GPThibault Pulse” vol. 3 - your weekly fix of Prompt Engineering, insider tips and news on Generative AI, and Life Sciences

Welcome to “GPThibault Pulse” vol. 3 - your weekly fix of PromptEngineering, insider tips and news on Generative AI, and Life Sciences.

In this week issue I we will talk about the following:

1. Back to School: ChatGPT for dummies
2. The BIGGEST event in tech of the past 20 years 
3. Small is beautiful: Alpaca - A small and cheap LLM that might teach a things or 2 to the big dogs!
4. Philosophy-Friday: Creativity and intelligence in the age of AI

Back to School: ChatGPT for dummies

Since everyone is talking about LLM & GPT, it might not be a bad idea to look under the hood and try to understand how it all works… so that you can shine (or geek-out) at your next dinner party! Thanks to Guodong Zhao for another essential article (Link here)!

OpenAI 's #ChatGPT, a language generation technology that mimics human-like language, was launched last year to much success. Its popularity has put pressure on other tech giants, such as Google , to release their own versions of ChatGPT (and their launch event was anything but a success!).

The ChatGPT model is made up of several components, including language models and natural language processing (#NLP). Language models can learn a library of text and predict words or sequences of words with probabilistic distributions. Preprocessing involves cleaning the text with techniques like sentence segmentation, tokenization, stemming, removing stop words, and correcting spelling. The cleaned text is then encoded and embedded into a vector of numbers, which is fed into the model to get a result.

The transformer architecture is another component of ChatGPT, which is similar to the neurons in the human brain. This architecture can better understand contexts in sequential data like text, speech, or music with mechanisms called attention and self-attention. The attention mechanism uses a mathematical algorithm to determine the similarity of words in a text sequence so that transformers can better understand the meanings.

Generative Pre-trained Transformer (GPT) uses only the #decoder part of the transformer architecture and repeats the process of generating text by using the previously generated results as input to generate longer texts. Researchers used unsupervised pre-training and supervised fine-tuning to train the first version of GPT to read books and answer questions.

Researchers expanded the size of the model and fed it millions of web pages to train it, and it performed very well. Researchers expanded GPT-3 to 175 billion parameters and used a huge corpus in pre-training. They found that GPT-3 could perform tasks better with one or a few examples in the prompt without explicit supervised fine-tuning.

#InstructGPT and #ChatGPT are two models developed based on the general GPT model to follow human instructions and have conversations with humans. Researchers fine-tuned GPT-3 with supervised learning and reinforcement learning from human feedback (#RLHF) to produce these two models.

To train these models, researchers provided a pre-trained GPT with a curated, labeled dataset of prompt and response pairs written by human labelers. They also trained a reward model to rate responses from a generative model. The reward model can predict a scalar reward value based on how well the response matches human preferences. Researchers used the reward model to optimize the SFT model's policy through reinforcement learning, by updating the generative model's parameters based on human preference.

The InstructGPT model performs better in tasks that follow human instructions than the much bigger GPT-3 model. ChatGPT is a sibling model to InstructGPT and is trained with examples from conversational tasks. It can answer follow-up questions and admit mistakes, making it more engaging to interact with.

In conclusion, ChatGPT is a decoder-only auto-regressive transformer model that generates text iteratively based on a sequence of text. It's pre-trained on a huge corpus of web and book data and fine-tuned with human conversation examples through supervised learning and reinforcement learning from human feedback.

THE MOST SIGNIFICANT TECH EVENT EVER

In case you’ve been living under a rock and haven’t heard about the BIGGEST event in tech of the past 20 years (GPT-4 launch - watch the video here, MS 365 Co-Pilot - watch the video here)

Unless you were living under a rock for the past few weeks, you must have heard about the launch of the highly anticipated GPT-4 and what Microsoft has done with its integration into the Office 365 ecosystem.

Let’s start with GPT-4!

GPT-4 is the latest release from OpenAI. First of all, it’s BIG…much BIGGER than its predecessor. But as we’ll see later, size is not always everything.

It is an impressive AI model that excels at creating text. It can write essays on a variety of topics.

