Takeaways from AI Dev 25
I attended the inaugural AI Dev 25 conference in San Francisco on March 14, the Pi day. The conference was organized by deeplearning.ai, an edtech company focused on AI education and developer community, founded by none-other-than professor Andrew Ng. I took the very first online ML class by prof Ng in 2011 on Coursera. I have been following prof Ng since then. In the summer of 2023, I took all the ~10 available AI classes on deeplearning.ai, and came away more impressed by the capabilities of LLMs than I expected. Now there are countless classes on deeplearning.ai and new classes are introduced almost on a weekly basis, all teaching the latest technologies, while indicating how fast and dynamic the space is developing.
The conference took place in Shack 15, on the upstairs in the Ferry Building, with nice views of the bay and the bay bridge. The venue was not large with several meeting halls plus a lounge area. There were ~500 attendees (tickets were sold out fast), consisting of mostly developers plus some business folks. I met a number of people flying in, even from abroad.
Prof. Ng kicked off with a 5 minutes keynote, which imparted several important messages:
This is the first app developer focused AI conference. There are academic AI conferences and there are conferences by specific companies show-casing vendors’ own technologies. But AI Dev 25 aims to serve the budding developers community for building AI apps, and equip them with the latest and vendor agnostic tools.
This is the golden age for AI app builders, with all the new tools as building blocks. Prof. Ng likened this to having a large variety of lego blocks which allow one to build ever more complex and interesting structures.
Addressing the common concern that with the AI coding assistants, one no longer needs to learn programming anymore, Prof. Ng says this will go down in history as one of the worst pieces of career advice. As programming gets easier, more people should code, not less. He again made an analogy that his friend who is an art historian can use AI to create much better art-work than he can, because she knows the language of art and can express her thoughts and command the AI tools much more effectively in creating better art-work.
With AI, building app prototypes can be 10X faster, while productizing an app would only be ~50% faster. He encourages people to prototype more as it takes far less time and resources than before, to move fast and be responsible.
Bill Jia from Google (previously at Meta), and Chaya Nayak from Meta also gave keynotes highlighting Google and Meta’s AI assets:
Google - Astra and Deep Research agents
Meta - Llama models and Llama Stack
The panel with Michele Katasta (Replit), Percy Liang (Stanford and Together AI), Roman Chernin (Nebius), Thomas Wolf (Hugging Face), moderated by Laurence Moroney (ARM) was filled with tips and prognostications:
We need to design new infra to run apps built by agents (Roman)
Current frontiers of agents - tool use, time spent thinking and problem solving, RL for self-improvement (Percy)
There will be different levels of agents, like the 5 levels of autonomous driving (Percy)
Agents for not only code-gen, but also deployment and monitoring, the full dev lifecycle
Evals with current benchmarks useful at the component level but not enough, need evals with end-to-end business metrics for the whole system, although Perplexity and MMLU do correlate with system quality (Percy)
Besides Agents, Robotics and AI for Science will also be huge in coming years
AI tech will keep changing, so we should really focus on data, insights, domain expertise and intuition.
I particularly enjoyed the two workshops:
Harrison Chase from LangChain held a workshop to hack out an automated email responding agent in an hour, using the new memory capabilities with LangGraph. There are three types of memories - semantic (e.g. facts), episodic (e.g. experiences), and procedural (e.g. skills). He gave examples of how to espouse these types of memories to make the agent smarter and more personalized in processing and responding to incoming email requests.
Joao Moura from CrewAI gave a workshop on how to build a multi-agent system, by orchestrating agents in a workflow to accomplish more complex tasks. He hacked two examples - 1) a business meeting prep app that gathers information about the business contacts, their responsibilities, positions and concerns, and craft a meeting agenda with topics and discussion points; 2) a research report app that fans out research tasks to multiple agents that gather info from different sources, and consolidate the findings into a well crafted research report.
Some takeaways from the startups hosting presentations and exhibit booths:
Replit offers a cloud platform for coding assistants and app building environment
Arize provides AI observability and eval frameworks
Pinecone is a fully managed vector database
Chroma provides open source database for AI apps, for embeddings, vector search, documents storage and full text search
Haystack is any open source dev framework for building agentic apps and pipelines
Codium Windsurf is an coding assistant IDE
Landing.AI (by Prof. Ng) provides a visual AI platform, transforming images and videos into intelligence
Vectara provides AI agents platform for enterprises
A number of public companies were also present:
MongoDB - recently acquired VoyageAI (from Stanford) to integrate its embedding and reranking models with rest of MongoDB offering.
Snowflake - while it’s a major data cloud for enterprises, I think there is room for it to increase its developer mindshare
Qualcomm - its AI Hub makes it easy for developers to test their AI apps running on hundreds of edge devices (with Qualcomm chips)
AMD - MI300 series for inferencing, and ROCm (AMD’s answer to Nvidia CUDA)
IBM Research - building Agent Communication Protocol to solve the agents discovery and interoperability problem
Overall, it was a very engaging, informative, and fun event for AI app developers.
