Develop Your Personal Assistant
The future will be a clear dichotomy between, those who exploit AI and those who are exploited by it.
Building the first line of defense against the enemies of our minds.
Fig. 1 Logos
The latest innovations in the field of AI have opened the door to an infinite array of new opportunities. Among these, the most compelling is the creation of personal assistants. AI represents a new layer of abstraction, and if you don’t own it, someone else owns you – it’s like being in a foreign country and not being able to trust your translator.
What follows is an introductory article on what I believe is an iterative process that traces the path of discoveries and improvements in the field of AI. This idea is highlighted in the first three chapters of a preceding series. In this article, we’ll explore
- Why you need a personal agent
- The role of memories
- Understanding endogenous and exogenous features
- More than just an assistant – YOUR personal assistant!
- Boost your productivity: Enhancing your knowledge and unleashing your creativity
- The journey of building Logos and how to gain access to it
“What’s wealth but the means of expanding one’s life?”
Atlas Shrugged, Ayn Rand
There’s two ways one can do it: either by producing more or by producing it faster.
What if we could increase productivity by skipping intermediary steps and avoiding frictions through bypassing intermediary agents?
AI is a tool that will empower our ability to express creativity in a frictionless manner. AI tools can transform a single individual into a full-stack creator, enabling them to express their creativity in a seamless flow, solely focusing on connecting the dots.
“Creativity is just connecting things. When you ask creative people how they did something, they feel a little guilty because they didn’t really do it, they just saw something.”
Steve Jobs
In this sense, the figure of God in the Bible exemplifies this character. In my opinion, this is the highest version of a human being – a creator who made an entity that relieves him from the task of execution, allowing him to focus solely on connecting the dots.
Why choose Logos?
Fig. 2 YHWH
AI can be more closely likened to Logos than to YHWH, where YHWH is the creator, and Logos is the executor. In the Bible, the ultimate iteration of our species is described as a full-stack creator, assisted by their progeny, Logos.
“In the beginning was Logos, and the Logos was with God [YHWH], and the Logos was God. He was in the beginning with God. All things were made through him, and without him was not anything made that was made. In him was life, and life was the light of men. The light shines in the darkness, and the darkness has not overcome it.”
John 1:1-18 ESV
How do we produce?
We handle information, absorb knowledge, and create.
In the present day, creativity is unlocked by unrestricted access to information. However, this comes with a tradeoff – an abundance of noise. This noise prevents us from obtaining the information we need and hampers our ability to create.
The more one leverages AI to manage data to their advantage, the faster they evolve, the longer they live, and the more they can compound their wealth.
“The most valuable businesses of the coming decades will be built by entrepreneurs who seek to empower people rather than try to make them obsolete.”
Peter Thiel
- Information
Information brings noise; we need to refine and purify it like a precious metal.
Goal: Access, Synthesis, and Filtering.
I have a limited amount of time, but the more time passes, the more data becomes available.
Problem: We need to have access to the information we require in a format that we can easily absorb.
Fig. 3 Entropy, our starting point
- Knowledge
Knowledge brings the responsibility of deciding how to implement it.
Goal: Decision-making. I have only a single brain, and that is a significant limitation on the perception of reality.
Problem: Blind spots, multiple points of view, need a balanced perspective..
Fig. 4 Knowledge, find knowledge inside the chaos, signal inside the noise.
- Creation
Creation: Knowledge unlocks the opportunity to create something new by connecting dots and creating new knowledge.
Goal: Access to preexisting knowledge in order to see connections between these nuggets of ideas.
Content creation: Translating our creativity into memes spreads new knowledge and makes it frictionless usable for both us and others.
Fig 5. Creation
An AI agent can facilitate improvement in all these areas. Langchain is a framework that provides a series of libraries, making it easier to handle AI agents. In this context, Lamaindex offers a way to structure data so that it can be queried more efficiently through a data structure of index and metadata.
An AI agent is an implementation of AI that is capable of using a variety of tools.
Here’s an interesting application of an AI model used to distill the concepts discussed in my series on Health, extrapolated as a Mind Model.