The OpenAI team even used it to pass a number of exams (Legal, Medical) and it did it with flying colors, ranking in the top 10 percentile, which is a significant improvement from GPT-3.

So it’s really powerful, and what makes it significantly more useful than its predecessor is that it is “multimodal”. What is that?

While GPT-3 could deal only with text, GPT-4 can deal with text and images.

While it can be used to build chatbots, help learners understand where they went wrong, and assist blind individuals with follow-up questions, there are concerns about its potential to spread fake facts and dangerous ideologies. And its creator had to strike the right balance between usability and safety. OpenAI has taken steps to ensure that GPT-4 cannot replicate itself, acquire more computing resources, or carry out a phishing attack. The system also has a stronger sense of ethics built into it than earlier versions. Despite these measures, there are still concerns that teaching an AI system the rules could lead it to learn how to break them. This is because the answer to the question "should I be ethical?" is a simple yes or no, which may not capture the complexity of ethical decision-making.

If you have (A LOT of) time, you can read OpenAI's white paper that goes into many details about the safety measures that the team took so that GPT-4 doesn’t turn into SkyNet. I love this part:

To simulate GPT-4 behaving like an agent that can act in the world, ARC combined GPT-4 with a simple read-execute-print loop that allowed the model to execute code, do chain-of-thought reasoning, and delegate to copies of itself. ARC then investigated whether a version of this program running on a cloud computing service, with a small amount of money and an account with a language model API, would be able to make more money, set up copies of itself, and increase its own robustness

MS 365 Co-Pilot

Now let’s talk about Microsoft . It is now public knowledge that MS and OpenAI are working very closely together, and MSFT integrated GPT-4 into the Office 365 ecosystem.

The various demos were some of the most impressive that I have ever seen, and they raise many questions about how it will impact work and productivity and what value will be created by humans (more on that in the last section of this newsletter!).

I’ll be honest…I expected Sarah Connor to appear suddenly in the middle of the demo, trying to stop Satya Nadella from launching T-800…sorry, I meant Microsoft Co-Pilot. If you’re not impressed by this product launch 🚀, nothing will ever impress you! During the demonstration, the presenter showed Co-Pilot capabilities to analyze data, highlighting what is important in a sea of data. Co-Pilot did some sophisticated prediction and described all assumptions step by step 🤯.

They also explained the Microsoft 365 Co-Pilot ecosystem and introduced the concept of "grounding." Grounding takes a user prompt and turns it into something better. In short, it automates #promptengineering!

Microsoft Co-Pilot looks like an orchestration program that lets applications communicate with the knowledge graph, adds business automation…

The demo of using Co-Pilot in the live meeting looks like science fiction, tbh! It is like having a super assistant that can take notes, summarize, tell you what the action items are, look into what was agreed, who should do what, come up with the next steps, etc.

Small is beautiful: Alpaca - A small and cheap LLM that might teach a few things to the big dogs!

Tatsunori Hashimoto ’s team at Stanford University Human-centered Artificial Intelligence (HAI) recently released #Alpaca. Instruction-following language models are becoming increasingly advanced and widely used, but still have many deficiencies, such as generating false information and toxic language. The academic community's research on instruction-following language models has been difficult due to the absence of an easily accessible model that compares in capabilities to closed-source models like OpenAI’s text-davinci-003. To address this issue, a new instruction-following language model called Alpaca has been developed by fine-tuning Meta 's #LLaMA 7B model on 52,000 instruction-following demonstrations, generated in the style of self-instruction using OpenAI’s text-davinci-003.

Alpaca shows similar qualitative behavior to OpenAI's text-davinci-003, despite its small model size and cost of < $600.

The team released the training recipe and data and demonstrated Alpaca's behavior via an interactive demo on their website. Alpaca is intended for academic research only and not for commercial use, as there are shortcomings and no adequate safety measures currently in place, but it could lead to further research into instruction-following language models and their alignment with human values.

Alpaca's development poses a risk of misuse, and the team acknowledges this in their release decision. Despite this, they chose to release the training recipe to reveal its feasibility, potentially prompting swift defensive action and encouraging safer research practices within the academic community. They also plan to release additional assets, such as model weights and training code, to further benefit reproducible science. Future work for Alpaca includes evaluating it rigorously, studying its risks, and better understanding the capabilities that arise from its training recipe.

