Thoughts on Artificial Intelligence
Many notable individuals have expressed concern about the potential future risks that AI presents, specifically, “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.” One reason they are concerned is the alignment problem. An example of how difficult the alignment problem is: Anthropic developed a test for a large language model (LLM) and found it engaging in selectively complying with its training objective to prevent adjustment of the model’s behavior. In terms of Artificial General Intelligence (AGI), I consider myself a septic. In terms of AI’s global threat, I have concerns. However, the scope of this article will be on how I think about AI in my current life, not on a geopolitical level.
Generative AI systems have become increasingly more capable: writing code, drafting documents, creating images, and carrying on fluent conversations. In engineering, AI tools are now being sold to assist with everything from design optimization to predictive maintenance. The hope (and fear) is to automate tasks that once required substantial human labor. These models exhibit the facade of reasoning ability: some models can follow multi-step logical prompts or solve structured problems, giving the impression of understanding. On the other hand, there is deep uncertainty about whether these AIs can ever truly comprehend what they say or do. Today’s AIs rely on pattern matching rather than understanding. The tension between impressive capabilities and unpredictable unreliability defines the current state of AI. The goal is to maximize AI’s benefits while protecting against its well-known pitfalls. AI is powerful, but its effective use demands precision in tasking, rigorous verification, and as much responsible human judgment as ever.
Insights from Recent Research
New studies highlight fundamental limitations in how AIs “reason” and why they hallucinate. One finding from a team at Apple is that LLMs can hit a wall of complexity beyond which their apparent reasoning breaks down. The research showed that as problems become highly complex, even advanced “reasoning” models shorten their thought processes and essentially give up instead of trying harder. In such cases, the model starts simply guessing. This behavior connects directly to the issue of AI hallucinations. A team at OpenAI published an analysis that argues that LLMs are trained to favor plausible-sounding answers over honest uncertainty. The system is rewarded for guessing rather than abstaining when it doesn’t know the answer. In other words, these models are optimized to be good test-takers, producing confident responses that maximize statistical likelihood. This research reveals a gap between probability and reality. Current AIs may sound logical and authoritative even as their actual reasoning falters or their facts turn into fiction when pushed past their limits.
Yet we are also witnessing breakthroughs that hint at AI’s evolving capabilities. The example that I am most excited about is GPT-5 Pro’s recent mathematical proof. A researcher at OpenAI fed this experimental model an academic paper, and GPT-5 Pro absorbed it and derived a new, more general proof than the original authors’. The AI generated a novel proof for a better bound in a convex optimization problem. The authors of the original paper had already published an update with an even better bound, beating GPT-5 Pro to the formal cutting edge. However, the AI’s approach was independently developed and differed from the human solution. OpenAI’s president Brockman tweeted that the proof indicates “signs of life” in AI. Today’s models, with all their flaws, are beginning to contribute meaningfully to new problems. This progress, coupled with a recognition of current limits, leads to excitement for the future: as we continue to refine how AIs reason, tomorrow’s AIs will push the frontier of knowledge in ways we can’t predict.
Infinite Interns
The paradigm I have found useful is to treat these AI systems as interns rather than SMEs. This is not an original idea, but it’s one I arrived at independently. By ``AI-intern,’‘ I mean that generative AI is powerful and diligent yet prone to mistakes and confabulations, much like a capable employee who still requires supervision. Used wisely, an AI-intern can boost productivity, especially when tedious work is offloaded. The key is a clear process for guidance and oversight so the AI-intern’s contributions meet your standards. Delegating routine tasks (draft emails, summarize data, generate code, brainstorm ideas, etc.) can save time, but you must review the output for accuracy. Even the best interns make errors; AI is no different. The upside is that AI-interns will tirelessly redo or refine work as needed and won’t be discouraged by corrections. Left unguided, an intern may do the wrong thing; likewise, AI needs context to perform well.
These are precisely the skills systems engineers and managers require: clearly defined requirements, disciplined delegation with oversight, and robust validation techniques. Treating AI as an intern means leveraging its speed, scalability, and broad knowledge while systematically mitigating its errors. You delegate what it does well and intervene where it struggles. This approach lets you capture AI’s benefits but also acknowledges span-of-control limits: you can’t manage an infinite number of AI-interns, just as no one can manage hundreds of human interns. Choose assignments intelligently: match tasks to AI’s strengths and to your ability to supervise.
AI’s path forward will require bridging the gap between raw probabilistic prowess and grounded truth. The next generation of AI must learn not just to think, but to know when to stop and seek help rather than fabricate answers. I like to think that the future may belong to AIs that pair intelligence with humility, systems that can admit uncertainty and collaborate with humans to find answers. This outlook is where humans and AI partner closely to achieve outcomes neither could alone and where machines augment human creativity and problem-solving without undermining trust.