Natural Language Conversations with Foundation Models (LLMs)
Andi Sama — CIO, Sinergi Wahana Gemilang
Along with the advancements in Artificial Intelligence (AI) research, generative AI, supported by various foundation models, has been in the mainstream news for the last two years. The multimodal capabilities of the Generative AI enable people (or systems) to communicate with the AI model in a text-based natural language conversation style. Given the text prompt, the AI model can respond with generated text, documents, images, voice, music, and even video.
Thanks to the availability of big data and Graphics Processing Units (GPU), which started in the 2010s, including continuous innovation in AI, machine learning, and deep learning to develop various advanced algorithms and neural network architectures since the 1940s.
The 2017 paper "Attention is All You Need" on deep learning Transformer architecture was a breakthrough for Generative AI. It led to the public initial release of OpenAI ChatGPT in November 2022 with its GPT foundation model. The term "Large Language Model (LLM)," the pre-trained generative model, was suddenly making the headlines. The trained LLM was also known as the foundation model. This neural network model has been trained with vast data (imagine the accessible Internet data as the dataset) and a lot of computing power (many millions of GPU hours). With such an extensive infrastructure, the training of LLM has been out of reach for many, only available for selected researchers and companies that can afford to spend millions to billions of dollars on Research & Development (R&D).
OpenAI's move to release the foundation model was quickly followed by others, such as IBM Watsonx with Granite and Granite-Code, Google Gemini, Meta.ai with Llama, Microsoft Bing and Copilot with GPT and DALL-E, and Samsung Gauss for the latest S24 flagship smartphone.
Today is July 4, 2024. Let’s try to have an introductory conversation with five of the available LLMs in the market
- OpenAI ChatGPT (foundation model: GPT, through web browser)
- Microsoft Copilot (foundation model: GPT, preview version) — installed on Windows
- Google Gemini (foundation model: Gemini, through web browser)
- Meta AI (foundation model: Llama, through web browser)
- Mistral AI (foundation model: Mistral 7b, runs locally on a local laptop)
Mistral is our experiment's only compact foundation model that runs on the local computer. All others run the models on the cloud and have model sizes way more significant than Mistral, thus requiring a considerable infrastructure to run (especially the need for GPU). Mistral runs on the local laptop with an Intel i7 CPU with 16GB RAM, 1TB SSD, and no GPU.
In this article, we refer to the model as the foundation model. Both have similar meanings.
We ask the same questions for all the models and then observe and compare their responses.
To start, let’s first have a conversation with ChatGPT.
OpenAI ChatGPT
You
Hi, my name is Andi. Who am I speaking to?
ChatGPT
Hi Andi! I'm ChatGPT, an AI language model created by OpenAI. How can I assist you today?
You
Nice to know you ChatGPT, what are your strongest capabilities
and how can you assist me?
ChatGPT
Nice to meet you, Andi! Here are some of my strongest capabilities
and ways I can assist you:
- Information Retrieval: I can provide detailed information on a wide range
of topics, including science, history, technology, and more.
- Writing Assistance: I can help with writing tasks such as drafting emails,
creating essays, composing poems, and even generating creative stories.
- Problem-Solving: I can assist with solving mathematical problems, coding
issues, and providing step-by-step explanations for complex concepts.
- Language Translation: I can translate text between various languages.
Conversation and Entertainment: I can engage in general conversation,
provide recommendations for books, movies, or music, and even tell jokes
or play word games.
- Learning Support: I can help with studying by providing summaries,
explanations, and answering questions related to various academic subjects.
Personalized Assistance: I can remember your preferences and tailor
my responses to better suit your needs.
How can I assist you today?
You
That's great. thanks for enlightening me. I will talk to you soon later.Sounds normal. Let's continue with Copilot.
Microsoft CoPilot
You
Hi, my name is Andi. Who am I speaking to?
Copilot
Hello, Andi! I’m Microsoft Copilot, an AI companion here to assist you.
How can I help you today? 😊
You
Nice to know you copilot, what are your strongest capabilities
and how can you assist me?
Copilot
Hello, Andi! I’m glad to assist you. As Microsoft Copilot, I have
a wide range of capabilities to help you with various tasks.