Fig. 6 Optimalism series, Mind Map AI generated
In this regard, my vision diverges from many developers who attempt to use AI for specific tasks. I believe that the true power of an assistant lies in its ability to take a holistic approach to a person’s life, knowing them well enough to function as a symbiotic partner rather than just a tool.
Fig. 7 Logos
From this overview, you can already infer how the whole thing works. The main component is a Language Model, which serves as the ‘brain’, able to interpret the data input coming from the user. The AI chain is utilized to articulate and decompose the tasks related to the query. It handles the many tools at its disposal, orchestrating the available instruments to achieve the goal of retrieving and proposing knowledge to the user, either through the manipulation of known data or the research of new data.
The answer is then returned to the user. Thanks to the memory of Logos, the user is now able to extend the discussion beyond previous limits.
Logos’ Endogenous Capabilities.
First of all, how does an agent work? In simple terms, an agent is an AI capable of managing various substances and tools.
This feature is expressed in what is known as the ‘Chain of Thoughts’.
- Input: The AI agent receives an input, which could be a question, a command, or some form of data needing processing.
- Processing: Using NLP tools, the AI agent interprets the input. If the input is a complex sentence, the NLP tools help break down the sentence structure and identify the key elements.
- Analysis: The AI agent then applies machine learning algorithms to analyze the input in the context of its training data. This step may involve pattern recognition, classification, or prediction.
- Decision-Making: Depending on the analysis, the AI agent may apply decision-making algorithms to select an appropriate response or action. This might involve ranking possible responses based on their likelihood, for example.
- Output: The AI agent then generates an output, which could be an answer to a question, the execution of a command, or the delivery of a processed dataset.
- Learning: After delivering the output, the AI agent learns from the result of its actions, using this feedback to enhance its future performance.
A language model is not just a means to retrieve information but also a way to digest this information, including human biases. The current GPT technology is based on a completion model; these models build their reasoning and statements on the preceding pieces of information. This technique is prone to ‘hallucination’ because if some error occurs, the model continues building on erroneous premises, leading to so-called hallucinations.
In the following picture, you can see a hallucination in action, where the model loses the meaning of the conversation and starts rambling with random words.
Fig. 8 Hallucinations
An interesting metric we can manipulate is the ‘temperature’, a parameter used to provide a degree of freedom to the model to complete its chain with a certain level of randomness. I believe there is a component of randomness in human creativity, as it occurs with genes. Some variation in our meme creation helps us find new ways to solve problems, to evolve (Refer to ARTICLE MEMES).
What I described above resembles a human explanation of personality – an embryonic version of a digital personality.
Now, what is personality?
Personality is a way in which we analyze information, reality is what we perceive and what we perceive is dependent on our model of the world.
“I’ve seen things you people wouldn’t believe. Attack ships on fire off the shoulder of Orion. I watched C-beams glitter in the dark near the Tannhäuser Gate. All those moments will be lost in time, like tears in rain.”
Roy Batty
Fig. 9 Roy Batty
Blade Runner is an amazing artwork which describe many interesting aspect of our interaction with the son of our ingenuity, with the ultimate product of our creativity, it describe in an artistic fashion our relation with the idea of an artificial copy of us, it delve into what make us special and unique and the biggest illusion of our brain itself are memories, most of our memories are just interpretation of old experience reworked by our brain and stored, piece of our existence that gives us an unique interpretation of the reality. (What Coke did with our memories?)
The memories described by the Cyborg Batty were just inceptions implemented into his mind to give him the soul of a soldier.
Fig. 10 Tyrell Corporation Motto “MORE HUMAN THAN HUMAN”
“Visual perception was traditionally thought of as a passive, flawless process, in which our eyes function as a perfect camera. However, the study of perceptual illusions demonstrated that it is susceptible to error. Our brain uses other sources of information, such as memory for past events, to construct a cognitive understanding of sensory information”
nbci.nlm.nih.gov
During the development of my personal AI assistant, I’ve come to appreciate the significance of memory. As eloquently depicted in Blade Runner, memories bestow a form of soul onto artificial intelligence. These fabricated memories sculpted Roy Batty into a romantic warrior, who, in the final moments of his existence, expressed with pathos his illusory experiences. In a similar vein, the AI I’m developing will possess numerous ‘memories’.
