What 40 YC AI founders say about building with AI today

Executive overview

AI tools are surprisingly strong at creative tasks but unreliable at distinguishing fact from fiction. Founders are shipping with probabilistic models layered on top of deterministic software — a fundamentally different engineering discipline. The gap between an 85% AI solution and a production-ready one requires iteration, prompt engineering, and humans in the loop.

The core challenge is not capability — it's trustworthiness and knowing when to steer the model.

What AI is actually good at

• Semantic search: finding relevant content from arbitrary text, previously unsolved
• Creative and generative work: voice synthesis, image generation, storytelling
• Coding acceleration: describe a UI change in plain English, get working code
• Reading and reasoning over arbitrary data to answer questions
• Delivering an 85–90% solution quickly using simple chained operations

The probabilistic problem

• Same inputs can produce different outputs — unlike traditional software
• Marrying deterministic systems with probabilistic models is the core engineering tension
• Variance can be a feature in entertainment; it is a liability in safety-critical applications
• Iteration on prompts is ongoing — solutions degrade as data and models change

Hallucination and trust

• Models cannot reliably distinguish fact from fiction despite being fluent storytellers
• Efforts to reduce hallucination create the opposite failure: models deny knowing things they should know
• Citations are still inconsistent — trustworthiness has a long way to go
• In high-stakes domains (e.g. medicine), verification overhead can negate time savings

Working with AI effectively

• Give the model one specific task and clear structure — it performs far better with constraints
• Engineer the process: map your own workflow, then encode it as a prompt sequence
• Debug and iterate prompts continuously; do not treat any solution as permanent
• Keep humans in the loop, especially where accuracy and trust matter
• Use randomness in outputs deliberately to explore solution spaces