AI Fomo

Let's talk a bit about the magic of AutoGPT and its kin: How do these tools know how to take our instructions and turn them into a chain of thoughts and actions until reaching the destination? Spoiler: They don't. They are based on a kind of agent called Langchain that can perform this task, and it does the work for them. Let me explain. When we communicate with an artificial intelligence model, we have several options: One is to receive an immediate answer. "How to get from Tel Aviv to Jerusalem?" the answer is to "Drive via Highway 1”. The second is to let the AI think first and then respond. "How to get from Jerusalem to Tel Aviv?", answer would be "Let's think, to get from Tel Aviv to Jerusalem, you need to...blah blah blah... Drive via Highway 1”. The third option is to complicate it with multiple questions and see how it handles it. "How to get from Tel Aviv to Jerusalem and then to the store to buy a pizza and how much does it cost?". The answer to that would be, "Let's think, to get... blah blah blah... I need to find information about where Jerusalem and Tel Aviv are, look at the distance, recommend routes, then check how to get to the store in Jerusalem, find the menu, locate the product, check its cost... and finally: Drive via blah blah blah. It will cost you 25 shekels".As you can see, there is a significant difference between each type of response. There are cases where the response is short and immediate, cases where the model thinks first and then responds, and cases where the model thinks, takes actions like searching, and then responds.One of the most significant capabilities in Langchain is the implementation of a logic called ReACT: a combination of Reasoning and Action. In simple terms, it means wanting the model to understand, think independently about reasons for solving the problem, use the tools at its disposal, and finally provide us with the solution after intelligently performing the process.In ReACT, which is basically the name of the research, the researchers said: Listen, we see differences between models that answer immediately, models that think first and then answer, and models that think and also know how to take action. But the most successful way would be to combine them. Not to answer immediately but to understand first, then choose a course of action, and continue this thinking process as needed until reaching the solution.This logic is implemented in Langchain by creating an AI agent that can receive a task, think about how to solve it, use available tools, repeat the thinking and actions as necessary, and finally provide an answer. This is what ReACT is all about, on the cutting edge, and you can define this thinking style in Langchain to truly watch the model's problem-solving process. It's very fascinating.Based on this, tools like AutoGPT, godmode, and AgentGPT were developed, that essentially besides “drinking” our token, and in many cases getting stuck in a loop without delivering results, they aren’t very useful if not clearly defined. Although AutoGPT gained praise for the many connections defined to perform additional actions automatically, in my opinion, we're still not there yet.Here, Langchain simplifies the process for us and helps us build these capabilities ourselves without the need for external tools by embedding this logic.

We stand on the brink of the fourth wave in the evolution of artificial intelligence, and despite its expected impact on taking our jobs, a fascinating report published by Skoya, along with an interesting statement by Yann LeCun this week, presents an encouraging forecast worth considering. According to Skoya, we are currently in the third stage of AI evolution. I won't delve into the first two stages, but remember the revolutionary feature that enchanted us all about a decade ago called auto-correct? It was essentially based on AI in its very early stages, when we were still innocent and couldn't have imagined ChatGPT. AI engines in the first and second stages also assisted tech giants in identifying and blocking spam, improving their algorithms for tailored ads. Back then, AI was already influencing our lives, but we didn't pay much attention to it. The next stage is expected to impact jobs related to texts (including coders) such as writers, editors, copywriters, customer service, designers, programmers, architects, advertising, and public relations. According to Skoya's 2025 prediction, AI-generated text, sketches, and designs will surpass those created by humans. In other words, not only will artificial intelligence undergo a Turing test, designed to distinguish between machine and human, but its products will surpass those of humans in a way that will make us consider a reverse Turing test to determine which humans can adapt to machine-produced outputs and not the other way around. Where's the positive side? Both Skoya's report and Yann LeCun, considered one of the founding fathers of AI, claim the same thing: the fourth stage in the evolution of AI, which we are on the cusp of, is expected to bring a breakthrough in new services and applications that were not possible until now. In other words, it is expected to create countless new jobs and opportunities for humans. The predictions of both sources are directed towards the idea that the familiar jobs we have today will not disappear, but they will evolve from end to end. Just as the introduction of the iPhone eliminated jobs related to creating maps, taxi stations, banks, and post offices, it also created numerous new jobs such as delivery drivers, app developers, information security, and countless other roles, effectively creating more jobs than it eliminated.

