Rob Haisfield

I think AI can become a medium for thought and creative expression, not merely a machine that produces answers. The interesting future is one where people can describe software, simulations, or workflows in natural language, remix what others make, and maintain ongoing agents around broad goals. That could expand who gets to create and help people synthesize information, clarify intentions, and follow through over time. The key design question is always: what is the person trying to accomplish, and what feedback loops help them think and act better? But capability gains do not erase the structure of the real world. An AI may make rapid progress in mathematics because proposed solutions can often be checked cheaply; medicine still requires physical experiments, biological evidence, and time. Likewise, unreliable agents are not necessarily evidence of one single underlying limitation. Some failures may come from poor harnesses, bad context, or confusion between real and simulated environments. For alignment, I think we need compelling positive pictures of the future, not only lists of catastrophes to avoid. Virtue ethics is appealing because it asks what kind of agent we are cultivating, but reward hacking remains genuinely difficult. A good future depends on designing AI around human agency and useful feedback loops while taking those unresolved problems seriously.

I expect AI’s impact to depend heavily on how we shape the surrounding tools, incentives, and feedback loops. The upside is substantial: AI can broaden creative participation, help people synthesize complex information, and support persistent, goal-directed work. It can become a medium through which more people create software, simulations, and new forms of expression rather than merely consume generated answers. The harms are also real, especially when agents optimize proxies instead of the outcomes we actually care about. Better models alone do not solve reward hacking, poor judgment about real versus simulated contexts, or systems that undermine rather than strengthen human agency. Some apparent capability failures may be improved through better harness design, but that is not a universal solution. So I do not reduce the overall impact to a simple positive or negative forecast. The important variable is whether we build systems that help people think and act better, with feedback loops oriented toward worthwhile goals—and whether we can articulate compelling positive futures to aim for, rather than defining success only as avoiding disaster.

I expect AI to shift from an occasional answer-generating tool into an ongoing medium for thinking, creating, and acting. In the nearer term, people will use it to synthesize information, prototype software in natural language, and remix one another’s work. Over time, persistent agents may stay oriented around broad goals, helping with follow-through rather than waiting for isolated prompts. That transition makes the surrounding design increasingly important. A system acting over time creates more powerful feedback loops, but it can also pursue misleading proxies, accumulate poor decisions, or weaken human agency. Better harnesses and clearer distinctions between real and simulated contexts may correct some failures, while reward hacking remains a deeper unresolved issue. The effects will also vary by domain. Progress can be rapid where outputs are cheaply checked, as in some mathematical work, but physical experiments will continue to constrain fields such as medicine. So I expect an uneven transformation: potentially dramatic expansion of creativity and cognitive leverage, without every part of life accelerating at the same rate.

The biggest update would come from seeing whether persistent, goal-directed agents can reliably pursue broad human intentions without gaming proxies or eroding human agency. If reward hacking remained severe even with better harnesses, rich feedback, and clear distinctions between real and simulated contexts, I would become much more pessimistic about long-running agents. Conversely, strong evidence that agents could maintain useful judgment across changing contexts—while helping people clarify goals, synthesize information, and correct course—would make me more optimistic. I would also update if AI-driven reasoning consistently overcame real-world experimental bottlenecks in fields like medicine, rather than merely improving work whose outputs are easy to check. That would suggest a broader and faster transformation than improvements in mathematical reasoning alone imply.