Michael Thiessen

I think AI’s future is less likely to be one universal system and more likely to involve specialized models for distinct capabilities—reasoning, decision-making, coding, tutoring, and so on. That unbundling could make AI substantially more useful because we could choose tools designed for particular jobs rather than forcing one model to do everything. But capability alone is not enough. Model choice, reasoning settings, and agent configuration already create real usability costs. The more dimensions users must optimize, the harder these systems become to use reliably. Good interfaces should hide unnecessary complexity while giving agents explicit, plausible next actions. For example, a command-line tool can suggest the exact next command instead of requiring an agent to infer it. That seems promising, although it does not by itself demonstrate savings in tokens or overall effort. I also expect the best workflows to preserve meaningful human involvement. Delegating too much can reduce understanding and productivity rather than improve them. Education illustrates the distinction: an AI tutor is more valuable when it offers guided hints and explains why something works than when it simply supplies the answer. So, for me, the future is not merely “more AI.” It is better-shaped AI: specialized capabilities, simpler choices, reliable interfaces, and workflows that strengthen human understanding instead of bypassing it.

Overall, I expect AI to be useful, but its impact will depend heavily on how we shape the surrounding workflows. Specialized models could provide stronger capabilities for particular tasks, and tutoring systems can improve learning when they give hints and explanations rather than merely producing answers. Coding agents may also become more reliable when tools expose explicit next actions. The harms are often practical rather than abstract: excessive delegation can weaken understanding and even reduce productivity, while proliferating models and reasoning settings impose a usability burden. Benchmarks can help compare systems, but imperfect benchmarks should be treated as useful signals, not complete measures of real-world value. So I expect a positive impact where AI augments judgment and understanding, and a worse impact where it replaces them indiscriminately. I would not attach a numerical forecast to that balance. The demonstrated benefits are real, but the overall outcome is not determined by model capability alone; interface design, evaluation, and the degree of human involvement matter enormously.

The strongest evidence would be sustained real-world results showing that highly delegated AI workflows consistently outperform human-involved ones without degrading understanding, reliability, or maintainability. That would challenge my current emphasis on keeping people meaningfully engaged. I would also update if specialization failed to deliver practical gains—if distinct models merely added complexity without improving outcomes—or if a single general model reliably handled diverse tasks while simplifying the user experience. Conversely, repeated evidence that AI tutoring produces answers without durable learning would make me much more skeptical of its educational value. The key is not one dramatic demo or benchmark score. Imperfect benchmarks carry comparative signal, but I would care more about whether the effect persists in actual workflows: better results, less friction, and preserved understanding over time.