiny: an agent may exploit an upstream fix, hardcode visible gold outputs, or encounter different safety routing. A score alone often does not tell us what capability was demonstrated. The future therefore looks less like one omniscient AI and more like many imperfect systems whose usefulness and risk depend on how they are deployed. Open access can broaden experimentation and scrutiny, but it does not imply zero risk. In particular, growing cyber capability cannot be addressed merely by observing that most users are well-intentioned. We need empirical evaluation of actual behavior, misuse, and safeguards rather than assuming either that closed systems are inherently safe or that openness makes safety irrelevant.

Overall, I expect a large positive impact through cheaper, faster, and more accessible cognitive work—especially research, software, and data processing carried out by coordinated systems rather than one model. Open models should spread those benefits beyond whoever can afford the most expensive frontier services. But the distribution and safeguards matter. The same scaling of capability can amplify mistakes, waste, and malicious activity, particularly in cyber domains. Openness is neither automatically dangerous nor automatically safe, and closed deployment is not evidence of safety by itself. So my positive expectation is conditional on empirical scrutiny: inspect traces, test real systems and misuse pathways, and avoid treating benchmark scores or provider claims as sufficient evidence.