Artificial Intelligence is a wildfire in today’s technological advances. The new concept out there is Constitutional AI (CAI) which will inflame new startups.
What is it all about, how does it work, and Why is it “Constitutional”?, We will find out in today’s blog.
What is Constitutional AI?
We all have experienced ChatGPT and its immense power.
It clearly has displayed its prowess and speaks volumes of how AI has advanced in the recent past. They have become capable and fast.
Leveraging this capability of AI, a group of Ex-ChatGPT team is working on a model to supervise other AIs.
What does it mean? Does AI have managers now?
In a way, YES!
The experimental AI model is training itself with self-improvement methods to become a harmless AI assistant.
The only human involvement will be through defined rules and principles.
This self-improving harmless AI training methodology is “Constitutional AI.”
In the days to come, this may become a true competitor or enhancer of ChatGPT.
Read the research here from Anthropic.
Why use the term ‘constitutional’?
When it comes to the development and deployment of general AI systems, Anthropic suggests, it is important to consider the concept of a constitutional approach. This approach emphasizes the need for establishing a set of guiding principles, also known as a constitution, to govern the system.
Furthermore, the team also suggests that they chose this term because it highlights the ability to train less harmful systems by specifying a short list of instructions or principles.
It is also important to note that even if the principles governing the AI system remain hidden or implicit, they still exist and impacts its behavior. That is why the term constitutional – to remind developers and stakeholders that when creating a general AI system, it is impossible to avoid the choice of some set of principles to govern it.
Implementing a constitutional approach to AI, in the long run, may aid in the creation of responsible, trustworthy, and transparent AI.
How does the Constitutional AI training happen in this model?
Two key phases:
- Supervised Learning Phase (SL Phase)
Step 1: The learning starts using the samples from the initial model.
Step 2: From these samples, the model generates self-critiques and revisions.
Step 3: Fine-tune the original model with these revisions. - Reinforcement Learning Phase (RL Phase)
Step 1: The model uses samples from the fine-tuned model.
Step 2: Use a model to compare the outputs from samples from the initial model and the ‘fine-tuned’ model.
Step 3: Decide which sample is better. (RLHF)
Step 4: Train a new “preference model” from the new dataset of AI preferences.
This new “preference model” will then be used to re-train the RL (as a reward signal).
It is now the RLAIF (Reinforcement Learning from AI feedback).
Using this methodology, the team at Anthropic can train the AI assistant, which specializes in harmlessness.
The model takes a step further. It engages harmful queries and politely declines harmful outputs.
In ChatGPT, that is not possible. Sam Altman’s OpenAI assistant answers every query by bypassing the harmless filter.
Follow us for more updates on Constitutional AI
If you wish to experiment some AI, try these…
1 – WordAI
2 – Rytr
3 – AI-Writer
4 – Writesonic
5 – Paragraph AI
6 – Pictory
7 – Inkforall
8 – Tugan.ai
Conclusion
In general, conversational and generative AI agents allow us to easily have a human-like conversation with a computer on the topic of our choice.
But with the introduction of the harmlessness feature where the AI self-improves harmlessly, it will have many use cases in the future with very little human feedback.
ChatGPT chatbot is now just over a month old (November 2022) and will soon be left behind with the release of this new harmless AI assistant.
Very soon, Google may be in trouble too. We will have to watch out for how this model self-improves.
The concept is new (December 15, 2022). Hop on to this train, and let’s see how AI self improves on this journey of the internet’s new machine learning language models and AI assistants.
It will leave GPT-3’s natural language behind with more transparency of AI decisions.
Don’t forget to subscribe to hear more about how the Constitutional AI concept will self-learn to grow over nonsensical answers and human-judged performance.
Hoomale has several sections, including Corporate Culture & Leadership, Generation Alpha Mindset & Behaviour, The Future of Work & Technology, and more. Every section features interesting and thought-provoking articles that are sure to appeal to anyone who is interested in learning more about these topics.
If you wish to receive an email when we post the next, consider using the below form.
Disclaimer: Some of the links in this post may be affiliate links, which means that if you click on the link and make a purchase, we may receive a commission at no additional cost to you. Please note that we only recommend products and services that we have personally used and believe to be of high quality. Thank you for your support.