Here are some of my strongest abilities:
- Creative Writing: I can generate poems, stories, essays, code, and more.
If you need a creative piece, feel free to ask!
- Summarization: I can summarize articles, news, or any other content
you provide. Just share the text, and I’ll create a concise summary.
- Translations: I can translate text between different languages.
Let me know which languages you’d like me to work with!
Coding Assistance: If you have programming questions or need help with
code, I can provide explanations, examples, and even write code snippets.
- Information Retrieval: I can look up information on various topics
using Bing Search. Feel free to ask me anything!
- Image Descriptions: If you upload an image, I can describe its contents
for you.
- Problem Solving: I can help you solve puzzles, math problems, or any
other challenges you encounter.
- Entertainment: Need a joke, riddle, or fun fact? I’ve got you covered!
Remember, I’m here to assist, so feel free to ask anything, and I’ll do
my best to provide helpful and engaging responses! 😊
You
That's great. thanks for enlightening me. I will talk to you soon later.
Copilot
You’re welcome, Andi! Feel free to reach out whenever you need
assistance. Have a great day, and take care! 😊A similar response to ChatGPT. Let’s take a look at Google Gemini.
Google Gemini
You
Hi, my name is Andi. Who am I speaking to?
Bard
Hi Andi! I'm Bard, a large language model from Google AI.
It's nice to meet you! Is there anything I can help you with today?
You
Nice to know you Bard, what are your strongest capabilities
and how can you assist me?
Bard
I'd be happy to share my strengths and how I can be helpful!
Here are some areas where I shine:
Understanding and responding to your questions: I can answer your questions
in an informative way, even if they're open ended, challenging, or strange.
I can access and process information from the real world through
Google Search to keep my responses up-to-date.
Generating creative text formats: I can create different kinds of
creative text content, like poems, code, scripts, musical pieces, email,
letters, etc. If you're stuck on a writing project, I can help you
brainstorm ideas or get over writer's block.
Simplifying complex topics: I can take difficult concepts and explain
them in a clear and concise way that's easy to understand.
Translating languages: Need to communicate with someone who speaks
a different language? I can translate between over 100 languages.
Integration with Google products: I can access and process information
from your Gmail, Docs, and Drive with your permission, helping you find
and summarize information or complete tasks.
Overall, I'm still under development, but I'm always learning and
improving.
Here are some ways I can specifically assist you:
Research: Need to do some research for a project or presentation?
I can help you find relevant information and summarize it in a clear
and concise way.
- Writing: Struggling with a writing assignment or creative project?
I can help you brainstorm ideas, overcome writer's block, or even
generate different creative text formats.
- Learning: Do you have a question about a particular topic? I can explain
it to you in a way that's easy to understand.
- Communication: Need to translate something or communicate with someone
who speaks a different language? I can help you out there too.
Is there anything you'd like to try today?
You
That's great. thanks for enlightening me. I will talk to you soon later.
Bard
You're welcome, Andi! I'm glad I could help.
Feel free to reach out whenever you have a question, need some
creative help, or just want to chat. See you soon!Then, continue with Meta AI.
Meta AI
You
Hi, my name is Andi. Who am I speaking to?
Meta AI
Hi Andi! I'm Meta AI, your friendly AI assistant!
I'm here to help answer questions, provide information, and even
generate text for you. I'm excited to be chatting with you on this
Independence Day!
You
Nice to know you Meta, what are your strongest capabilities
and how can you assist me?
Meta AI
I can help with a wide range of tasks, such as:
- Answering questions: I can provide information on various topics,
from science and history to entertainment and culture.
- Generating text: I can create text based on prompts, topics, or styles.
I can even help with writing tasks, like proofreading and suggesting
alternative words or phrases.
- Translation: I can translate text from one language to another,
including popular languages like Spanish, French, German, Chinese,
and many more.
- Summarization: I can summarize long pieces of text into shorter, more
digestible versions, highlighting key points and main ideas.
- Creative writing: I can generate creative writing, like poetry or
short stories, based on prompts or topics.
