Back to previews chapter 3 core principles
<aside> <img src="https://prod-files-secure.s3.us-west-2.amazonaws.com/d8baae53-7dc0-4bad-a65f-26898d6a633d/0c5a5d12-87b1-4a0e-b08d-3e2bd3994e3a/Silex_Brand_Symbol.png" alt="https://prod-files-secure.s3.us-west-2.amazonaws.com/d8baae53-7dc0-4bad-a65f-26898d6a633d/0c5a5d12-87b1-4a0e-b08d-3e2bd3994e3a/Silex_Brand_Symbol.png" width="40px" /> Sending the same input 3 times to AI will give you 3 different outputs.
</aside>
Generative AI, at its core, produces text by predicting the most likely next word based on previous inputs. There’s an inherent element of randomness involved, stemming from the probabilistic nature of the technology. Similar to weather forecasting or human behavior, generative AI is stochastic—meaning that identical inputs won’t always yield the same results.
This inherent randomness will inevitably influence your AI-driven features. As a result, it’s crucial to understand that you won’t always achieve the expected output with absolute certainty.
When selecting your use case, keep in mind that achieving near-perfect results will require extensive effort. You’ll need to continuously iterate, evaluate, and refine your system to approach a success rate as close as possible to 99.99%.
The real challenge for teams is to embrace this randomness and adapt to this new paradigm. Success will depend on building a deep intuition about the model’s behavior while maintaining the resolve to harness the technology and achieve your goals.
“Companies succeeding in shipping great AI features iterate relentlessly to build an intuition about the models’ behavior without giving up” - Olivier Godement, Head of API Product at OpenAI during Hexa event
<aside> 💡 What does this mean in practice?
The steps you follow to build an AI product look different to those for a traditional product.
Traditional product development follows a mostly linear path, since the expected output of the feature is predictable and provide a consistent user experience. It might look something like this:
Problem discovery > Specification > Technical development > Quality assurance
AI changes things. Introducing a non deterministic technology at the center of the delivery changes the game. Teams now need to:
<aside> <img src="https://prod-files-secure.s3.us-west-2.amazonaws.com/d8baae53-7dc0-4bad-a65f-26898d6a633d/e055bf2a-4109-4d41-95c0-09f440c11d1d/Silex_Brand_Symbol.png" alt="https://prod-files-secure.s3.us-west-2.amazonaws.com/d8baae53-7dc0-4bad-a65f-26898d6a633d/e055bf2a-4109-4d41-95c0-09f440c11d1d/Silex_Brand_Symbol.png" width="40px" /> The technology changes so quickly that every six months, your product will be noticeably different
</aside>
Particularly with respect to the foundational models developed by OpenAI, Google, Anthropic, Mistral, and others, AI is constantly in flux.
