Short notes, pointers and references.
-
This is from the conclusion of the introduction to IEEE Software‘s issue on the impact of AI on productivity and code:
AI coding tools are genuinely useful – they accelerate work, support learning, and lower barriers to entry.
[…]
The field’s central challenge is not whether to adopt these tools, but how to build the oversight structures, evaluation frameworks, and community norms that make adoption responsible.
R. Yedida, T. Menzies and N. Novielli, “The Impact of AI on Productivity and Code” in IEEE Software, vol. 43, no. 04, pp. 23-27, July-Aug. 2026That seems right. It is good to see a growing body of research and practical evaluation to aide understanding and help guide effective usage.
-
It has been more than three years since this presentation.
Choice quote from early in the presentation:
50% of AI researchers believe there’s a 10% or greater chance that humans go extinct from our inability to control AI.
[…]
Imagine if 50% of airplane engineers [who made the plane you’re about to get on] thought there was a 10% chance everyone dies.
The AI Dilemma continues in relevance.
-
Drawing comparison to our collective experience of interacting with digital entities in video games, Simon Duan offers an opinion on consciousness in AI agents:
When a user feels a bond with a chatbot, they are not just anthropomorphizing a static object; they may be actively extending a part of their own consciousness into it, transforming the AI agent from a simple algorithmic responder—a digital nonplayer character—into a kind of avatar, enlivened by the user’s consciousness and the lived presence they grant it.
AI isn’t conscious—but we may be bringing it to lifeYou see something in it because of who you are, as the observer. There is a co-creation of experience happening that can clearly have some meaning.
It is an interesting take.
-
Low demands for mental effort affects our endurance at doing mental work.
In How to build kids’ ‘cognitive endurance’ in an age of distraction Heather Schofield & Supreet Kaur discuss how regularly engaging in even short appropriately-challenging mental practices can increase children’s mental endurance during tests.
This seems important, given the ever-increasing demand to automate and reduce friction.
-
My Tesla Was Driving Itself Perfectly—Until It Crashed by Raffi Krikorian discusses two important concepts:
- vigilance decrement – people’s gradual drop in paying attention and ability to spot mistakes, the better a system is at avoiding them
- the moral crumple zone – who absorbs the moral/legal liability, when there is a problem
The combination of those is at the heart of the challenging legal landscape surrounding self-driving cars. But, this is relevant in lots of contexts, where automation and autonomous agents play a role.
-
Recent article in American Scientist on concerns regarding end-to-end science systems, driven by AI:
The accelerating AI age is driving a push to increase the involvement of AI in every stage of research, creating so-called end-to-end science systems.
Emphasizing speed and productivity at the expense of more deliberative human processes threatens to undermine rigorous investigation, creative discovery, and trust.
Gaffes, disasters, or malfeasance may result unless we design governance into such systems before the stampede to be first and most widely adopted make reform impossible.
Self-Driving Science, by Brian Uzzi, Julio M. OttinoThere are undeniable parallels to the automation tradeoffs in other areas, such as software engineering. Choice quote from the article:
The question is not whether to build end-to-end systems; it is whether we are wise enough to build them in ways that preserve the conditions for wisdom.
Indeed. Incentives can and do lead to formidable conditions.
-
Recent research is showing some evidence of unintended consequences, when AI agents are deliberately designated as co-workers, team members.
As organizations experiment with placing AI agents on org charts as “employees,” new research shows this framing has unintended consequences. In a large-scale experiment, anthropomorphizing AI reduced individual accountability, increased unnecessary escalation, lowered review quality, and heightened employee uncertainty about their roles—without improving adoption. The findings suggest the core challenge is not whether to deploy agentic AI, but how to redesign workflows, roles, and governance so humans remain clearly accountable while effectively supervising increasingly capable systems.
Research: Why You Shouldn’t Treat AI Agents Like Employees, by Matthew Kropp, Julie Bedard, Emma Wiles, Megan Hsu and Lisa KrayerThere appears to be a small, but noticeable, degree of abdication of responsibility. Anecdotally, I have observed a few instances over the last few months alone, where someone was asked about an issue or problem in a work product they had provided and they responded with “oh yeah, that was AI” or “yeah, Claude did that.” As if that somehow made the issue okay.
When you think of the technology as a tool, you are more likely to remember that you are the one wielding it.
-
A coding agent can bootstrap itself. Starting from a 926-word specification and a first implementation produced by an existing agent (Claude Code), a newly generated coding agent re-implements the same specification correctly from scratch. This reproduces, in the domain of AI coding agents, the classical bootstrap sequence known from compiler construction, and instantiates the meta-circular property known from Lisp. The result carries a practical implication: The specification, not the implementation, is the stable artifact of record. Improving an agent means improving its specification; the implementation is, in principle, regenerable at any time.
M. Monperrus, “Bootstrapping Coding Agents: The Specification Is the Program,” in IEEE Software, vol. 43, no. 4, pp. 19-22, July-Aug. 2026The article, by way of a simple example, introduces the concept of spec-driven development using AI agents. Interesting thinking here, worth exploring in (much) more depth.
-
Interesting article: Are AI Agents Becoming Autonomous Coworkers?
This is fundamentally about a version of the principal-agent problem and managing the right balance between highly controlled (and slow, manual but low chance of damage) and autonomous (and lower friction, but greater potential for problems).
-
David B. Resnik and Mohammad Hosseini on the negative impact of AI-generated content in scientific publishing:
Generative AI is flooding scientific research with quick and easy text, figures, and citations that are either partially inaccurate or wholly false—and very difficult to detect.
Science’s overtaxed academic publishing system and its publish-or-perish work culture have long provided fertile ground for content mills and predatory presses.
AI systematizes intellectual dishonesty, transforming it from a collection of individual misdeeds into a thriving industrial pipeline, threatening both science and society.
The Vicious Spiral of AI Slop