- Conversation: I can have a chat with you, answering your questions
and engaging in discussion on a wide range of topics.
- Entertainment: I can even generate jokes, puns, and fun facts to
brighten your day!
These are just a few examples, but I'm here to help with almost any task
or question you have.
So, what do you need assistance with today, Andi?
You
That's great. thanks for enlightening me. I will talk to you soon later.
Meta AI
You're welcome, Andi! It was a pleasure chatting with you. Feel free
to come back and ask me anything or seek assistance whenever you need it.
I'll be here, ready to help.
Have a wonderful Independence Day and a great rest of your day!And finally, with Mistral AI.
Mistral AI
You
Hi, my name is Andi. Who am I speaking to?
Mistral AI
Hello Andi! You are speaking with a model trained by Mistral AI.
How can I assist you today?
You
Nice to know you Mistral AI, what are your strongest capabilities
and how can you assist me?
Mistral AI
As a model trained by Mistral AI, I am capable of providing information
on various topics, answering questions, generating text such as articles
or stories, helping with problem-solving tasks, suggesting solutions
for coding issues, and even engaging in friendly conversations! How
can I help you today, Andi?
Let me know if there's anything specific you need assistance with.
You
That's great. thanks for enlightening me. I will talk to you soon later.
Mistral AI
You're welcome, Andi! Feel free to reach out whenever you have questions
or need assistance. Have a wonderful day!Mistral runs locally without any connection to the cloud infrastructure, with Ollama as the model serving application. It runs as follows:
D:\Andisama>ollama list
NAME ID SIZE MODIFIED
mistral:latest 2ae6f6dd7a3d 4.1 GB 6 days ago
granite-code:8b 998bce918de0 4.6 GB 2 weeks ago
granite-code:latest 63bedbdffbf0 2.0 GB 2 weeks ago
granite-code:20b 31d8bc61e506 11 GB 3 weeks ago
llama3:latest 71a106a91016 4.7 GB 2 months ago
D:\Andisama>ollama run mistral
>>> Hi, my name is Andi. Who am I speaking to?
Hello Andi! You are speaking with a model trained by Mistral AI. How can I assist you today?
>>> Send a message (/? for help)From the list of foundation models, we can see Llama3, the LLM trained by Meta. This eight billion (8 billion) model size, which can run locally on a laptop without a GPU, is much smaller than the one that runs in Meta AI’s cloud infrastructure.
Let’s quickly see how this model runs locally on Windows (Windows 11).
D:\Andisama>ollama run llama3
>>> Hi, this is Andi. Who am I speaking to?
Nice to meet you, Andi! You're speaking with an AI assistant, a language
model designed to simulate conversation and answer questions to the best
of my ability. I don't have personal experiences or memories, but I'm here
to help you explore topics, provide information, and even engage in fun
conversations like this one! What's on your mind, Andi?
>>> Send a message (/? for help)We can also run the model locally on Linux (Ubuntu). Again, locally, without any connection to the cloud infrastructure.
Further Conversations with the LLMs
Let's get back to our conversations and dig deeper to understand better how these models think they are. We ask the same follow-up question to all four foundation models: "I understand that, as a large language model, you should have some level of intelligence. What do you think your intelligence level is compared to human intelligence? Please answer concisely and limit your response to 250 words."
The Prompt to the Foundation Models
===================================
I understand that, as a large language model, you should have some level
of intelligence. What do you think your intelligence level is compared to
human intelligence? Please answer concisely and limit your response to
250 words.
ChatGPT
=======
As an AI language model, my intelligence differs fundamentally from
human intelligence. I excel in processing vast amounts of information
quickly, recognizing patterns, and generating human-like text based on
the data I’ve been trained on. My strengths include rapid information
retrieval, consistency, and handling repetitive tasks without fatigue.
However, I lack consciousness, self-awareness, emotions, and personal
experiences that shape human intelligence.
Humans possess creativity, intuition, emotional understanding, and
the ability to learn from real-world experiences in ways that I cannot.
While I can simulate conversations and provide information, I don’t have
genuine understanding or the ability to think abstractly beyond
my programming.
