Questions Considered

Notes on thinking, learning, decision making, and occasionally running. Simple ideas, mostly obvious.

Tag: artificial intelligence

  • Per ACM’s policy on authorship (further clarified here), authors may use AI to prepare manuscripts and are not required to disclose such usage.

    […] some articles provoke a vocal response, with critics accusing them of being “AI Slop”—a vague but heavily pejorative label claiming that a machine wrote an article, that it contains commonplace observations, and that it lacks human insight. In some cases, the evidence appears to be little more than grammatically correct, stylistically polished prose.

    […]

    For Communications, the important distinction is not between articles written with and without AI. It is between using AI to express human thought and using AI to avoid human thought.

    AI and Communications of the ACM, by James Larus

    If something reads as if it is Claude generated (be it structural, certain words or turns of phrases), it likely affects the opinion you form about the text and perhaps you wonder about its provenance or substance.

    After the onslaught on AI-generated text over the last couple years, it is pretty easy to find oneself second-guessing.





  • Much has been and will continue to be written about how the field of mathematics is changing because of AI and how its community is holding up.

    Socrates is said to have worried that writing would ruin our memories—and Socrates’ crowd probably did remember longer stretches of the Odyssey than any of us do. But the Fields letter argued that the use of A.I. was not like merely off-loading the memorization of poems or phone numbers; it was more like off-loading, or bypassing, daydreaming and debating and thinking. It’s not that humans will come to know too much too quickly but that they will understand less and less—eventually less than we understand now.

    The Year Math Changed Forever, by Rivka Galchen.

    Optimizing away cognitive friction is a double-edged sword.





  • Research at Okta examined the impact of using GitHub Copilot in software engineering and found a gap between individual gains and overall throughput.

    Despite increased engineer motivation and perceived skills, as well as reduced working hours, GitHub Copilot adoption revealed no statistically significant improvement in engineering output metrics. We call this disconnect the efficiency-throughput gap.

    The Organizational Ohm’s Law (OOL) provides a practical and holistic framework for measuring engineering productivity, integrating subjective surveys with objective metrics to assess AI’s impact as a function of individual potential and organizational resistance.

    Achieving organizational productivity gains requires more than deploying an AI-powered tool. It demands a deliberate effort to learn, adapt, and reinvent teams, processes, and infrastructure to align with AI-augmented developer capabilities.

    Beyond the Hype: The Efficiency-Throughput Gap with GitHub Copilot, by Fei Liu and Todd McKinnon.

    The usual caveat here: The research looked at tool usage for a time period in 2024. Given the many advancements since that is of course a very long time ago indeed.

    It does highlight an important property often seen in systems thinking and constraints management: Speeding up one part of a larger system does not automatically make the whole more performant. Optimizing the whole does not usually mean optimizing the individual parts.





  • We are all operating with bounded rationality. The AI chatbots you interact with can cater (and even encourage) narrow views – and it could actively help you grow and broaden them.

    An emerging line of work, sometimes described as machine psychology, studies how generative models exhibit and interact with cognitive shortcuts and how systems might support better reasoning rather than amplify bias. This motivates a research pathway that connects design, evaluation, and governance to reduce AI harms and bias risks in practice.

    Bounded Minds, Generative Machines, by Jiqun Liu.

    I think there is a lot of opportunity in using AI to help expand the bounds of a person’s rationality. That is what can make it an effective thinking partner.





  • The lens of elephant-in-the-room and the blind men who are individually only perceiving parts of it make for powerful framing to examine systems and change.

    Generative artificial intelligence (AI) coding tools are transforming software production and work. To benefit, organizations must integrate AI into processes, enforce stricter quality controls, manage automation risks, and adapt training, career paths, and governance to sustained AI-assisted development.

    AI: The Blind-Men Elephant in the Software Developers’ Room, by Diomidis Spinellis.

    AI is one elephant in the software development room. Less clear is what we are blind to see. The more convinced you are of your vision, the more surprising your blind spots.





  • Having learned to distrust AI output, the author of the following essay came to draw productive parallels to alchemy.

    We are now in the pre-chemistry era of AI. We know how these models work, because we crafted their code ourselves, and yet in most cases we cannot say why: why this word and not another, why this particular path through the machine. The crucible is closed, and like the alchemists we are not going to stop using it. The alchemists were the ones in history who had to pull real value from a process they could not see into. Now it’s our turn.

    AI Made Me Look Foolish. Then I Learned, by Paul M. Sutter.

    Alchemy led to chemistry. Deliberate scientific approaches in our use of AI, will lead us to better understanding of the technology and improved outcomes. Arguably, the journey from early vibe coding to current coding harnesses illustrates some of that.





  • Data generated by defunct companies is becoming increasingly interesting as training material for AI companies.

    Demand for defunct companies’ data emerged suddenly last year, as AI training labs hit the limits of public and synthetic datasets.

    […]

    This past August, as part of Spirit Airlines’ bankruptcy, Google outbid the AI data company Mercor to buy the carrier’s decades of data, including 100 million emails, 500 million Teams messages, and 176,000 employee records. The price tag: $10 million. And Bloomberg reported in September that SpaceX’s AI unit is weighing buying data from struggling startups.

    AI labs are buying data from dead startups to train their models, by Matthew King.

    This sounds like a potentially expensive data privacy nightmare in the making. Even just mapping out the parties affected would be difficult, given that conversational data may contain all kinds of references – not only to employees of the company, but also vendors, competitors, customers, et cetera.





  • In a recent working paper, researchers report on how to enable practitioners in organizations to use AI more effectively.

    Companies often struggle to identify where AI can produce meaningful business gains. Research involving 515 companies found that those exposed to examples of how other organizations redesigned work around AI identified 44% more use cases, completed 12% more tasks, were 18% more likely to acquire paying customers, and generated nearly twice as much revenue. Leaders should map workflows from beginning to end, identify constraints that limit performance, and redesign surrounding processes rather than simply automating individual tasks. Because solving one constraint can create another elsewhere, companies should continually reassess their processes. They should also encourage experimentation and learn from AI applications outside their own organizations.

    For Real Performance Gains with AI, Start by Mapping Your Workflows

    Systems thinking (and theory of constraints) continues to matter a lot.





  • In a recent paper, researchers showed how to train to be better able to tell AI-generated face images from those of real humans.

    Compared with real faces, AI-generated faces tend to be more symmetrical, more proportional and more attractive—while also being less expressive, less distinctive and significantly less memorable. When the researchers taught participants to look for these six markers instead of fleeting artifacts such as malformed ears or mismatched jewelry, their accuracy in spotting an AI face almost doubled, from about 40 to 80 percent success on a comparison task.

    How to spot an AI-generated face, according to science, by Sam Macdonald.

    This is interesting, though I suspect that it is also, over time, a losing battle.





  • The Opening Argument of The Atlantic’s November issue is about the quickly growing resistance against data center construction.

    More than 500 jurisdictions have passed moratoriums or bans. This grassroots movement rose so fast and spread so widely that politicians in both parties, including Iowa’s candidates for governor and senator, are struggling to catch up with their constituents. The meteoric emergence of the data-center rebellion is a response to the speed and scale of construction, which in turn reflect the sudden ubiquity of generative and agentic AI.

    […]

    By pushing so hard and promising so much, tech companies, venture capitalists, and their political allies ensured that AI would create a firestorm.

    The Great Data-Center Rebellion, by George Packer.

    Can those firmly inside a tech bubble and living away from the affected communities really understand the concerns?