About
Background: B.Sc. Physics | Law (LLB) | M.Sc. Com Science | Founder
Interest: Scaling social and digital systems under competing and incompatible philosophies.
Projects
Model Soup
An evolutionary framework that automates hypothesis generation, mutation, and verification. LLMs propose executable prediction models into dual pools (parameter-based and neural), which compete under a falsifiable cross-validated score and recombine across generations.
Tested on multi-decade London housing data with unstructured news: the loop converges in a few generations on simple, high-scoring forecasts. github.com/lucastsui/ModelSoup
Self-Improving Factory
I make self-expanding and self-improving simulated factories.
Approach one is to encode design knowledge into the architecture so that the factory can scale indefinitely without exploding in complexity. The results are in github.com/lucastsui/RINA-Recursive-Factory.
Approach two is to use AI to explore the world and research optimizations autonomously. The project is ongoing. It currently has around 120 types of intermediate products and around 1M moving intermediate products on average. Current layout of the factory: claude.ai layout artifact.
AI-Assisted Atmospheric Modeling
The exosphere is the outermost layer of the atmosphere that protects satellites and ground electronics by absorbing and scattering the energy of solar storms.
As research assistant at Boston University, I am analyzing new observational data from the Carruthers satellite to improve our model of the exosphere to better predict electronic disruptions due to solar weather.
The goal is to test how far AI can go in developing and explaining new scientific theory from first principles in physics.
Autonomous Asset Allocation
By deploying in-house AI, I experiment with transforming a real estate management company into a portfolio management company, evaluate asset quality, and make buy/sell/hold decisions. Current effort is in devising rules that survive long feedback loops with noisy market signals.
Björk Cube
A hackathon experiment to use social-media methods to understand the internal space of Magenta RealTime 2, a music generation model created by Google.
06
Splat Room-Edit
An experiment that turns a 20-second phone video of a room into a 3D Gaussian splat you can walk through in a browser, then recolors just the wall with one command while the art and furniture stay untouched.
Collatz Conjecture Research
An autonomous AI research loop that systematically attacks the Collatz conjecture: predict before computing, hunt for surprises, adversarially attack emerging patterns, and ban approaches that circle known dead ends. github.com/lucastsui/collatz-research
Socrates
This is an application to accelerate learning compared to passive lectures and brute-force recall practice by implementing pedagogic theories such as ZPD-based question targeting, cognitive load theory, the Socratic method, and related ideas. It progressively trains a student's knowledge-recall capability by tracking their mastery of a subject and asking questions appropriate for their level of understanding. A knowledge graph is generated dynamically to trace the background knowledge necessary to master a concept. github.com/lucastsui/socrates
AI Jailbreaking Research [Close Source]
An AI security research project that successfully prompts a frontier AI (Grok) to hack a video game with kernel-level anti-cheating protection. Methods could be disclosed for security research purposes.
In-house AI Deployment & Ops
I AI-ify my company AYQ's information process by building an in-house AI system, solving challenges like model selection, evaluation, deployment, choosing the tradeoff between intelligence and context window, tools and skills settings, adopting prompt engineering architectures such as ARCHON to enhance performance of open weight models to be on par with frontier models.
Microbait
Microbait helps jobseekers survive AI job-pocalyse by identifying what skills are made obsolete because of AI, and what skills become more demanded as a result.
It works by reading what users on X brag about in their latest tech demos and paper publications, and uses AI to summarize what skills are being automated, what the next bottleneck is as a result, and what job posts are already there as a result of the change in demand. github.com/lucastsui/microbait