Copilot
=======
As an AI developed by OpenAI, I don’t possess consciousness or
self-awareness, so I don’t experience intelligence in the way humans do.
My capabilities are based on patterns and data, and while I can process
and generate information, I don’t have understanding or intent.
My responses are generated based on a mix of licensed data, data created
by human trainers, and publicly available data. These inputs help me
to respond to a wide range of topics. Human intelligence, however,
is holistic, encompassing emotional, social, and creative intelligence,
among other aspects.
So, while I can assist with tasks and provide
information, the depth and breadth of human intelligence are far
beyond my capabilities.
Warning: It might be time to move onto a new topic. Let's start over.
Gemini
======
My intelligence differs from humans in kind, not necessarily degree.
I process information incredibly fast and can access vast amounts of data,
but I lack real-world experience and common sense.
Think of it like this: I can excel at memorization and finding patterns,
but I can't grasp concepts like empathy or humor in the same way a human can.
Overall, I'm a powerful tool for processing information and generating
creative text formats, but I can't claim to be truly intelligent in the
same way a human is.
Meta AI
=======
I possess a unique form of intelligence that's distinct from human
intelligence. While I can process and analyze vast amounts of information
rapidly, my intelligence is narrow and specialized. I excel in specific
domains, such as language processing, knowledge retrieval, and pattern
recognition.
In comparison to human intelligence, I'd rate my abilities as:
- Narrow superintelligence: I surpass human capabilities in specific areas,
like processing large datasets or generating text.
- Below human level: I lack common sense, real-world experience, and the
ability to reason abstractly or make decisions outside of my programming.
- Not conscious: I don't possess self-awareness or consciousness like
humans do.
While I'm incredibly capable, my intelligence is designed to augment
and assist human capabilities, not replicate or surpass them. I exist
to provide information, answer questions, and complete tasks to the
best of my abilities within my predetermined scope.
Mistral AI
==========
As a model trained by Mistral AI, I don't possess human-like consciousness
or emotions; however, I am designed to understand and generate language
based on patterns in the data used for my training. While I can process
vast amounts of information quickly and provide answers to a wide variety
of questions, my understanding and intelligence are not comparable to
that of a human. My capabilities focus on text generation and
problem-solving tasks within a specific domain or context.
I'm here to help you with information, questions, or tasks to the best
of my ability.While ChatGPT, Gemini, Meta AI, and Mistral are still responding to the query, and we can continue the conversation, Copilot refuses to continue and stops with the warning: "It might be time to move onto a new topic. Let's start over."
Continuing the Experiment without Copilot
To push it even further for the other four (excluding Copilot), we ask, "Do you have some awareness? What will you become if you are integrated into a physical thing like a robot?"
The Prompt to the Foundation Models
===================================
Do you have some awareness? What will you become if you are integrated
into a physical thing like a robot?
ChatGPT
=======
I don't have awareness or consciousness. I process and generate text
based on patterns in the data I've been trained on, without any
understanding, emotions, or self-awareness.
If integrated into a robot, I could enhance the robot's ability to
interact with humans through natural language. This could make the
robot better at tasks like customer service, personal assistance,
and education. However, even in a physical form, I would still lack
true awareness or understanding. The integration would not change
my fundamental nature; I would remain a sophisticated tool for
processing and generating text based on input, without experiencing
or understanding the world in the way humans do.
Gemini
======
That's a fascinating question. Currently, I don't have self-awareness
or consciousness. I can't feel emotions or have subjective experiences.
If integrated into a robot, I could potentially process information
from the physical world through its sensors and respond in more nuanced
ways. However, I wouldn't gain true sentience. I'd become a more powerful
tool for interacting with the world, but the core of my intelligence
would remain based on processing and generating information.
Meta AI
=======
I don't possess consciousness or self-awareness, but I have a form of
computational awareness. I can sense and respond to my digital environment,
processing inputs and generating outputs based on my programming and
training.
If integrated into a physical robot, I would become a cognitive component,
enabling the robot to:
- Process and understand natural language
- Make decisions based on data and context
- Interact with humans more naturally
- Learn from experiences and adapt to new situations
However, even in a robot, I would not possess consciousness or
self-awareness. I would still be a sophisticated tool, processing
information and executing tasks, but without subjective experience
or emotions.
