How did you hear about this role?

I’ve heard about Elicit from the Latent Space podcast back in April where Jungwon and Andreas told their founding stories. I was surprised to hear that they had researched trying to emulate advanced AI with humans and learned a lot about topics around segmented intelligence, which I later experienced while building LLM applications. I also found another episode insightful where James explained the principles of AI engineering is a fault-tolerant mindset and distributed system skills.

What's something you're learning now or have recently learned?

I recently examined bolt.new's open-source codebase to understand how they achieved the performance, and learned three key insights of their success:

• They boldly didn’t impose any pre-defined workflows or agentic flows that can get complicated quite quickly, but just used a single LLM call with a custom parser.
• They boldly kept all information in one message history chain removing the finicky problem of managing context, but sacrificing the cost of LLM call, potential context limit issue, and steer-ability.
• They’ve done a good job of breaking down what is necessary for full-stack LLM agent: create files, edit files, execute shell commands. This made the output as simple as for LLMs to follow the structure and later to be parsed.

LLMs are improving at a rapid pace, and Bolt designed their product with this trajectory in mind. While many understand this concept in theory, executing this vision successfully is far more challenging. It's tempting to overcomplicate the system to make it work. More importantly, it's difficult to create the right abstractions that allow next-generation models to reach their full potential. Bolt's team successfully achieved this with theses three points.

However, I suspected that this approach has a significant limitation: the system's controllability is reduced since everything relies on a single LLM call with a simple message history. Although this simplicity approach will shine in a short term especially right after the model’s capability improves, it will hit a wall very soon.

I've subscribed to Bolt and tried it for a few weeks and experienced problems related to this. It was good at starting a project from scratch but struggled with improvements, meaning the longer the conversation gets, the more it overlooks relevant code snippets and loses sight of the entire architecture. This limitation mainly stems from its state management — a single message history. The system needs a better solution that can effectively retrieve code snippets related to the user’s request.

Through this analysis, I reaffirmed that task decomposition is essential for building an AI assistant capable of complex tasks. Yet I also discovered the power of simplicity in unleashing LLMs' full capabilities. I’m exploring how to merry these two approaches.

Which of your past experiences or projects are most relevant to the work you will do at Elicit?

I’d like to share my side-project “Interactive Story Game”

Talk about

• prompt engineering
• parallelize LLM calls
• Having a fallback model
• Using function calling, string parser, type check,
• using different models: GPT mini, 4o, Claude

Based on what you've read here or already know about Elicit, what interests you most about this role in particular?