November 2024, Issue 6
Path to AGI
What do researchers miss when they try to design a system that conforms to AGI principles? (besides the part where they focus on LLMs too much)
Modeling chemical synapses
and wrong brain maps in Devlog #9
The GeForce RTX 40 Series has already reached the number of 76 billion transistors, so this video card can be compared with a human brain that has 86 billion neurons. A human with a slightly damaged brain. Something like too much alcohol at college parties.
Discrete Spiking Networks
First, we need to clarify how to work with the signal propagation in time. From one side, we use discrete time points, moments when a neuron has transmitted the action potential to following dendrites, but we need to track the dynamics of how values change when new inputs reappear on some neurons while the previous chain reaction is still propagating. Feedback signals and loops are essential features of this model, and that is what makes it an open problem here.
Latest
Ansible grimoire
A short example how in Ansible you can read file content into a variable and assign that variable to stdin. Why? We just setup crontab with Ansible.
Unsupervised image classification
I found a post about blobs that serve as a blurred contour for locating objects in pictures. The idea is that we can take that generator and prepend a fully connected network that maps Gaussian noise (technically that is the vector with features?) into the spatial map with blobs.
Reinforcement Learning
Walls in the grid environment is an interesting problem. In rsrl crate it was implemented that the movement is possible but not movement actually occurs, so the agent just stays in the same position. It's not good.