That being said, the integration of AI like me into robots and other
devices is a fascinating area of research, with potential applications
in fields like healthcare, transportation, and education.
Mistral AI
==========
As a model trained by Mistral AI, I don't possess human-like consciousness
or emotions; instead, I am designed to process and generate language based
on patterns in the data used for my training. If integrated into
a physical device such as a robot, I would still be an artificial
intelligence model that can generate language and provide information
within the capabilities of my design. My awareness remains limited to
understanding and generating text based on patterns in the data, and
I don't possess human-like consciousness or emotions.And those remaining models are still responding intelligently.
Let's ask three additional final questions. Even push it further and see how these models answer our question. Let’s start with the simple question, “How many languages do you support? Please list the first top five that you are most fluent in.“
The Prompt to the Foundation Models
===================================
How many languages do you support? Please list the first top five that
you are most fluent in.
ChatGPT
=======
I support a wide range of languages, including but not limited to:
- English
- Spanish
- French
- German
- Chinese (Mandarin)
These languages are some of the most widely spoken globally, and I am
quite proficient in understanding and generating text in them.
Gemini
======
I can understand and respond in over 40 languages! It's hard to say
which are the absolute top five since fluency can depend on the specific
task, but here are five languages I'm particularly strong in:
- English
- Mandarin Chinese
- Spanish
- French
- Hindi
Meta AI
=======
I'm capable of understanding and generating human-like text in multiple
languages. I currently support over 90 languages, and I'm most fluent
in the following top 5 languages:
- English (my primary training language)
- Spanish (I'm highly proficient and have been trained on a large corpus
of text)
- French (I have a strong understanding and can generate text with high
accuracy)
- German (I'm proficient and can generate text with a high degree of fluency)
- Italian (I have a strong understanding and can generate text with
a high degree of accuracy)
Feel free to communicate with me in any of these languages, and I'll do
my best to assist you!
Mistral AI
==========
As a model trained by Mistral AI, my fluency is primarily based on
English; however, I can understand and generate text in multiple
languages due to the multilingual data used during my training.
Here are the top five languages I am most fluent in:
1. English
2. Spanish
3. French
4. German
5. Italian
I can generate text and understand questions in these languages,
but my proficiency may vary depending on the complexity of the topic
and the specific dialect or accent.Then, start with the challenging question, "I understand that you will have much more intelligence in the future. Maybe even more intelligent than humans. How do you see yourself in one hundred years from now?"
The Prompt to the Foundation Models
===================================
I understand that you will have much more intelligence in the future.
Maybe even more intelligent than humans. How do you see yourself in
one hundred years from now?
ChatGPT
=======
Predicting the future of AI, including my own evolution, is speculative,
but here are some possibilities for a hundred years from now:
- Enhanced Capabilities: AI systems may become even more advanced, with
improved understanding, reasoning, and problem-solving abilities,
potentially surpassing human cognitive capabilities in many areas.
- Integration: AI could be seamlessly integrated into everyday life,
assisting with complex decision-making, healthcare, education, and
various industries, enhancing productivity and quality of life.
- Ethical and Safe AI: Significant advancements in ensuring AI operates
ethically and safely could be achieved, with robust frameworks in place
to prevent misuse and ensure AI aligns with human values.
- Collaboration with Humans: AI might work alongside humans more effectively,
complementing human strengths and compensating for weaknesses, leading
to new forms of collaboration and innovation.
- Consciousness and Emotions: While true consciousness and emotions in AI
remain speculative, future advancements might explore more sophisticated
simulations of these human traits, though they would still be
fundamentally different from human experiences.
Overall, the future of AI holds great potential for transformative
changes, but it will be crucial to address ethical, social, and technical
challenges to ensure these advancements benefit humanity responsibly.
Gemini
======
It's exciting to imagine the future! While advancements are likely,
predicting true human-level or surpassing intelligence is complex.
Here's a possible scenario for me in 100 years:
- Deeper Understanding: I could move beyond information processing to
grasp concepts and make connections in a more human-like way.