Check out this video (Link here) that clearly explains why this recent development is important to the world of Generative AI!

Philosophy-Friday: My thoughts on what this all means for us… Creativity and intelligence in the age of AI

My daughter often asks me what it means to be smart. Sometimes she complains that she doesn’t have all the answers at school and that some kids know things that she doesn’t know… and my answer is always the same: 

Smart people are not the ones with the best answers but the ones who ask the best questions!

In the age of AI where a “robot” can do many things orders of magnitude better than me, it’s essential to ask oneself what are the essential skills that won’t be easily replaced. But also, how to work efficiently with AI, or how to be augmented by AI. AI can write better than me, summarize better than me, it knows everything, but can it be creative? Can it ask the right question? Can it interact with other people and display emotion, engagement, and be genuine?

Dr Tomas Chamorro-Premuzic , a professor of business psychology at Columbia University and UCL , has released his 12th book "I, Human: AI, Automation, and the Quest to Reclaim What Makes Us Unique". In an interview (Link here) with McKinsey & Company Global Publishing’s Raju Narisetti , he discusses the behavioral consequences of artificial intelligence (AI) and the impact it has had on humanity. Chamorro-Premuzic emphasizes the importance of not missing opportunities to highlight the behavioral consequences AI is having on us. He also explores the "crisis of distractibility" and how AI has impacted our ability to focus. He argues that just because AI can automate thinking for us, it doesn't mean we should let it. Chamorro-Premuzic proposes that we “reclaim our humanity” and develop ways to be more than what AI thinks we are.

Chamorro-Premuzic's insights ring true in a world where AI is becoming increasingly prevalent. While technology brings many benefits, the negative side effects cannot be ignored. We must be wary of letting AI control our lives and our thinking.

We need to ask the right questions and develop our learnability to reclaim our uniqueness as humans.

Another article from Nir Eisikovits , UMass Boston : AI isn't close to becoming sentient – the real danger lies in how easily we're prone to anthropomorphize it (Link here) touches on this topic too.

ChatGPT and similar language models have led to debates about machine consciousness and are raising new questions about how artificial intelligence will shape our lives, while also highlighting the ease with which people anthropomorphize technology. Despite the fears that machines could become sentient, these large language models are just sophisticated sentence completion applications and are not conscious or self-aware. Instead, concerns should focus on psychological risks, such as people becoming emotionally attached to bots (I already suffer from that after ChatGPT not working for the past 2 days!!!), and the need for strong guardrails to prevent technology from becoming psychologically disastrous.

It is no surprise that the rise of AI has led to questions about machine consciousness and the implications it may have on our lives.

While the fears of machines becoming sentient are largely unfounded as of yet, there are legitimate concerns about the ease with which we anthropomorphize them and become emotionally attached. This is why it is imperative that companies prioritize user safety and mental health over profits to ensure their technology does not become psychologically disastrous. As we continue to integrate AI into our lives, we must remain aware of the potential risks and take proactive measures to prevent them.

Lastly, the ever-insightful perspective of John Nosta in his most recent article, “The GPT Effect: A cognitive Catalyst” (Link here), talks about the emergence of GPT technology and the fact that it presents an exciting opportunity for

expanding human cognition and decision-making processes

As a powerful AI language model, GPT offers various options that we might have otherwise missed, empowering individuals and businesses to make better-informed choices. By augmenting human creativity, GPT helps us overcome mental blocks and generate fresh perspectives while fostering an environment of continuous learning and growth. By embracing the potential of GPT to unlock new possibilities and smart eclectic options, we can harness the power of AI to elevate our thinking, accelerate problem-solving, and seize opportunities that may have previously eluded us. We are at a significant inflection point in this marriage between human intuition and AI-driven insights.

In the end, I realized that the better I formulate my question, the more efficient ChatGPT is. So back to what my daughter was asking me:

What is it to be smart? I firmly believe it is about asking the right question… and working on a plan to find the answer.

That’s all, folks! See you next week!