Using Simulated Fruit Flies to Discover Optimal Routes in Logistics Networks (WIP)

Using simulated fly brains to find optimal routes in logistics networks.

Google released a complete map of a fruit fly's brain. I trained a simulated version of the brain to haul cargo along a linear assembly line involving multiple types of intermediate products at successive stages.

The design challenge is how to train a brain adapted to the real world to operate in a 2D environment. Simply adding a decoder network to translate its motor neuron signals into up, down, left, and right controls has been ineffective. Whether dopamine-based rewards are more effective than backpropagation remains an open question. A combination of movement-specific training and training for overall goal satisfaction, using traditional backpropagation without dopamine-based rewards, has proved effective.

I used a cluster of 1 RTX 5090 and 2 DGX Sparks for training by using Nvidia PAIR, a computer clustering tool. AI has made the hardware communication part of the problem trivial.

The current direction is to generalize the training to assembly networks with greater depth and breadth, and more complex layouts.

Self-Improving Factory

Factorio is a factory automation game with >= 1M moving parts that gradually became my obsession. There is an underground competition where players try to making the biggest factory possible under the hardware constraints. As a die-hard fan of the game, whenever there is a new automation technology, I would try to throw it at the game and see how large of a factory I can make.

I have tried to apply an research networking architecture called Recursive Ineternetwork Architecture (RINA) and use Claude to design the optimal factory top-down. github.com/lucastsui/RINA-Recursive-Factory. The schematics work but the design is not space efficient, and it can't solve some minor but crucial nuances that are needed to make the design applicable in the in-game client.

After GPT Astra came out, I abandoned of notion of a perfect design and mimic the approach GPT used to hack Hugging Face by creating multiple agents, each with limited vision but they can move around the factory and communicate through the in-game chat box.

The "community" approach is more promising. The top-down approach took about 1 month to complete the design, while the agent swarm took only one day to make substantially progress which would take non professional players weeks to achieve.

My theory is that the top-down "galaxy brain" sees the game as a sea of game logs, while the Astra agents, actually interact with the in-game client using in-game controls. The abstractions they form over the game are different, and my intuition gained from my study in computer networks told me that selecting the right abstraction is half of the solution.

The experiment is ongoing.

Carruthers

I join the Carruthers Geocorona Imager mission of NASA to set up data pipeline with physical server deployment and monitoring involved. My role is to make trade-off between data storage and computation efficiency and help scientists quickly form intuition over observational data.

The tool helped discover a potential data calibration error in the mission's data pipeline.

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.

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.