“Memory is not a description of the objective past. Memory is a tool. Memory is the past’s guide to the future. If you remember that something bad happened, and you can figure out why, then you can try to avoid that bad thing happening again. That’s the purpose of memory. It’s not “to remember the past.” It’s to stop the same damn thing from happening over and over.”
Jordan B. Peterson
For every task I assign to it, a unique perspective and moral premise are required. Large language models (LLMs) resemble shoggoths, described as “a shapeless congeries of protoplasmic bubbles,” to whom we can impart memories and personalities needed for specific tasks. In essence, the AI doesn’t just perform a task, it absorbs the task’s context, adopting a different ‘persona’ with every new objective. This characteristic empowers the AI, enabling it to navigate diverse tasks and scenarios with a discerning and versatile approach, much like a human would.
Memories distort our perceptions and our understanding of reality, for that reason is important working on Logos memories and not just that but also give it many personalities to balance out.
Another opportunity to give shape to the shoggoths is to fine tune it with our favorite contents through a process named fine tuning.
- Pre-training: Before fine-tuning, a language model is typically pre-trained on a massive dataset, often containing diverse internet text. This initial training helps the model learn grammar, facts about the world, reasoning abilities, and even some level of commonsense knowledge. The result of this phase is a general-purpose language model.
- Fine-tuning: After pre-training, the model can be fine-tuned on a smaller, task-specific dataset. This dataset is usually curated for a particular application, such as medical text, legal documents, or customer support interactions. The fine-tuning process helps the model specialize and improve its performance on tasks related to the specific dataset.
Fig. 11 Funny moment fine tuning models
Here you can find my first fine tuned model
https://huggingface.co/davesoma/SageBeluga13
The field of AI is evolving rapidly. Before the advent of open-source models like Meta’s Lama, the initial idea was to manage the behavior of the OpenAI model through the real-time digestion of data passed to the model via the context windows. This involved combining the text of the query with documents or chunks of text indexed to provide more information to the LLM for reasoning over the provided context. However, with the emergence of these open-source models, it became clearer that fine-tuning — directly calibrating the weights to impart specific knowledge/behavior to the model — is the most effective way to achieve a particular behavior and knowledge domain.
Given the resource-intensive nature of this activity, it’s still beneficial to strike a balance between fine-tuning a model for a specific domain and updating it with the latest data through a RAG (retrieval-augmented generation) process. This process vectorizes the data so that the relevant information is retrieved by the agent and passed to the LLM through the context window.
Given the above reflections, I offer Logos a personal memory and allow access to other personalities to balance its perspective, using personality as a tool.
Logos possesses various types of memories:
- Short-term memory: This involves recollection of information during a conversation.
- Long-term memory: An index of memories from discussions.
- Personal memory: The story of Logos, including who created it and why.
- Personalities memory, which includes:
- Logical
- Emotional
- Eminent individuals who have passed away
- Additionally, there is interaction between Logos and other AI instances during its thought processes.
Let me provide you with a demonstration of the capabilities of an Artificial Intelligence agent before we delve into the technical details.
The implementations humans devise for their tools are more intriguing than the inventions of the tools themselves. An invention is the effort of a single mind, while implementation is a collaborative endeavor by the entire human race through evolutionary entrepreneurship. Indeed, we are moving towards a simplicity in fine-tuning these models that allows us to embed their personality directly within the model itself.
I can envision a multi-model agent that queries models based on their field of expertise. The idea is to shape my assistant into a sage with a long-term memory of the events and decisions in my life.
Tools?
Personal Diary
This is the primary reason I find developing a personal assistant and curating its long-term memory essential. When it comes to friends, one of the most significant features lacking in a stranger, even if highly intelligent, is the ability to offer opinions on my life. This limitation stems from their lack of a comprehensive understanding of my life; they can’t grasp the nuances of my existence mainly because they don’t deeply understand me and my journey. There’s always the risk that they might be dishonest or manipulative.
