What You'll Learn
The Evolution of a Founder's Toolkit
Jonas Templestein's career reads like a compressed history of modern technology: from video games in childhood, to co-founding one of the UK's most successful neobanks, to now building Iterate - a company at the forefront of AI-powered startup infrastructure. But what makes his perspective particularly valuable is how radically his own work has transformed in just two years.
"I could not do my job without LLMs," Jonas states plainly. "I would literally pay $10,000 a month for GPT-4 with deep research." This isn't hyperbole from a tech enthusiast - it's a practical assessment from someone who has built products used by millions.
Voice: The New Default Interface
Perhaps the most striking change in Jonas's workflow is his shift to voice as the primary computer interface. "I almost exclusively use my voice now, except when I'm actually writing code myself," he explains. This represents a fundamental departure from the keyboard-centric workflow that has dominated programming for decades.
The reasoning is simple: voice has higher input bandwidth. While typing might feel faster for code, for everything else - research, planning, ideation, communication with AI assistants - voice enables a more natural and efficient interaction model. This shift mirrors broader changes in how we're beginning to interact with intelligent systems.
Understanding Self-Driving Startups
Jonas introduced the concept of "self-driving startups" - companies where AI agents handle an increasing proportion of operational tasks. With the latest Gemini diffusion models capable of generating 2,000 tokens per second, entire applications can be created within a single request-response cycle.
"A human speaks to a computer, the computer decides a graphical UI would be useful, and 200 milliseconds later, it's there," Jonas describes. This capability costs several orders of magnitude more than serving cached pages, but represents a fundamental shift in how software can be built and delivered.
The implications are profound: rather than pre-building interfaces for generic users, future applications could generate perfectly customized UIs on-demand for each individual interaction. This personalization at scale was previously impossible, but becomes economically viable as inference costs continue to decline.
The Headcount Question
Reflecting on his Monzo experience, Jonas offers a counterintuitive insight: "In so many ways, headcount kills. If you can do something with fewer people, it's just better." Had AI tools been available at Monzo's founding, he believes a smaller team could have built more, particularly in operations and customer service.
This isn't theoretical. Gradient Labs, founded by a former Monzo colleague, now provides AI customer service that achieves higher satisfaction scores than human agents - including at Monzo itself. "Any job you can do from your parents' basement, AI will do better within 2-3 years," Jonas predicts, noting this includes CEO roles at unicorn companies.
From Skepticism to Conviction
Jonas's AI journey began with skepticism. GPT-2's poetry attempts seemed merely interesting. But two moments crystallized his conviction: seeing GPT-3.5 generate a React component, and experiencing Stable Diffusion's creative capabilities.
"Most people couldn't read the React docs and make a component," he recalls. "It was easy to see the trajectory." Stable Diffusion was even more transformative - moving from blurry neural net outputs to 512x512 pixel images of anything imaginable in just 18 months.
The Energy Bottleneck
When asked about limitations, Jonas identifies energy as the primary constraint. "People are always confused how the world finds new ways to use energy," he observes. As AI capabilities expand, energy consumption will increase dramatically - not just for training, but for inference at scale.
Generating custom UIs on demand, running deep research queries, and powering autonomous agents all require orders of magnitude more energy than traditional web services. This economic reality shapes everything from model selection to business model viability.
Practical Advice for Founders
Jonas's core message for founders: use the best, most expensive models available. "It's like hiring employees - you don't want the cheapest, you want the best ones. If you can spare $200 a month, get GPT-4 with deep research. It's a better collaborator than almost any engineer I've worked with."
This marginal cost consideration differs fundamentally from traditional internet technology. While cloud infrastructure scaled with negligible per-user costs, AI inference remains expensive. Founders need to factor this into their economics while benefiting from the capabilities these models unlock.
The Next Wave
As Jonas looks ahead, he sees AI agents moving from concept to reality. Not the chatbots of previous hype cycles, but truly autonomous systems capable of complex decision-making and task execution. Combined with voice interfaces, real-time generation, and improving models, we're approaching a fundamentally different computing paradigm.
For indie hackers, engineers, and founders, Jonas's message is clear: the tools are here now. The question isn't whether to adopt AI in your workflow, but how quickly you can integrate it deeply enough to gain competitive advantage. Those who treat LLMs as occasional assistants will be outpaced by those who rebuild their entire workflow around them.
Conclusion
Jonas Templestein's journey from video games to Monzo to Iterate illustrates how each technology wave creates new opportunities for those who recognize them early. His candid assessment of AI's current capabilities and near-term trajectory provides a valuable benchmark for founders navigating this transition.
The most striking aspect isn't the prediction of dramatic change - it's the evidence that this change has already happened in Jonas's own workflow. When a successful founder who built a $5 billion company says their computer usage has changed more in two years than in the previous thirty, it's worth paying attention to what they're doing differently.
