Jeremy Howard argued that AI should augment human creativity and understanding instead of removing the effort that builds autonomy and mastery.
2
Annie Vella's research found that engineers reported higher productivity while their flow state and other parts of developer experience declined.
3
Mic Neale argued that idle computing power in consumer devices could support a more sovereign alternative to relying only on hyperscalers.
Summary
The second-day keynote program connected AI engineering with human agency, the craft of software work, and control over computing infrastructure. Jeremy Howard used self-determination theory to argue that work shapes people through autonomy, mastery, relatedness, and purpose. He warned that agentic coding can create a kind of dark flow, where the feeling of progress is separated from verified results, while showing an alternative in which AI helps people understand papers, test ideas, and build working systems. Annie Vella presented research with professional software engineers in 28 countries. Participants felt more productive with AI, but many reported worse flow and higher cognitive load as their work shifted toward supervision and verification. Mic Neale then described the large amount of idle compute in phones and laptops and connected it to sovereignty and local control. The stream ends shortly after Jishwan Lee begins introducing Z ai and the Chinese model-lab context.
Human flourishing depends on work that builds autonomy and mastery
Jeremy Howard began with self-determination theory and its account of people as curious, active, and motivated to learn. He contrasted eudaimonia, the fuller development of a person's capacities, with hedonia, which he described as pleasant ease and passive enjoyment. The research he cited connects authentic motivation with interest, confidence, persistence, creativity, self-esteem, vitality, and well-being. He focused on autonomy and mastery. Mastery means developing a genuine ability through effort and learning, rather than simply producing more output. Howard's point was that the work engineers choose changes who they become, so AI tools should be judged partly by whether they support those conditions.
AI coding can create dark flow when the feeling of progress loses contact with results
Howard distinguished ordinary flow from what Mihaly Csikszentmihalyi called junk flow or dark flow. Flow involves a meaningful challenge, adequate skills, clear goals, rules, and useful feedback. Dark flow can begin with a similar feeling of absorption, then become an addictive loop that does not help a person grow. He compared this with casinos creating an illusion of control. In AI-assisted coding, developers may receive a strong dopamine hit from agents producing code quickly while losing contact with external validation. One example reached 95 percent of a project in five hours, then spent 15 more hours finding problems without knowing where the project really stood. Another team had 200,000 lines of vibe-coded software and discovered during management reviews that its apparent progress did not match shipped results or customer outcomes.
AI can support mastery when it makes people understand and test ideas
Howard showed a different use of AI through Solve It, a system he used to study recursive language models and web styling. While reading a paper, he asked about figures, requested concrete examples, followed citations, wrote code, used a sub-agent, and compared the results with his own work. He then reproduced the paper's tasks and investigated whether the method was simply a tool loop with particular tools. In another example, he read Julia Evans's article about moving away from Tailwind, then rebuilt styles, components, colors, and typography inside the environment. The AI helped explain and test his ideas, but Howard kept deciding what he wanted to learn and checking the code as it ran. He presented this as a way to augment human creativity and understanding.
AI is changing software engineering toward supervisory engineering
Annie Vella described a longitudinal study of professional software engineers using AI at work or home. She surveyed participants at two points six months apart, with people from 28 countries. Most engineers said they spent less time on common development tasks, except code review, which shifted toward more time. Vella named the emerging work supervisory engineering. It includes directing AI, evaluating its output, and operating in a new middle loop between intention and implementation. The dimensions of craft remain, but engineers apply them to a different kind of work. This means that more automation does not simply create free time. It changes what people must pay attention to and how they practise software engineering.
Reported productivity can rise while the felt craft of development declines
Vella said 84% of engineers in her study reported feeling more productive with AI at both measurement points. At the same time, developer experience worsened for some participants. She measured cognitive load, flow state, and feedback loops. The share reporting a decline in at least one of these areas almost doubled to 27% at the second measurement. Flow state was affected most negatively, followed by increasing cognitive load. Feedback loops improved, but more frequent feedback could interrupt flow. Vella's conclusion was that productivity alone is an incomplete measure. Engineers may produce more while losing some of the conditions that make the work feel like a craft. She urged leaders to account for this when they assess AI-assisted work.
Self-efficacy gives engineers a practical way to improve their experience with AI
The strongest predictor in Vella's study was self-efficacy, meaning a person's belief that they can accomplish the work. Engineers with greater confidence in their software ability were over 10 times more likely to report higher productivity gains. Vella treated this as actionable because beliefs can change more readily than seniority, company size, or work conditions. She connected self-efficacy with mastery experiences: experimentation leads to learning, learning builds confidence, and confidence supports further experimentation. She then described possible future roles including the artisanal developer, the clerical coder, and the orchestrator or code conductor. Her advice was to design work so that people retain pride and joy, whether they focus on domain understanding or on building the systems that direct agents.
Idle consumer devices could provide a more sovereign source of compute
Mic Neale argued that substantial computing power already exists in phones, laptops, desktops, and other consumer devices, much of it idle. He connected this unused capacity to the energy cost of manufacturing devices as well as the electricity used to run data centers. Neale described sovereignty as having options about where computation and data run, rather than depending on connections to distant infrastructure or a small group of hyperscalers. He suggested that local models and personal devices could support collaboration without relying entirely on a public service with a single incentive structure. The talk was brief and presented the technology as a proof point rather than a finished answer. The stream then moved to Jishwan Lee's introduction of Z ai before cutting off.
"The work that we're all doing is changing. And I want to share with you some key findings from the last 50 years of research about how your work makes you."19:21
Who should watch
Software engineers deciding whether AI assistance is helping them learn or only making the first part of a task feel faster.
Engineering leaders measuring AI adoption through productivity metrics and wanting to understand flow, cognitive load, and pride in the work.
People interested in local compute, device ownership, and alternatives to sending every AI workload to hyperscaler infrastructure.
