LLMs aren't the end-all and be-all of AI

Roie Schwaber-Cohen from Pinecone on the "Practical AI" podcast. Full audio 👉 practicalai.fm...
Subscribe for more! 👇
Apple: practicalai.fm...
Spotify: practicalai.fm...
Android: practicalai.fm...
Overcast: practicalai.fm...
Email: practicalai.fm...
Twitter: / practicalaifm
Mastodon: changelog.soci...
#podcast #ai #machinelearning #datascience #artificialintelligence #deeplearning #ml #mlops #nlp #dataengineering

[@WmJames-rx8go]

I have an idea and observation I would like to share with you. It's about developing a particular type of neural network to be used in large language models.
In many neural networks each input is mapped to each node in the hidden layer and each hidden layer is mapped to another adjacent hidden layer until the output layer.
I propose that the hidden layers be partitioned into groups and that in between input layer and the hidden layers there are placed logical circuits or the equivalent thereof. And that their output would be sent to the adjacent hidden layer, etc..
By training on a network with this type of configuration each partition of the neural network would take on qualities that are specific to a category of sorts. So if for example an XOR circuit came between two partitioned layers it would be able to prevent both layers from operating at the same time, because of course, XOR is equivalent to saying statement A or statement B but not both.
Of course you would want to consider using any logical circuit such as , NAND, OR, NOT, AND. As you well know any logical circuit can be built from any of these logical components so it would not necessarily be useful to mix all the components, however, it would probably be a considerable help to be able to mix them so that when the analysis of the neural network was done you would have some clue as to what section was working and choosing to produce the output and why it may have done so.
This concept is very similar to a famous NSAT problem, although it is strictly different as you are aware. However, it does have some of that flavor.
As an extremely simple visual aid I ask that you picture a neural network divided into sections and separated by a NOT circuit.
If I were training the circuit with an input, "a cartoon of a dog is not an actual dog", the NOT circuit would prevent a section of the neural network from outputting something like, "a cartoon is an actual object".
Note:
Emailed to various parties developing neural networks including, Open AI, Google, Meta, Anthropic, Amazon and various comment sections on the internet.
Second note.
Many of these companies make it difficult to contact any person. If for some reason this idea should prove useful but did not reach your company, I hate to be so straightforward and bold, but I say it's all on you.