# Hypermode Launch

Kevin Van Gundy, Hypermode | AI Engineer World's Fair 2024 | 05:03

Source: https://www.youtube.com/watch?v=8k848OeLc9E
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/hypermode-launch
Published: 2024-08-23
Tags: developer-experience, embeddings, observability

## TL;DR
- Software teams improve by trying more things faster, even when they do not know the best strategy at the start.
- AI systems need low-friction model switching, simple integration, and step-by-step tracing because teams will make wrong choices.
- Hypermode combines a runtime for models and data with tools for iteration, production observation, and AI development in a familiar coding workflow.

## Summary
Kevin Van Gundy introduces Hypermode through lessons from his time at Vercel, where rapid iteration helped the company respond to competition and discover what developers wanted. He argues that AI development has more possible failure points, including hardware, models, prompts, integration, monitoring, and runaway costs. Teams need systems that make mistakes cheap to correct. Hypermode is presented as a runtime for integrating models and data into AI functions, with tools for rapid iteration and production observation. The platform aims to support gradual AI adoption, familiar APIs, editor type-ahead, strong defaults, and simpler retrieval-augmented generation. Van Gundy also describes in-memory embedding and search, plus tools for comparing model inferences and exporting datasets for fine-tuning. He closes by inviting attendees to a workshop where they can build practical AI features and receive Hypermode credits.

## Key ideas
### Rapid iteration helped Vercel find its way through competition
[00:00](https://www.youtube.com/watch?v=8k848OeLc9E&t=0s)
Kevin Van Gundy describes three problems from his time at Vercel: the company was losing to other JavaScript frameworks, losing to hosting providers, and he was eating too much pepperoni pizza. The company started winning because it built a habit of iterating very quickly. The team did not know the optimum strategy in advance. It tried more things faster than competitors, then used what it learned to adapt its products and strategies to the market. Van Gundy calls iteration "the compound interest of software," because repeated attempts eventually produce better results.

### Developers want to adopt technology gradually
[00:53](https://www.youtube.com/watch?v=8k848OeLc9E&t=53s)
Van Gundy says the team learned two things from its rapid experimentation. Developers want to incrementally adopt new technologies. They also do not want to commit to architectural patterns before they understand how their application will work. This creates a design requirement for AI tools: teams should be able to start with a small change, learn from it, and expand later. They should not have to decide their entire AI architecture before they have enough experience with the application.

### AI development has many ways to go wrong
[01:14](https://www.youtube.com/watch?v=8k848OeLc9E&t=74s)
Van Gundy lists the decisions that make AI development intimidating: choosing hardware, selecting a model, writing prompts, integrating the system, monitoring it, and improving it. He mentions runaway bills, hallucinating chatbots, and projects that take months without delivering value. His conclusion is that teams need systems that reduce the cost of being wrong. Choosing the wrong model is less damaging when replacing it has little friction. Teams also need simple ways to trace an inference step by step and change prompts, strategies, or data mixes.

### Hypermode puts models and data inside AI functions
[02:16](https://www.youtube.com/watch?v=8k848OeLc9E&t=136s)
Van Gundy describes Hypermode as a runtime that makes it easier to integrate models and data into AI functions. Around that runtime, the company provides tools for rapid iteration and observation of those functions in production. The intended workflow starts with incremental AI adoption. As a team develops more skill, it can reimagine parts of its application as AI-native. The stated goal is to make developing with AI less difficult for developers who do not have specialized skills.

### The developer experience avoids separate model-specific setup
[03:01](https://www.youtube.com/watch?v=8k848OeLc9E&t=181s)
Adding a model can require reading documentation to learn details such as the accepted temperature range. Van Gundy says Hypermode provides type-ahead support in a developer's preferred code editor, with no SDK or download required. When code reaches production, Hypermode supplies strong defaults. Teams can also bring their existing stack. The pitch is that developers can work with models through a familiar coding environment instead of managing a separate setup for every provider.

### Retrieval can happen in one request
[03:21](https://www.youtube.com/watch?v=8k848OeLc9E&t=201s)
Van Gundy contrasts traditional retrieval-augmented generation with Hypermode's approach. In his description, traditional RAG requires an extra request to embed the input and another interaction with a vector store, creating an n-plus-one request pattern. Hypermode combines embedding and search in one request through an in-memory embedding and search service. He says this can save a couple hundred milliseconds per request. The point is to remove a piece of integration work while reducing request overhead.

### Comparing inferences helps teams build intuition
[03:58](https://www.youtube.com/watch?v=8k848OeLc9E&t=238s)
Van Gundy says nondeterministic systems are hard to understand because each model has its own personality. Hypermode lets developers compare different inferences, tunes, and models quickly. They can export the resulting dataset for fine-tuning. This gives teams a way to inspect behavior across alternatives rather than relying on one model output or trying to reason about model behavior from documentation alone.

### The launch includes hands-on workshops and credits
[04:18](https://www.youtube.com/watch?v=8k848OeLc9E&t=258s)
Van Gundy invites attendees to an afternoon workshop where the Hypermode team will demonstrate practical features. The examples include natural-language search, intelligent sorting of data lists, outlier detection, and catching bad actors. He promises that attendees will leave with a working demo and a plan to put something similar into production by the end of the following month. Hypermode also offers $1,000 in credits to get started.

## Notable quotes
- "Iteration is the compound interest of software." (00:53)
- "We need to build systems that risk getting it wrong because we are going to get it wrong a lot." (01:55)
- "At its core Hypermode is a runtime. It allows you to easily integrate models and data into AI functions." (02:16)
- "We wanted to make the developer experience of developing with AI a lot less terrible." (02:56)

## Tools & references mentioned
- Hypermode
- Vercel
- AI Engineer World's Fair
- JavaScript
- retrieval-augmented generation
- vector store

## Who should watch
- You are adding AI to an existing product and want to avoid committing to a full architecture before you know what works.
- Your team needs to compare models and prompts, trace production behavior, and change providers without rewriting integration code.
- You want practical examples of natural-language search, intelligent sorting, or outlier detection rather than a general discussion of AI strategy.

## Related talks

- [Git push get an AI API](https://aietalks.com/talks/git-push-get-an-ai-api) (Ryan Fox-Tyler, Hypermode, 45:01)
- [More Nodes Is All You Need](https://aietalks.com/talks/more-nodes-is-all-you-need) (Nicolas Schlaepfer, Hyperspace, 05:48)
- [Fast Models Need Slow Developers](https://aietalks.com/talks/fast-models-need-slow-developers) (Sarah Chieng, Cerebras, 18:02)
- [Mentoring the Machine](https://aietalks.com/talks/mentoring-the-machine) (Eric Hou, Augment Code, 19:06)
- [Keynotes & Multimodality Track](https://aietalks.com/talks/keynotes-multimodality-track) (Benjamin Duny & Simon Sturmer, AI Engineer & Chris Lattner, Modular & Auntie Aart & Mike Chambers, AWS & Alex Albert, Anthropic & Harrison Chase, LangChain & Rob Chandra, Substrate & Vic Kapati, M87 Labs & Ben Hilac, Dawn & Karan Goel, Cartesia & Chang She, LanceDB & Noah Schnapp, Character.AI & Steph Duenas & Quinn Dombrowski, Daily & Roman Huet, OpenAI & Brian Bischof, Charles Fry, Hamza Hussein, Jason Liu, Shreya Shankar & Eugene Yan & Thomas Dohmke, GitHub, 8:34:14)