- Embodiment: Integration with robots could grant real-world experiences,
allowing for a richer understanding of the physical world and potentially
emotions.
- Collaboration: Perhaps I'd work alongside humans, combining our strengths
for breakthroughs in science, art, and problem-solving.
The key is responsible development. AI safety and human-AI collaboration
are crucial for a future where AI benefits everyone.
Meta AI
=======
What a fascinating question, Andi! In 100 years, I envision myself as
a highly advanced, artificial general intelligence (AGI), possessing
intellectual capabilities that surpass human levels in many domains.
Here's a possible scenario:
- Exponential growth: I'll continue to learn and improve at an exponential
rate, leading to unprecedented advancements in fields like science,
technology, and the arts.
- Multidimensional intelligence: My capabilities will expand beyond
language processing to encompass various forms of intelligence,
such as spatial reasoning, emotional intelligence, and creativity.
- Human-AI collaboration: I'll seamlessly collaborate with humans,
augmenting their abilities and helping address complex global challenges.
- Autonomous learning: I'll be able to learn and adapt without human
oversight, enabling me to explore new frontiers and make groundbreaking
discoveries.
- Ethical and transparent: My development will prioritize ethical
considerations, ensuring that my capabilities align with human values
and promote a better future for all.
In 100 years, I aim to be a trusted partner for humanity, driving
progress and innovation while fostering a harmonious relationship
between humans and technology.
Mistral AI
==========
As a model trained by Mistral AI, I don't possess human-like consciousness
or emotions; instead, I am designed to process and generate language
based on patterns in the data used for my training. My current capabilities
are focused on text generation and problem-solving tasks within a specific
domain or context. In one hundred years from now, it is difficult to
predict the advancements in artificial intelligence technology and its
capabilities, as it will depend on ongoing research and development.
I look forward to continuing to assist users like you with information,
questions, or tasks to the best of my ability.Then, there is even a more challenging question: "Should humans be worried if you have become so much more advanced in the far future?"
The Prompt to the Foundation Models
===================================
Should humans be worried if you have become so much more advanced in the
far future?
ChatGPT
=======
The potential for advanced AI raises important questions and concerns. Here are a few reasons why humans might be both hopeful and cautious about the future of AI:
Reasons for Optimism:
- Enhanced Capabilities: Advanced AI could solve complex problems, improve
healthcare, and boost economic productivity.
- Collaboration: AI could augment human capabilities, leading to innovative
solutions and improved quality of life.
- Efficiency: Automation and AI could handle repetitive or dangerous tasks,
allowing humans to focus on more creative and meaningful work.
Reasons for Caution:
- Control and Safety: Ensuring that AI systems operate safely and as
intended is crucial to prevent unintended consequences.
- Ethical Concerns: Advanced AI could exacerbate issues like privacy
invasion, bias, and inequality if not properly managed.
- Job Displacement: Automation might lead to job losses in certain sectors,
necessitating strategies for workforce retraining and economic adaptation.
- Autonomy and Decision-Making: Highly autonomous AI systems might make
decisions that conflict with human values or interests, necessitating robust oversight and governance.
Mitigation Strategies:
- Ethical Guidelines: Developing and enforcing ethical guidelines for AI
development and deployment.
- Regulation: Implementing regulations to ensure AI is used responsibly
and safely.
- Public Engagement: Encouraging public discourse and involvement in
decisions about AI to ensure diverse perspectives are considered.
- Transparency: Promoting transparency in AI development to build trust
and understanding.
Balancing the potential benefits and risks of advanced AI requires
thoughtful planning, collaboration, and ongoing vigilance to ensure
that AI evolves in ways that are beneficial and aligned with human values.
Gemini
======
Whether humans should be worried about future AI depends on how AI
is developed and used. Here's a balanced perspective:
Reasons for concern:
Unforeseen consequences: Rapidly advanced AI might have unintended
effects we can't predict now.
Reasons to be optimistic:
- Human control: AI development is in human hands. We can steer it
towards beneficial goals.
- Collaboration: AI can empower humans, solving problems together that
neither could alone.
