Wait, #LLMs or #NeuralNets can't reason!? This is so surprising! How could that happen!? This is unbelievable! I'm in shock!

But they surely have feelings, right...right!?

#PromptEngineering is a task best left to #AI models | The Register @theregister / Thomas Claburn @thomasclaburn
https://www.theregister.com/2024/02/22/prompt_engineering_ai_models/

"(...) Machine-learning boffins find #OpenSource #NeuralNets can optimize their own queries (...)

enlisting an #LLM to refine prompts for improved performance on benchmark tests. (...)

"The point of the research was to discover if smaller, open source models can also be used as optimizers," explained Battle, "And the answer turned out to be yes." (...)"

I'm giving a short advertisement for my work on #NeuralNets at #JMM2024 in about 15 minutes! I applied a bit late and ended up in one of the contributed paper sessions. These are already a bit more subdued, but this one is particularly poorly attended. Perhaps it's for the best, since the videos on my YouTube channel (link in bio) are much more detailed than what I'm about to say.

My talk from yesterday on a categorical semantics for neural nets for the New York Category Theory Seminar has already been posted on YouTube! You can find it at https://www.youtube.com/watch?v=FKkpVKuspmA and you can see more about this seminar at https://www.sci.brooklyn.cuny.edu/~noson/Seminar/. The preprint I mention is https://arxiv.org/abs/2308.00677 and the talk by Joyal which was mentioned at the end can be found at https://www.youtube.com/watch?v=MxClaWFiGKw.

I made my first post to the #AI section of the arXiv this week! You can find the preprint "Discrete neural nets and polymorphic learning" at https://arxiv.org/abs/2308.00677.

In this paper a learning algorithm based on polymorphisms of finite structures is described. This provides a systematic way to choose activation functions for neural nets whose neurons can only act on a fixed finite set of values. These polymorphisms preserve any specified constraints imposed by the learning task in question.

This paper is the result of a 2021 REU at the University of #Rochester. I am working with a great group of students on a follow-up project right now, so videos of talks and a sequel preprint should be out soon!

Data science & Groovy using #ApacheBeam #ApacheCamel #ApacheCommons @ApacheGroovy #ApacheIgnite #ApacheMXNet #ApacheOpennlp #ApacheSpark #ApacheWayang #Datumbox #deepjavalibrary #DeepNetts #EclipseDL4J #graalvm #gradle #stanfordnlp #TensorFlow #Tribuo #smile #tablesaw #opencsv

https://www.javaadvent.com/2022/12/groovy-and-data-science.html

Covers data manipulation, regression, clustering, classification, natural language processing & object detection #jsr381 #visrec #ai #ml #groovylang #neuralnets #deeplearning #javaadvent22

After the initial excitement has set by now, more and more people have raised the issue that the GPT-3 based chatbots are able to make statements that sound convincing, but may have no bases in reality.

This makes a lot of sense to me, as AI, in its core, is based on one thing: pattern recognition.
So really, all it does is cargo cult, not learn or necessarily understand.

As a personal project, I'd like to learn #julialang and develop simple #neuralnets with eventually increasing (?) complexity and biologically inspired dynamics. Or at least without backprop!
I've only used #numpy #keras and #tensorflow in uni or on the job, and then some #pytorch studying GANs on coursera.
I think Julia could be fast and smooth working with ODEs and matrix multiplications.
Anyways, suggested sources and tips?

Conjecture that is likely true, and damning for large language models presuming that it is: An LLM trained strictly on truth will still confabulate, because it will break the bindings in what it saw, in the interpolation process, and continue to fabricate.

Further conjecture; if the above is (as I suspect true), will we never see fully honest LLMs (though they might be used as components in larger systems)