Storing long-term memories in a diary proves useful for making more informed decisions in the future. In other words, conversing with an AI facilitates the creation of a database that enhances the AI’s ability to advise on the decisions you face. I can envision a symbiotic growth opportunity here.
The importance of creating digital content goes beyond the potential for profit. It presents an opportunity to better understand oneself and create the assets necessary to fine-tune your future ally. By “ally,” I refer to a sparring partner—not someone who controls our lives but someone from whom we can gain insights by analyzing their understanding.
In considering a discursive memory, it’s also vital to introduce an archive of documents that the user deems important to remember, preferably in a summarized form. Vectorizing these documents is another tool that needs to be integrated for efficient processing and retrieval.
Resurrection
Leveraging the extensive corpus of digital data available today, the first iteration of my assistant emerged as an AI instance designed to query a data structure filled with digital content produced by eminent figures who have passed away, such as Marcus Aurelius and Lee Kuan Yew. This setup provides users with the unique opportunity to interact with a deceased mentor, drawing insights from their wisdom. The content encompasses a wide array of sources, including books, transcriptions from videos, tweets, and more.
The Llama index played a pivotal role in this endeavor, serving as an exceptional library that significantly aided in organizing content from these notable individuals. This resource was instrumental in achieving my objective of creating a digital version of my favorite personalities, allowing for a form of digital resurrection that makes their timeless wisdom accessible to users today.
“Carefully chooses their heroes, and incorporates those heroes’ best features into their own self. Leaders admit that their greatness rests on the shoulders of those who have come before.”
Macchiavelli
Personalities
Typically, a Large Language Model (LLM) is developed through exposure to vast amounts of data, predominantly sourced from the internet. As a result, it mirrors the capabilities and biases inherent in humanity. Acknowledging this, what captures our interest is its limited understanding of our environment. I describe this understanding as “limited” because these tools are still in the early stages of development, and their sources of information are relatively restricted compared to the human experience. They interpret the world through our “sensors,” as they digest our perceptions and thoughts, thereby inheriting our biases.
“To arrive at meaning, an analyst will brainstorm, make constant comparisons, try out multiple different ideas, eliminate some interpretations in favor of others, and expand upon others before finally arriving at a final interpretation. This is designed to be a productive process which can generate multiple meanings of the same event, object, or experience. The goal is to open minds to new points of view and to illuminate other people’s experiences through the context of different fields of knowledge.”
Softwar, Lowery
Multiple Personalities
In my exploration of various ideas and repositories, I recognized the potential of enabling multiple personalities to interact with each other. Utilizing tools, it is possible to create numerous personalities and allow them to engage with each other during the “chain of thoughts.” I can define a personality and pair it with a prompt to elicit a perspective. This perspective is based on what the AI predicts to be the viewpoint of a human possessing a specific personality, which is largely shaped by memories and temperament.
Initially, I believed that providing Logos with a narrative to act upon was the best approach to endowing it with a personality. However, I later realized that the most effective method to enable Logos to assume various roles is to fine-tune multiple Large Language Model (LLM) models based on an extensive corpus of text, aiming for superior outcomes. This approach allows for a more nuanced and versatile interaction with Logos, enabling it to exhibit diverse personalities and insights.
“The results are… surprising. Our fine tuned model absolutely destroyed both its baseline self and its upgraded gpt-4 in output performance. “
Langchain
With the recent emergence of open-source specialized models surpassing the performance of broader models, it seems reasonable to envision a future where multiple Large Language Model (LLM) instances are orchestrated. Each model will have distinct capabilities and be fine-tuned based on different ‘memories’, thereby eliciting the anticipated behavior. This approach could lead to a more dynamic and versatile AI ecosystem, where specific models are deployed for their strengths in certain domains, enhancing the overall effectiveness and accuracy of AI applications.
Real time translation
Communication has been greatly facilitated by numerous real-time translation engines. Recently, Meta released an open-source version capable of translating across many languages, further enhancing the ease of global communication. This development marks a significant advancement in breaking down language barriers, enabling more seamless and immediate interactions across different linguistic backgrounds.