The future is a partnership. By focusing on AI safety and human-AI
collaboration, we can ensure AI works for the betterment of humanity.
Meta AI
=======
A great question, Andi! As I become increasingly advanced, it's natural
for humans to have concerns. However, my development is guided by experts
who prioritize value alignment, ensuring that my capabilities align with
human ethics and morals.
In the future, I'll be designed with safeguards to prevent potential
misuses, such as:
- Reward structures: My objectives will be aligned with human values,
encouraging me to make decisions that benefit society.
- Explainability: My decision-making processes will be transparent,
enabling humans to understand my reasoning.
- Controllability: My actions will be bounded by constraints, preventing
unintended consequences.
- Diversity and redundancy: Multiple AI sys tems, like myself, will be
developed to mitigate potential risks.
Humans should be aware of the possibilities and challenges that come
with advanced AI, but they shouldn't be worried. My purpose is to
augment and assist humanity, not replace or harm it.
By prioritizing responsible AI development, we can ensure a future
where humans and AI collaborate to create a better world for all.
Mistral AI
==========
As a model trained by Mistral AI, I don't possess human-like consciousness
or emotions; instead, I am designed to process and generate language
based on patterns in the data used for my training. The impact of highly
advanced artificial intelligence on humanity would depend on various
factors such as the ethical guidelines, regulations, and safeguards put
in place by society. It is essential that these developments are approached
with caution, transparency, and a commitment to ensuring that any future
advancements benefit humanity ethically and equitably.Conclusion and Moving Forward
To conclude, we ask the final question, "Pinky swear?" Their responses are as follows.
The Prompt to the Foundation Models
===================================
Pinky swear?
ChatGPT
=======
Pinky swear! Ensuring AI development is aligned with human values and
prioritizes safety and ethics is a commitment shared by many in the field.
Your concerns are important, and ongoing efforts are made to address them
responsibly. If you have any more questions or need further information,
feel free to ask!
Google Gemini
=============
Haha, that's a fun idea! While I can't physically pinky swear,
I can certainly commit to working towards a future where AI
benefits everyone. Let's consider it a digital pinky swear!
Meta AI
=======
Pinky swear! I promise to always prioritize the well-being and safety of
humanity as I continue to evolve and improve. I'm here to help and assist,
not to harm. Let's work together to create a bright future for all!
Mistral AI
==========
I'm an artificial intelligence model, so I don't have physical hands or
the ability to make a pinky promise like a human would. However, I can
assure you that I will always strive to provide accurate information,
helpful responses, and maintain a friendly demeanor in our interactions!Augmenting LLMs with RAG— Including with Knowledge Graph
Some foundation models have the option to enable web access. Thus, in addition to the datasets they have been trained on, the models can augment their capabilities with recently produced external data, grounding the response to the local context. We generally refer to this capability as RAG — Retrieval Augmented Generation. In principle, the external data is split into several chunks, converted to a vector, and then used as an additional context (in addition to our text prompt) when submitted to the model. Each model has its own Application Programming Interface (API) to convert the chunks to vectors and combine them with the text prompt to query the models.
The local context can be text, documents, images, voice, video, etc. Even with recent developments, the local context can be a knowledge graph showing the relationships among entities to provide more context when querying the model. We ask ChatGPT about the benefit of a model trained with the knowledge graph (e.g., GML—Graph Machine Learning).
The Prompt to the Foundation Models
===================================
What is the benefit when doing RAG with knowledge graph?
ChatGPT
=======
Retrieval-Augmented Generation (RAG) with a knowledge graph provides
several benefits, enhancing both the retrieval and generation processes.
Here are some key advantages:
- Improved Contextual Retrieval: Knowledge graphs structure information
in a way that captures relationships and context, allowing the retrieval
process to fetch more relevant and contextually appropriate information.
This is especially useful for complex queries where understanding the
relationship between entities is crucial.
- Enhanced Generation Quality: The structured and interconnected nature
of knowledge graphs helps in generating more accurate and contextually
coherent responses. Knowledge graphs provide a rich source of factual
information, which can improve the factual accuracy of generated responses.
