AI Engineer Summit 2023
Announcing the AI Engineer Network
Loading slides…
Read the talk
Announcing the AI Engineer Network
At the inaugural AI Engineer Summit, Network adds LLM-based attendee matching to an event app, using the problem in a participant’s profile as the starting point for a connection.
From a talk by Benjamin Dunphy
An emerging ecosystem in the room
Who has gathered to build the emerging field of AI engineering? The AI Engineer Summit 2023 opens with two days of talks from founders and engineers, with Swyx due to set the technical context. Benjamin Dunphy begins by looking around the room. AutoGPT supplies an immediate example of how quickly the ecosystem is changing: a few months after being an open-source project, it is the summit’s presenting sponsor. Dunphy calls it the fastest-growing open-source project in history, without specifying a growth metric or comparison period, and points ahead to announcements from Toran.
Supabase brings a different combination: playful social media and product launches, alongside databases whose ease and trustworthiness Dunphy associates with being open source. He says the company became a diamond sponsor despite its policy against event sponsorships, then asks Paul and Ant to let him know how the policy change works out.
Fixie makes the possibilities more concrete. Dunphy wants to ask its booth sidekick whether it can see what he is wearing and recommend wardrobe updates; he nicknames the sidekick Derek. The visual interaction remains hypothetical. If vision is not available yet, he offers Matt and Ben an introduction to Logan Kilpatrick for a possible Vision API preview. This is an October 2023 prospect: OpenAI’s public vision API announcement followed in November with GPT-4 Turbo image inputs through gpt-4-vision-preview.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From investment to conversations
Dunphy credits Microsoft’s investment in OpenAI with helping commercialize the movement. Separately, a later Microsoft-hosted report describes $13 billion in cumulative funding across 2019–23; the contemporaneous partnership announcement described a multiyear, multibillion-dollar investment. For founders in the room, the immediate opportunity is a conversation with Microsoft for Startups representatives in the Carmel Room, past the elevators behind the audience. GitHub’s neighboring booth provides a destination for Copilot questions. These are introductions to teams, rather than a promise of startup funding.
Cloudflare gets a belated birthday greeting, with Dunphy joking that it will be going through some changes. Its booth is next to Microsoft in Carmel, across the elevators. He then recognizes Alessio Fanelli, the venture-capital half of the Latent Space Podcast, whom he describes as no less technical than Swyx. Perhaps the talks will draw Fanelli back into founding—though Dunphy suggests a software startup this time, leaving hardware to the experts.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Can already useful conversations become more useful?
Dunphy describes the inaugural summit’s audience as 500 selected attendees. They are experienced software engineers experimenting with production AI, or founders building AI-enhanced or AI-native applications and companies. That shared experience matters when attendees meet: the person beside them is likely working through related practical questions. He also recognizes livestream viewers and points them toward the next day’s opening keynote for announcements about expanding participation in 2024.
Conferences already do some of the work of finding relevant people. They bring experts from around the world together around a narrow subject, making many conversations useful before anyone has arranged an introduction. The next question is whether that concentration of expertise can produce better connections still. Shared interests make a conversation relevant; a stated problem gives a connection a purpose.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A profile problem becomes the basis for a match
The Network application starts with the practical information attendees need on their phones: the conference schedule, the full expo partner list, other event information, and recommended restaurants, bars, and cafés within walking distance. On top of that event companion, Dunphy announces AI-enhanced attendee matching.
The feature is called the Generative Matching Algorithm, or GMA. Dunphy gives the name a plain explanation: it uses LLMs to match attendees with the right people. The announced interaction is straightforward:
- Enter the problem you want help solving in your Network profile.
- Let GMA use that stated problem as the basis for matching.
- Connect with people the system identifies as able to help.
The problem field is the substantive input described here. The announcement does not specify a model, prompt, ranking method, or evaluation of match quality.
Dunphy directs attendees to the conference’s Network download page. At launch, matching is limited to people attending this summit in person. That restriction defines the available pool for the announced feature, even though the app’s eventual ambitions extend beyond this event.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Opening the app and matching algorithm to other events
The final announcement expands the scope: Dunphy announces that both the event app and the matching algorithm are being open sourced. The goal is a simple, powerful mobile experience for attendees at events around the world, with the matching component helping those attendees find one another.
The implementation credits separate the application, matching, and infrastructure responsibilities:
- Mobile app: Simon Sturmer, credited as lead engineer.
- Generative Matching Algorithm: Suizette Steller.
- Authentication: Descope.
- Database and vectors: Supabase and pgvector.
Dunphy invites attendees to talk with the partners outside about pgvector. Its inclusion identifies part of the infrastructure, but the announcement does not explain how vectors participate in GMA’s matching decisions.
After thanking the developers and partners, Dunphy announces that the repository is available that day through the Network landing page and GitHub. The closing slide supplies the destination: aiDotEngineer/Network. The release invitation carries the idea beyond the summit: other events should be able to provide the mobile experience and help participants connect around the problems they bring.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Source repository for the mobile event application and its attendee-matching project.
Historical event page with the program, speakers, workshops, venue, and presenting organizations.
Postgres extension for storing vectors and performing similarity searches, with installation and SQL examples.
Further reading
Current guide to enabling pgvector and storing embeddings in a Supabase database.
Current starting guide for adding authentication with Descope Flows or SDKs.
Updates since the talk
DevDay announcement introducing GPT-4 Turbo image inputs through the Chat Completions API.
Full transcript
Loading transcript…