# Inferring Creativity

Nick Hasty, GIPHY | MLOps Podcast | Episode 198 | 55:30
Hosted by Demetrios Brinkmann

Source: https://www.youtube.com/watch?v=89gd1r_dVuU
Channel: MLOps Community, now AAIF Live (https://www.youtube.com/@AAIFLive-x1r). Summarised by MLOps Talks.
Page: https://mlopstalks.com/talks/inferring-creativity
Published: 2023-12-26
Tags: embeddings, product-strategy, search, startups

## TL;DR
- Nick Hasty says his move from music and internet art into engineering came from learning to work through uncertainty rather than assuming technical skills would appear naturally.
- GIPHY used machine learning embeddings to connect searches, images, emotions, and cultural references, including relationships that traditional tags could not capture.
- Nick argues that AI product teams should choose models based on the user value they need to deliver, while accepting that large language models still bring latency and scaling limits.

## Summary
Nick Hasty describes a career that moved from music and internet art to engineering, then to product and machine learning at GIPHY. He joined GIPHY in 2013 when its search engine was still small, and helped build a service that treated GIFs as a communication format distributed through other networks. As the catalog grew, manual tagging stopped scaling. GIPHY used computer vision, language models, and embeddings to improve search and recommendations, learning relationships among images, queries, clicks, emotions, and pop culture. Nick explains why he prefers product engineering: he wants to make things that people use and that create a shared experience. He sees large language models as a way to build prototypes and interfaces that handle messy language, but says teams should use older methods when they already solve the problem. He also discusses GIPHY's experiments with generative GIFs, startup advising, and the pressure for new companies to focus on specific users and gaps that large technology companies will not address.

## Key ideas
### Nick's path into engineering began with music, art, and uncomfortable technical work
[08:38](https://www.youtube.com/watch?v=89gd1r_dVuU&t=518s)
Nick Hasty grew up in rural North Georgia, studied English in Athens, and initially wanted to spend his life playing music. At NYU's Interactive Telecommunications Program, he encountered circuit design, prototyping, and programming. He had to work through the assumption that he was not the kind of person who could program. Afterward, he worked at Rhizome.org, an internet art nonprofit connected with the New Museum of Contemporary Art, where he became its only engineer and learned how to archive and work with GIFs as an internet-native art medium.

### GIPHY treated GIFs as a communication layer across existing networks
[12:12](https://www.youtube.com/watch?v=89gd1r_dVuU&t=732s)
Nick joined Alex Chung at GIPHY in February 2013, when the first prototype was a GIF search engine backed by a MySQL database and crawled Tumblr data. The site was already popular enough to crash under demand. The team saw people using GIFs as reactions, analogies, and responses inside comments and posts. They decided to build a content distribution network that plugged into places where people already communicated, rather than creating another social network. That meant integrations, plugins, and a willingness to handle whatever work the growing company needed.

### He handles scaling by treating uncertainty as a reason to research
[18:39](https://www.youtube.com/watch?v=89gd1r_dVuU&t=1119s)
Nick says he recognizes how much he does not know, and that awareness drives a constant search for information and people who can help solve a problem. He does not expect insight to arrive on its own. He researches, applies what he finds, sees when it fails, and then asks what he misunderstood. At GIPHY, this approach mattered because the team had to keep infrastructure running while the service grew quickly. The pressure included users publicly complaining when GIFs did not load, even though the consequences were very different from failures in life-critical systems.

### Machine learning replaced an increasingly unmanageable manual tagging process
[22:05](https://www.youtube.com/watch?v=89gd1r_dVuU&t=1325s)
As GIPHY pulled in more content, a small editorial team could no longer tag everything by hand. Nick began investigating computer vision and machine learning as a way to extract useful metadata from GIFs. Early object detection was too general for GIPHY's culturally specific and abstract content, but the team continued adapting available tools. They used Word2Vec-style embeddings with anonymized session information, treating sequences of searches, GIF views, and clicks like text. This produced semantic relationships between GIFs, queries, and expressions, including playful vector relationships involving cultural references.

### Product engineering matters to Nick because products create shared experiences
[28:10](https://www.youtube.com/watch?v=89gd1r_dVuU&t=1690s)
Nick describes himself as a product engineer because he wants ideas to emerge as things people can use. He compares product work with music: making something can bring personal satisfaction, but it also communicates an experience to other people. A product can address a pain, solve a problem, or create joy. This preference explains why he moved away from infrastructure problems that were necessary at GIPHY's scale but were not where he wanted to spend his time. He moved toward special products and R&D, where he could explore new interactions and internal tools.

