AI Agents Programming
Here are some important concepts to think about when developing AI Agents.
If LLMs and AI Agents do not work reliably, then what needs to be improved and changed?
My mini-articles which I am usually publishing in the AI groups
Humans Sensory Memory principles could be used in the Transformers based neural networks too.When processing time point data, the model remembers few previous time points latent data as well.The model learns to incorporate these earlier memories when working with current moment data.
Two independent neural networks could be trained together and their collaboration could happen via shared trainable memory. This is similar to JEPA approaches where you work with latent states.
Transformers Tokens could be fused for classification intelligently instead of using Mean Pooling, or picking last Token or [CLS] Token. Or picking last hidden_state of further constructs.Especially interesting approach is making a hologram from source Tokens.
USA Army needs will lead to more complex LLMs architecture development.None of current AI providers can deliver what Secret Services need to do their job.
Humans Sensory Memory principles could be used in the Transformers based neural networks too.When processing time point data, the model remembers few previous time points latent data as well.The model learns to incorporate these earlier memories when working with current moment data.
Two independent neural networks could be trained together and their collaboration could happen via shared trainable memory. This is similar to JEPA approaches where you work with latent states.
Transformers Tokens could be fused for classification intelligently instead of using Mean Pooling, or picking last Token or [CLS] Token. Or picking last hidden_state of further constructs. Especially interesting approach is making a hologram from source Tokens.
USA Army needs will lead to more complex LLMs architecture development. None of current AI providers can deliver what Secret Services need to do their job. Government LLM needs information and reasoning segments for various audiences. That leads to 10K narrowly specialized LLMs, and special LLMs to talk to various audiences - be it a citizen, government employee, or military officer. None of these specialized LLMs can be trained with all publicly available information like right now. In many government situations real information and propaganda will conflict. That will lead to reasoning beyond simple stupidity. No RAG databases or fine-tuning here can help. For Secret Services we need CLEAN branch of information, and to train LLMs to reason in completely different reality. At this USA Army and Anthropic dispute moment we have to make a decision - are we creating 2 realities LLMs, or just absolutely honest one like Elon Musk promotes. This should be public policy makers vote to decide which way we will go. OR, this is referendum level topic where the whole country votes. Good question is how to explain to the not technical people what this all means.
Neural Network could be made from the parts. Each part could be trained separately. Each part has it's designated function. Parts communicate via Interfaces.