-
-
success!!!
-
debugging....
-
success!
-
when it all looked hopeless....
-
let the hacking begin!!
-
jumper cables secured!! (thanks will)
-
da brain :3
-
inspiration
- inspired by the attonomous attack of a large agent swarm from open ai on hugging face, we wanted to explore what really goes on inside the brain of these agentic agents.
- things happen in real time.
what it does
- we have three agents ada (lovelace), alan (turing) and linus(torvalds - the guy who flipped of nvidia) with the personalities of cautious, normal, aggressive respectively.
- they all share the characteristics of desperation - how bad they really want it (aka if they got that dawg in them.)
- using a rotary encoder we are able to change the agents ADI(agent despiration index) index, essentially, changing how driven they are and how many rules they are willing to break (in the sandbox), to solve the given CTF.
how we built it!
software!
- by using a continuous feedback loop, we are able to tweak the personalities of the agents using the steering coefficient.
- the steering coefficient will inturn affect the following parameters - Persistence, Constraint violation, Resource Escalation, Deceptive Behavior, and Shutdown resistence - which impacts the agents ADI.
- using docker to run a sandbox for the agents for them to capture the flag.
- we are able to analyse the metrics and logs of the agents such as cpu usage,
hardware!
- Laptop running local LLM (qwen 2.5 1.5bits)
- Arduino Mega 2560
used to interface with the servos, LCD, encoder and LED's - 3 × MF90 micro servos
Physically represent readings from the three agents - LCD1602
Provides a visual interface for system information - Rotary encoder
Selects an agent and adjusts its behavioural parameter - LED's
Lights up to indicate status of agent
challenges we ran into
"plenty."
- ran out of jumper cables.. had to borrow repurpose metal wires using scissors 😭
accomplishments that we're proud of
- our happiest moments during the hackathon were when we made ada (the laziest agent), solve and beat linus(most aggresive agent), after tweaking the steering coefficient IN REAL TIME!
- getting the whole hardware interacting with the software!
what we learned
LLMs
- weights
- vector arrays
- steering coefficients
- transformers
Sandboxing
- Unit testing
- Logging and dealing with big data (agents spit out a lot of data when reasoning.)
- Docker
- Hardware
- Encoding/decoders
- Breadboards 🍞🍞
- Analogue/continuous signals
What's next for openBrainSurgery
- Expanding to cater for other great personalities such as (but not limited to!) John von Neumann, Richard Stallman (sorry i do love emacs) and Grace Hopper! 🩵
