The task-completion time trajectory, 3 seconds in 2019 to nearly 5 hours in late 2025, with the doubling rate accelerating from every 7 months to every 4 months, is the most concrete capability benchmark I've seen that doesn't rely on benchmark scores. But two predictions sit in tension without resolution: Inference Famine suggests compute constraints will limit deployment, while the labor displacement prediction assumes agent adoption scales fast enough to reshape the workforce. If inference stays scarce, which of your 18 predictions breaks first, and what's the condition that resolves the contradiction?
The physical AI prediction is where compute shortages meet a deployment reality that's rarely discussed: edge connectivity. Autonomous agents and robots don't just need raw intelligence -- they need persistent, low-latency data paths to function reliably. The class divide between premium and free-tier AI users will mirror what we already see in IoT: organizations with eSIM-based carrier orchestration, private 5G, and OTA firmware governance can deploy physical AI at scale, while others remain stuck in demo mode. The real UX moat in physical AI isn't the interface -- it's the invisible connectivity and edge inference stack that determines whether a robot actually works on a factory floor or fails silently.
The 'Review Fatigue' concept is something I'm actively fighting in my agent system. Built explicit verification loops and human approval gates for anything irreversible. Your point about UX becoming the primary business moat resonates - as AI capabilities commoditize, the differentiator becomes how seamlessly humans can direct these systems. https://thoughts.jock.pl/p/wiz-personal-ai-agent-claude-code-2026
The shift from a conversational UI to a delegative UI means we won’t just be asking AI for information we’ll be able to assign it goals and have it plan and execute tasks autonomously. This has an impact in all roles within companies: imagine managers eventually delegating not only to their team members but also to AI agents.
It’s exciting to think about the new workflows and efficiencies this could enable, but it also raises important questions about oversight and trust as AI takes on decision-making responsibilities.
Fantastic article. One topic I felt was missing: edge models & improving intelligence density on edge devices. Would be very interested in your take around this. It seems to me to be a space ready for disruption - depending on if reasoning potential on edge models can continue to improve, & if anyone discovers decent product ideas for that space.
Hey Jakob, after finish reading your article on the predictions, we went to bed. Only to wake up at 3 AM with some of our own thoughts 💭.
While we agree with your thoughts on how the different AI tools will become really good and may occasionally overtake each other, we don't quite agree with the whole "switch from one to the other every few months," approach.
Here's why: Context & Memory
Now that AI tools allows us the option to ✅ switch on the Context/Memory using the chat history, switching AI tools mid-stream on any large scale project work can be potentially disruption. Whether it's design or engineering.
Even if a designer or engineer is done with a single project in one season using their AI tool of choice and decides to switch after, there is always a chance that they may need to revisit the project later to make updates or improvements and changes.
Again, context and chat session memory does matter as well.
In our very own case, we are working on a major product vision (our next digital masterpiece, "magnum opus") that could span the next 2-3 years. Or shorter once we can get our hands on the right hardware. Personally, we wouldn't be considering switching, even if that means continuing to use an AI tool that is slightly behind on the others.
At the end of the day, all these AI tools are nothing but accelerators. What still matters is the designers/engineers themselves.
Anyway, that is our 3 AM take on this topic. It's 3:11 AM, for us to go back to sleep.
The task-completion time trajectory, 3 seconds in 2019 to nearly 5 hours in late 2025, with the doubling rate accelerating from every 7 months to every 4 months, is the most concrete capability benchmark I've seen that doesn't rely on benchmark scores. But two predictions sit in tension without resolution: Inference Famine suggests compute constraints will limit deployment, while the labor displacement prediction assumes agent adoption scales fast enough to reshape the workforce. If inference stays scarce, which of your 18 predictions breaks first, and what's the condition that resolves the contradiction?
The physical AI prediction is where compute shortages meet a deployment reality that's rarely discussed: edge connectivity. Autonomous agents and robots don't just need raw intelligence -- they need persistent, low-latency data paths to function reliably. The class divide between premium and free-tier AI users will mirror what we already see in IoT: organizations with eSIM-based carrier orchestration, private 5G, and OTA firmware governance can deploy physical AI at scale, while others remain stuck in demo mode. The real UX moat in physical AI isn't the interface -- it's the invisible connectivity and edge inference stack that determines whether a robot actually works on a factory floor or fails silently.
Here's how Prediction 9: Generative UI (GenUI) and the Disposable Interface unfold: https://adamvalek.substack.com/p/the-new-era-of-digital-experiences?r=1uvjax&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true
The 'Review Fatigue' concept is something I'm actively fighting in my agent system. Built explicit verification loops and human approval gates for anything irreversible. Your point about UX becoming the primary business moat resonates - as AI capabilities commoditize, the differentiator becomes how seamlessly humans can direct these systems. https://thoughts.jock.pl/p/wiz-personal-ai-agent-claude-code-2026
The shift from a conversational UI to a delegative UI means we won’t just be asking AI for information we’ll be able to assign it goals and have it plan and execute tasks autonomously. This has an impact in all roles within companies: imagine managers eventually delegating not only to their team members but also to AI agents.
It’s exciting to think about the new workflows and efficiencies this could enable, but it also raises important questions about oversight and trust as AI takes on decision-making responsibilities.
Must read article not only for designers!
Fantastic article. One topic I felt was missing: edge models & improving intelligence density on edge devices. Would be very interested in your take around this. It seems to me to be a space ready for disruption - depending on if reasoning potential on edge models can continue to improve, & if anyone discovers decent product ideas for that space.
Hey Jakob, after finish reading your article on the predictions, we went to bed. Only to wake up at 3 AM with some of our own thoughts 💭.
While we agree with your thoughts on how the different AI tools will become really good and may occasionally overtake each other, we don't quite agree with the whole "switch from one to the other every few months," approach.
Here's why: Context & Memory
Now that AI tools allows us the option to ✅ switch on the Context/Memory using the chat history, switching AI tools mid-stream on any large scale project work can be potentially disruption. Whether it's design or engineering.
Even if a designer or engineer is done with a single project in one season using their AI tool of choice and decides to switch after, there is always a chance that they may need to revisit the project later to make updates or improvements and changes.
Again, context and chat session memory does matter as well.
In our very own case, we are working on a major product vision (our next digital masterpiece, "magnum opus") that could span the next 2-3 years. Or shorter once we can get our hands on the right hardware. Personally, we wouldn't be considering switching, even if that means continuing to use an AI tool that is slightly behind on the others.
At the end of the day, all these AI tools are nothing but accelerators. What still matters is the designers/engineers themselves.
Anyway, that is our 3 AM take on this topic. It's 3:11 AM, for us to go back to sleep.