15th Month @Stanford
December 5, 2024

TL;DR

At the lab, I caught up on research trends in Test-time Computing, attended conferences, and updated my knowledge of cutting-edge AI SaaS development. Additionally, I immersed myself in US culture through events like Thanksgiving and Election Day.

1. Research

11/01 Google DevFest Silicon Valley
11/05 Stanford Graph Learning Workshop
11/17-22 React Summit & JS Nation @NY

Conferences

The React Summit was a highly rewarding event where I got to hear live talks from @theo and @jherr, who have been pivotal in my self-taught web development journey. Most of the talks centered on new features in React 19 and Next.js 15, particularly enhancements to server actions and React Server Components (RSC), which have had a major impact.

A significant takeaway was the shift in form-building techniques with the adoption of Partial Pre-Rendering (PPR). The previously standard react-hook-form + zod combo is no longer viable for server-side code, prompting a reevaluation. A panel discussion featuring library developers from react-hook-form and tanstack revealed plans to adapt their tools to align with the trend of server-centric web app development. I’m excited to follow their progress.

Google DevFest focused heavily on AI, with the sessions on Vertex AI in Firebase standing out for me. Its well-designed management console offers inspiration for my SaaS UI design, and the AI Testing Agent is a tool I’d love to adopt for DevOps-intensive projects. Writing tests is tedious, so being able to resolve this with natural language is a game-changer.

Over the past two years, even though PPR (Partial Pre-Rendering) became partially available, it wasn’t practical for widespread adoption. I’ve primarily developed web apps centered around tRPC, but excessive API calls have led to performance issues, as seen with VIHub, affecting user experience. With Next.js 15 finally making PPR feasible, I plan to integrate it into my next projects.

Test-time Computing

This continues to be a hot area of research. One common approach involves generating multiple answer candidates using LLMs, then employing a Verifier to evaluate and select the best solution. Increasing the number of candidates boosts coverage, while a well-designed Verifier ensures precision.

The key lies in Verifier design. For tasks where correctness can be strictly validated (e.g., parts of mathematics), rigorous Verifiers akin to unit tests can be created, making this method highly effective. However, many problems are ill-suited for simple unit tests, presenting the challenge of building a high-performance Verifier (often utilizing LLMs for this). Questions arise, such as whether prompting the same LLM with chain-of-thought (CoT) reasoning for detailed verification suffices or if a dedicated Verifier model is necessary.

Among the many methods, I found OpenAI's Prover-Verifier Game particularly intriguing. Similar to GANs during the deep learning boom, this approach designs the Verifier using a Generator-Discriminator framework. This incentivizes the Generator to produce responses that are easier for the Verifier to validate, potentially enabling AI beyond AGI to generate responses aligned with human readability and validation standards.

At the end of the month, Alibaba released QwQ, an open-source model rivaling o1's performance. This development is expected to accelerate research in test-time computing. I also plan to explore LLaVa-o1, a multimodal OSS model.

Future Directions

2. Life

11/01 Diwali @Stanford
11/02 Bar Mitzvah @🇺🇸Harrison
11/05 US Election
11/09 Wedding Ceremony @🇨🇳Alan, Safari
11/28 Thanksgiving Party @🇺🇸Tom, Steffi

11/08,13,15 Dinners w/ Friends
11/17-22 @NY

I served as the Best Man at a friend's wedding, a moving ceremony held at Stanford Memorial Church. It was gratifying to see the website we collaboratively built gain recognition among many guests.

I also had the invaluable opportunity to experience US culture through events like the US Election 🔗, Thanksgiving, Bar Mitzvahs (Jewish), and Diwali (Hindu).