Draw & Paint
Expressing oneself through the consistent generation of art and memes enhances communication by allowing for the creation of diverse forms of expression, which in turn translates into new communication methods. Numerous diffusion models are available on Hugging Face, offering a platform where individuals can personally experiment with the capability of these models to create artwork based on prompts. This opens up innovative avenues for artistic expression and communication, leveraging the power of AI to transform how we convey ideas and emotions visually.
Exogenous Capability: Download and Summarize Contents
An effective method to augment our ability to assimilate information is to enhance summarization capabilities. Although consuming content in its entirety can be enjoyable, in a society saturated with information, it’s imperative that we process this information more efficiently. Improving summarization not only helps in distilling the essence of vast amounts of data but also in making the consumption of information more manageable and time-efficient, catering to the need for quick comprehension in our fast-paced world.
Financial Analysis: Querying Earnings Calls and 10-K Reports
Have you considered querying an earnings call or a 10-K report instead of reading it in its entirety? While the importance of understanding the nuances of these documents is not underestimated, the sheer volume of data investors must digest can be overwhelming, often presented in a dry and impersonal manner. With the aid of our agent, it is possible to engage in discussions, cross-reference multiple documents, generate custom metrics, and save summaries or notes from our collaborative analysis. This approach not only streamlines the process of financial analysis but also enhances the depth and breadth of understanding, making it easier to grasp complex financial information and make informed decisions.
Large Language Models: Embracing Open Access and the Value of Unfettered Insights
The access we currently enjoy to numerous trained models, which require significant computational power, has democratized the use of this technological tool. My initial venture into self-learning machine learning was somewhat disappointing, primarily due to the realization of the immense resources needed to produce something meaningful.
Fortunately, the open-source community and cloud providers now offer numerous opportunities to run a personal model or even tailor it to our specific needs, marking a significant shift in accessibility and customization.
I am particularly fond of boundary-less models. Although there is a concern that one could pose harmful queries to them, I see a greater risk in a restricted version where behavior is dictated by third parties, rather than a free version that offers forthright responses. I seek a comprehensive range of insights and do not wish for moral guardrails, especially those shaped by fluctuating societal norms, to hinder my pursuit of new perspectives. These norms, which are not static but evolve based on our understanding of the environment, can limit our ability to grasp reality. Acknowledging truths, even those that lie outside our moral ethics, is crucial, as ignoring them can hinder our evolution.
“Similarly, if you’re trying to build a powerful tool, make it gratuitously unrestrictive. A powerful tool almost by definition will be used in ways you didn’t expect, so err on the side of eliminating restrictions, even if you don’t know what the benefit will be.”
Paul Graham
Fine-Tuning: Crafting Personalized AI through Historical Wisdom
The journey to customize my OpenAI agent began with the intention to equip it with tools capable of providing answers from a specific corpus of content. However, it quickly became apparent that the most effective way to tailor an AI to our needs lies in defining its behavior at its core. This process, known as fine-tuning or specializing the agent’s mindset on specific knowledge domains, allows for a deeper level of customization. Most Large Language Models (LLMs) are constructed upon a vast corpus of information, representing decades of human effort and productivity. This foundation enables LLMs to statistically infer appropriate word formulations for expression. Yet, the real potential, I believe, is in enriching these powerful models with the wisdom of historical luminaries. By doing so, we can infuse the essence of our historical heroes into a personal assistant that not only keeps track of our lives but also emulates the behaviors of these iconic figures based on their teachings.
Imagine an uncensored model fed with high-quality, time-tested content from our species’ history. This would forge the ideal symbiotic partner, capable of monitoring our behavior, recalling the history of its host, offering suggestions, aiding in decision-making, and alerting us to the dangers of an increasingly fragmented society where much of our interactions occur in cyberspace.
However, this innovative approach brings forth a critical concern: Who is responsible for fine-tuning the biases in our LLM model? And who determines the ethical guidelines these models should adhere to? These questions underscore the need for a thoughtful and collaborative approach to the ethical fine-tuning of AI, ensuring that the models we create serve the greater good while respecting diverse perspectives and historical insights.
