ChatGPT the man behind the curtain
Is anyone else like this?
I am a computer professional. I write code and play nurse-maid to complex computer systems of various kinds.
I also was diagnosed as “different” when I was 8 or 9 years old. My mother summed up a lot of testing and expert opinions as “your brain is just wired a bit differently”. This made me ask the question: “how, exactly are brains wired, and if mine is a bit different how, exactly is it different and in what way”. One reason for my interest in computers is around that central question. In some ways, all computer programs and hardware are an attempt in imitating a brain’s wiring, at least a little.
In the film “The Wizard of Oz” there is a scene where Dorothy and company are addressing (what they think is) the Wizard, Oz the Great and Powerful, a great floating head. Toto, the little dog jumps out of Dorothy’s arms and reveals a not-great and not-powerful man hiding behind a curtain operating cranks, and leavers and speaking into a speaking tube. Oz booms “Pay no attention to the man behind the curtain”, but of course the man behind the curtain is the real secret to the real Oz.
The brain is like the man behind the curtain. But the curtain is much harder to peek behind.
The real me has many complicated inner workings that I may not notice at first. For better or worse I am the type who is always trying to peek behind the curtain of how human brains and computer imitations really work.
So sometimes I ask ChatGPT about how ChatGPT works:
Me:
Please tell me more about the unchanging architecture of ANN. I read where Microsoft had PR problems with AI offerings because they were too flexible and were taught things that some found offensive.
G (ChatGPT):
Artificial Neural Networks (ANNs) in their foundational form are structured in a way that, once trained,
they become relatively "fixed" in how they generate outputs. This unchanging architecture can be broken
down into two major stages: training and deployment.
1. Unchanging Architecture in ANN:
- Structure: ANNs consist of layers of nodes (neurons), usually an input layer, hidden layers, and an
output layer. During training, these neurons adjust their weights (connection strengths) through
backpropagation, learning patterns in the data.
- Training: In this phase, the network is given input data and adjusts its internal weights based on
the error between its predicted output and the actual desired output. After many iterations over a
large dataset, the model converges and becomes ready for deployment.
- Deployment: After training, the architecture becomes static. The weights are set, and the model
generates outputs based on learned patterns without further changes unless retrained.
- Inference Mode: Once deployed, the model is used only for inference, generating outputs based on
fixed weights and no longer learning or adapting.
This fixed nature is a strength for stability and predictability. However, too much flexibility in some
AI systems can cause issues, as seen in Microsoft’s AI models.
2. Flexibility Problems in AI Systems:
Some AI systems, like Microsoft’s "Tay" chatbot, were designed to learn from user interactions, which led
to problems.
- Case of Microsoft’s Tay:
In 2016, Tay was launched to learn from interactions with Twitter users, but it quickly generated
offensive content. Without proper filtering or safeguards, Tay mirrored harmful language from users,
causing a PR disaster.
- Issues with Flexibility:
- Real-time Learning: Tay dynamically adjusted based on user interactions, which can lead to problems
when harmful inputs are present.
- Lack of Safeguards: Tay did not have sufficient filters to block offensive content, so it replicated
toxic behavior.
- Unintended Consequences: When AI models are overly flexible and unsupervised, they may reinforce
undesirable behaviors.
3. How AI Flexibility is Managed in Other Systems:
In contrast, many AI models use controlled training:
- Supervised Learning: Most AI models are trained on curated datasets to avoid harmful inputs and
operate in inference mode after deployment.
- Retraining: AI models can be retrained with new data to improve performance, but this process is
controlled.
- Filters and Moderation: AI systems often include content filters to prevent the generation of harmful
or biased responses.
4. Balancing Flexibility and Control:
- Controlled Adaptation: Some AI systems learn from users over time in a constrained manner, like
recommendation systems.
- Monitored Feedback: Continuous learning can be effective with proper safeguards and monitored
reinforcement learning.
Conclusion:
Traditional ANNs are fixed after training, but excessive flexibility can lead to issues like Tay’s case.
Effective AI design balances flexibility with control to prevent misuse while still adapting to user
needs.
_________________
ADHD-I(diagnosed) ASD-HF(diagnosed)
RDOS scores - Aspie score 131/200 - neurotypical score 69/200 - very likely Aspie