### Large language models make prototypes easier, but they do not remove product judgment
[29:07](https://www.youtube.com/watch?v=89gd1r_dVuU&t=1747s)
Nick agrees that language models let product engineers build useful demonstrations without waiting for a full machine learning or platform effort. A prompt can classify messy language that previously required incomplete substring lists or custom rules. These systems also allow interfaces to move beyond strict true-or-false buttons toward more natural, statistical interactions. He still chooses technology from the product's needs. A proven smaller model is the right choice when it already delivers the value. An LLM is useful when no existing tool fits or when it removes a large amount of implementation work.

### GIPHY experiments with generative media while accepting that quality is uneven
[39:59](https://www.youtube.com/watch?v=89gd1r_dVuU&t=2399s)
GIPHY uses machine learning throughout search, including embeddings that find matches beyond tags. It also has a model that recommends tags for uploaded GIFs based on existing tagging structure and popular queries. Nick says the team has been experimenting with generative products, including an internal generative GIF maker that works through a Slack-style bot. He does not consider generative video solved. Quality varies, and the team is waiting for better speed and output while using open source models such as AnimateDiff for experiments. GIPHY's history also makes the team more tolerant of strange or low-quality results when they create playful reactions.

### Startup teams need a specific user and a gap that large technology companies will ignore
[48:56](https://www.youtube.com/watch?v=89gd1r_dVuU&t=2936s)
When Nick advises startups, he helps with prototypes, code, product reviews, and engineering team structure. He says alignment and respect matter more than simply assigning tasks. Engineers need to understand why the work matters and then have room to solve the problem. He also thinks the funding environment has changed since GIPHY raised money. Investors now care more about revenue, user traction, and numbers, while large technology companies provide increasingly broad infrastructure and AI capabilities. His advice is to build for a specific user and look for edge cases that larger companies are unlikely to prioritize.

## Notable quotes
- Nick Hasty: "I don't understand this, I don't know enough, and it takes a lot for me to be satisfied that I actually know something." (18:39)
- Nick Hasty: "I like to make things for people to use." (28:10)
- Nick Hasty: "The tools are chosen out of necessity and not because you want to use some tech." (33:16)
- Nick Hasty: "I think you've got to have your North Star: what am I actually trying to build and what experience am I going for here?" (38:09)
- Nick Hasty: "You have to be real specific about your user and you probably have to be building at those edge cases a little bit." (51:23)

## Tools & references mentioned
- GIPHY
- Alex Chung
- NYU's Interactive Telecommunications Program
- Rhizome.org
- New Museum of Contemporary Art
- Tumblr
- Slack
- AWS
- Word2Vec
- K-Nearest Neighbors
- OpenAI
- DALL-E
- AnimateDiff
- Facebook
- Reddit
- Twitter

## Who should watch
- You are building search or recommendation features over messy cultural content and want an example of using behavioral data and embeddings.
- You are a product engineer deciding whether an LLM is justified for a feature that could also use a smaller, proven model.
- You advise an early-stage startup and want Nick Hasty's practical view on team alignment, prototyping, funding, and competing with large technology companies.

## Related talks

- [Creative AI: Using ML to Create Art, Music, and Jokes](https://mlopstalks.com/talks/creative-ai-using-ml-to-create-art-music-and-jokes) (Suyash Joshi, MLOps Community, 1:12:57)
- [Adventures in Building CLIP & Other (Largeish) LMs](https://mlopstalks.com/talks/adventures-in-building-clip-other-largeish-lms) (Sachin Abeywardana, Canva AI, 1:06:38)
- [Building Effective Products with GenAI](https://mlopstalks.com/talks/building-effective-products-with-genai) (Faizaan Charania, LinkedIn, 53:14)
- [Navigating Through the Generative AI Landscape](https://mlopstalks.com/talks/navigating-through-the-generative-ai-landscape) (Azin Asgarian, Georgian, 20:38)
- [Founding, Funding, and the Future of MLOps](https://mlopstalks.com/talks/founding-funding-and-the-future-of-mlops) (Mihail Eric, Storia AI, 57:31)