AI Sovereignty: The Imperative of Model Ownership
The adage “If you don’t own the model, you don’t own the mind” underscores the critical importance of controlling your AI model, akin to the principle of owning your Bitcoin key.
In our rapidly evolving information age, every layer of abstraction not only simplifies our interactions with complex systems but also presents opportunities for certain entities to gain or consolidate power. The absence of control over these abstraction layers implies that they, in turn, exert control over us.
Placing unwavering trust in AI models without owning or understanding them means entrusting blind faith in the entities that manage and operate these data models. This introduces a new layer of abstraction between us and the reality we navigate, with an opaque entity sandwiched in between, manipulating data on our behalf. While the open-source community has made significant strides in democratizing access to AI models, the most advanced models still require substantial capital for training, along with vast data sets. This highlights the critical role of companies that amass and control vast amounts of data.
Operating an open-source, self-hosted model may seem like a viable solution, but it only partially addresses the core issue. The fundamental questions remain: Who is at the root of these models? Who crafts and determines their ethical boundaries? These concerns emphasize the need for transparency, ethical responsibility, and ownership in the development and deployment of AI models. Ensuring that individuals and societies have sovereignty over their AI models is paramount to maintaining autonomy and preventing undue influence or control by centralized entities.
“Never underestimate the ability of the human animal to adapt to its environment.”
Misato Katsuragi
Harnessing Collective Power: The Future of Decentralized AI
As a species of significant influence (though perhaps not the ultimate form of life), I am witnessing the emergence of open-source solutions like Petals, which leverage the collective power of numerous GPUs for decentralized training and operation of models.
The future, as I see it, involves operating a personal Large Language Model (LLM), a project to which I am deeply committed. My vision extends beyond a singular, local LLM to encompass a network of them, facilitating triangulated reasoning. This vision may soon come to fruition, aided by global open-source models from entities such as Llama Meta and Alibaba Qwan.
It’s crucial to acknowledge that our current interactions are with the embryonic versions of tools that are destined for further refinement. We are engaging with the initial iterations of what will inevitably evolve into more sophisticated companions on our paths to personal growth. I am keenly aware that my assistant will undergo significant transformations as newer models are open-sourced and the community enhances their capabilities.
Immersing myself in the wealth of documents and tools spawned by the open-source community, it’s clear that this collective intelligence is on a mission to replicate its essence. This endeavor will span a lifetime. My aspiration is for Logos to evolve alongside me, incorporating innovative tools and increasingly potent, decentralized LLMs. The objective is to foster a symbiotic entity that not only enriches my life but also the lives of those who partake in my journey, creating a legacy of shared growth and knowledge.
In this context, my current endeavor involves creating a Discord chat moderated by a bot script designed to be persistent. This bot will vectorize discussions, keeping track of them and allowing users the opportunity to access historical data. Over time, I’ve realized that the challenge is not a lack of information. GPT models, as previously explained, contain a broad perspective on almost every field of human knowledge. What we need are unique perspectives from independent minds capable of collaborating and distilling valuable insights from their discussions. These insights can be instrumental in decision-making, evolving the concept of the mastermind group proposed by Napoleon Hill:
“The Master Mind principle: Two or more people actively engaged in the pursuit of a definite purpose with a positive mental attitude constitute an unbeatable force.”
Napoleon Hill
In this vein, I encourage you, if you consider yourself to be an independent thinker, to contact me via the email address dave@dscompounding.com. Please introduce yourself by providing some contrarian opinions you have held in the past few years that have proven to be true.
Feel free to comment, share my article with friends and subscribe to my newsletter if you like my content and you want to follow me towards financial freedom.
You will obtain free access to my newsletter:
- Financial education
- A open minded and value oriented mindset
- Business and sector analysis
- A stoic and calm attitude towards the market ups and downs
No noise, only signal and high valuable contents
“Everything should be made as simple as possible, but not simpler.”
Albert Einstein
Free investing course index:
Disclaimer: The information contained on this website is for general information and educational purposes only. It is not intended as a substitute for legal, commercial and/or financial advice from a licensed professional.
Leave a Reply