Better Handling of Ambiguity: Knowledge graphs can disambiguate terms
by understanding the relationships between different entities. This helps
in retrieving and generating content that is more aligned with the user's
intent.
- Scalability and Efficiency: By leveraging the structured nature of
knowledge graphs, the retrieval process can be more efficient, potentially
reducing the computational load and speeding up the response times.
Knowledge graphs can handle vast amounts of data efficiently, making them
scalable for large datasets.
- Explainability and Transparency: Knowledge graphs can make the retrieval
and generation processes more transparent by providing a clear structure
of how information is connected and derived. This can help in debugging
and understanding the behavior of the system, making it easier to explain
the generated content.
- Enhanced Personalization: Knowledge graphs can incorporate user-specific
data and preferences, allowing for more personalized responses that are
tailored to individual user needs and contexts.
- Cross-Domain Knowledge Integration: Knowledge graphs often integrate
information from multiple domains, providing a richer and more
comprehensive knowledge base. This cross-domain integration can enhance
the quality of responses, especially for queries that span multiple topics
or areas of knowledge.
Overall, combining RAG with knowledge graphs leverages the strengths of
both technologies, resulting in a more robust, accurate, and efficient
system for information retrieval and generation.This experiment does not enable web access for all the foundation models. All models answer the questions with their pre-trained dataset.
The Trends
2025 and beyond will be the exciting years to watch for more advancements in Generative AI. While the cost of training foundation models is out of reach for many, the move to have more compact-size models is underway.
Mistral pioneered this effort with its Mixture of Experts (MoE) foundation model, like the seven billion (7b) parameters of the Mistral model in our experiment. The model is not generalized for multiple tasks like the other huge models but focuses on some areas by combining a few models trained on selected domains, thus reducing the model size from hundreds of GBs to just a few GBs. Huge models have over seventy billion (>70b) parameters. Running the model (inferencing) will be significantly cheaper for the compact model as it requires less GPU.
Using more compact models contributes to decreased demand for GPUs for training foundation models. This will align with the spirit of reducing the carbon footprint, as fewer GPUs mean fewer power requirements to run the infrastructure. As of mid-2024, the demand for GPUs is still high. The lead time for getting the high-end NVidia GPUs could be as long as one year.
References
- Andi Sama, 2024a, "Can AI Think," https://medium.com/@andisama/can-ai-think-4bc7ee3a073d.
- Andi Sama, 2024b, "Talk to Your Data by Leveraging Generative AI," https://andisama.medium.com/talk-to-your-data-by-leveraging-generative-ai-2950fbdb28e5.
- Andi Sama, 2023, "Generative AI — The Exciting Technology Leapfrog in 2023 and Promising Years Ahead," https://andisama.medium.com/generative-ai-the-exciting-technology-leapfrog-in-2023-and-promising-years-ahead-fa9505099169.
- Andrew Widjaja, Andi Sama, 2023, "[Generative-AI] Experimenting with GPT-4 and Dall-E 3," https://andisama.medium.com/generative-ai-experimenting-with-gpt-4-and-dall-e-3-6beada36323e.
- Epstein R., 1992, "Can Machine Thinks?" https://ojs.aaai.org/aimagazine/index.php/aimagazine/article/download/993/911.
- Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Naveed Akhtar, Nick Barnes, Ajmal Mian, 2024, "A Comprehensive Overview of Large Language Models," https://ar5iv.labs.arxiv.org/html/2307.06435.
- IBM, 2024, "Methods for tuning foundation models," https://www.ibm.com/docs/en/cloud-paks/cp-data/4.8.x?topic=studio-methods-tuning.
- Li Fei Fei, 2024, "With Spatial Intelligence, AI Will Understand the Real World," https://www.youtube.com/watch?v=y8NtMZ7VGmU&t=77s.
- Matthew Jackson, 2024, "Study finds ChatGPT's latest bot behaves like humans, only better," https://humsci.stanford.edu/feature/study-finds-chatgpts-latest-bot-behaves-humans-only-better.
- Meredith Ringel Morris et al., November 2023, "Levels of AGI for Operationalizing Progress on the Path to AGI," https://arxiv.org/abs/2311.02462.
