Simon Willison: AI Agents Made Me More Productive. So Why Am I Working Harder?
Audio Brief
Show transcript
Can an artificial intelligence actually understand how to make a video game fun? Simon Willison realized it could when an autonomous agent iterated on game design until it was genuinely enjoyable. In this episode, host Ksenia Se sits down with Willison to discuss the rapidly shifting landscape of modern software. You will learn how to evaluate models, build custom software in an hour, and protect your systems from emerging security threats.
Standardized public benchmarks are rapidly being beaten by the newest frontier models, making custom, personal benchmarks essential for evaluating real-world utility. Willison explains that it is now much easier to build your own custom benchmarks to evaluate models on the exact tasks you need them to perform. You should build your own personalized, task specific evaluations instead of relying on public leaderboards.
The era of autonomous software builders that work for hours on their own has arrived. Since late twenty twenty three, coding agents have reached a massive inflection point in reliability. These fable class models can be given a clear, achievable goal and the necessary tools, and they will reliably complete the task even if they must run autonomously for hours.
Non technical storytellers are suddenly uncovering stories from public datasets. A massive productivity leap now allows Willison to produce a piece of software in just one hour that would have taken him a week beforehand. Artificial intelligence drastically reduces the time required for cleaning messy data and generating complex structured query language, which is the database language known as SQL. This allows data journalists to quickly uncover major stories from massive public datasets.
Creating video games is turning into a simple conversation about vibes. A new paradigm called vibe game development is where creators can describe gameplay ideas to an AI agent, which then refactors and refines the code over hours to make the game genuinely 'fun'. You describe gameplay ideas to an agent. The assistant then refactors and refines the code over several hours until the game becomes truly engaging.
A lethal trifecta occurs when a large language model is exposed to three specific conditions at the same time. If a model has access to your private data, untrusted user instructions, and a mechanism to send data back out, attackers can hijack the system. You must ensure your automated systems never combine private access, untrusted inputs, and data sending capabilities at the same time.
You can now run top tier artificial intelligence directly on your personal laptop. Small, open weights models like Qwen can now run efficiently without any internet connection. These local systems offer near frontier capabilities without cloud dependency or privacy risks.
Remember to force yourself to document and publish your findings for every project you complete to cement your learning. Deep domain knowledge remains critical for guiding and debugging agents. I really can, in an hour, produce a piece of software that would have taken me a week beforehand.
Episode Overview
- Simon Willison sits down with Ksenia Se to discuss the state of AI models, custom benchmarks, the evolution of coding and game development agents, security challenges, and the potential of local LLMs.
- The conversation highlights how rapidly the capabilities of AI models are shifting, from 3D modeling and game design to text-to-SQL and autonomous agents.
- Willison emphasizes the transition from relying on standardized public benchmarks to building personalized, task-specific evaluations.
- The episode frames the future of software development not as a threat to developer jobs, but as an opportunity to meet an explosive demand for custom, hyper-niche software.
Key Concepts
- Task-Specific Custom Benchmarking: Standardized public benchmarks are rapidly being beaten by frontier models, making custom, personal benchmarks essential for evaluating real-world utility.
- Fable-Class Models: AI systems that, when given a clear, achievable goal and access to necessary tools, will reliably complete the task, even if it requires running autonomously for hours.
- The Intersection of AI and Data Journalism: AI drastically reduces the time required for data cleaning and SQL query generation, empowering non-technical journalists to uncover stories from public datasets.
- Vibe Game Development: A new paradigm where creators can describe gameplay ideas to an AI agent, which then refactors and refines the code over hours to make the game genuinely "fun."
- The Lethal Trifecta of Prompt Injection: A security vulnerability that occurs when an LLM is simultaneously exposed to private data, untrusted user instructions, and a mechanism to exfiltrate data.
- The Rise of Powerful Local Models: Small, open-weights models like Qwen can now run efficiently on consumer laptops, offering near-frontier capabilities without cloud dependency or privacy risks.
Quotes
- At 3:32 - "It is worth the time now to build your own custom benchmark because it is not as difficult as it used to be." - explaining why developers should shift from public leaderboards to personalized evaluations.
- At 7:04 - "This year has been the year of Coding Agents." - highlighting the massive inflection point in agent reliability starting in late 2023.
- At 8:53 - "I really can, in an hour, produce a piece of software that would have taken me a week beforehand." - demonstrating the massive productivity leap enabled by modern LLM coding assistance.
- At 11:56 - "I thought fun was the last remaining uniquely human trait. Like no machine can make something fun. That is for us. Turns out they can." - describing the realization that AI can autonomously iterate on game design to optimize for enjoyment.
Takeaways
- Build your own task-specific custom benchmarks to evaluate models on the exact tasks you need them to perform, rather than relying on generic public leaderboards.
- Embrace a "write-up rule" where you force yourself to document and publish your findings for every AI project you complete, helping to cement learning and prevent half-finished loops.
- Seek opportunities to understand the lower levels of your technology stack, as deep domain knowledge remains critical for guiding and debugging AI agents when they encounter complex failures.