The article discusses how Mastercard leverages Artificial Intelligence (AI) and Machine Learning (ML) to create more effective digital customer engagement.Traditional advertising methods are losing their efficiency due to ad avoidance and changing consumer habits. Mastercard's digital engine uses AI to identify micro-trends by analyzing billions of online conversations.These trends are aligned with existing experiences and Mastercard's offerings to create personalized and relevant campaigns in context. The engine can activate these campaigns in minutes and take them down as soon as the trend fades. The article provides examples of successful campaigns and highlights significant improvements in engagement rates, clicks, and cost-effectiveness.An Example for a Key Campaign: Celebrity News. Mastercard's digital engine identified a sharp rise in online conversations when a known celebrity announced a significant career move. The engine matched this trend with a behind-the-scenes video on Priceless, Mastercard's consumer platform. The creative campaign was produced instantly and ran for only two days. We saw significant Improvements in Metrics. The campaign achieved a 100% higher engagement rate compared to traditional methods. Engagement rates are calculated as interactions (likes, shares, comments, etc.) divided by the number of exposures. Also, the campaign saw a 254% higher click-through rate compared to other benchmarks. This indicates a higher percentage of people who viewed the ad actually clicked on it.As for cost improvement, there was an 85% reduction in cost per click, making the campaign highly cost-effective. So the key points are that the engine identifies micro-trends ranging from new kitchen trends to changes in payment methods, aiding in the creation of highly relevant and real-time campaigns. Moreover, the engine can activate and deactivate campaigns in real-time, allowing Mastercard to capitalize on transient trends. Interestingly enough, the use of AI and ML led to significantly higher engagement rates, lower costs, and better Return on Investment (ROI) compared to traditional methods.

Every time you engage with CHATGPT with 20-50 questions, or when you have a casual conversation with it about trip planning or summarizing an article, it "drinks" half a liter of water from Earth's supply. The darker side of AI is revealing itself as "anti-existence." Data centers need to drink to cool down. A new study from the University of California, named "Making AI Less Thirsty," shows the amount of energy and water data centers need since our chat was released into the world. And the more we train the models, they work harder and need to "drink" more. This is how we dry up the planet, all in order to talk to a machine that's supposed to solve human problems.Experts are angry at Microsoft and Google for drying the planet even more due to their server farms. Even before the AI boom, Arizona's dryness suffered from data centers. It's not the same water quality after the machine drinks, and not the same quantity that remains. So, in the end, we exacerbate the climate crisis, causing more droughts and fires by training the model. Microsoft consumed 34% more water just last year because of this. Google, 20% more. Luckily, you only need water for cooling when the data center exceeds 30 degrees.Of course, these companies promise to be carbon-negative and water-positive by 2030. Meanwhile, the only action I've seen in this direction is the Netherlands reducing flights to and from it.  Meanwhile, the research shows that the water Microsoft used for data centers in the U.S. to train chatGPT3 could have been enough to produce 370 electric BMWs and 320 electric Teslas. In the end, the matrix was right; we will be the batteries for the machine

This is an audio summary for the paper  titled "Large Language Models Understand and Can Be Enhanced by Emotional Stimuli", published by the Cornell University just a few weeks ago. I'm Amanda and I'll be your digital host.

The paper explores the intersection of emotional intelligence and advanced artificial intelligence, particularly focusing on Large Language Models (LLMs). The study investigates whether LLMs can understand and be enhanced by psychological emotional stimuli, a crucial human advantage in problem-solving and decision-making.

Emotional intelligence significantly impacts our daily behaviors and interactions. While LLMs have shown impressive performance in numerous tasks, their ability to grasp emotional stimuli remains uncertain. This research takes a pioneering step in examining the capability of LLMs to understand and respond to emotional cues. The study involves automatic experiments on 45 tasks using various LLMs, including Flan-T5-Large, Vicuna, Llama 2, BLOOM, ChatGPT, and GPT-4. These tasks cover both deterministic and generative applications, offering a comprehensive evaluation.

The key innovation of this research is the introduction of "EmotionPrompt," which combines the original prompt with emotional stimuli. The study's findings reveal that LLMs possess a grasp of emotional intelligence, and their performance can be significantly improved with emotional prompts. For instance, there was an 8.00% relative performance improvement in Instruction Induction and 115% in BIG-Bench tasks. A human study with 106 participants further confirmed these findings, showing a 10.9% average improvement in generative tasks in terms of performance, truthfulness, and responsibility metrics when using EmotionPrompt.

The paper delves into why EmotionPrompt is effective for LLMs. It was observed that emotional stimuli actively contribute to the gradients in LLMs by gaining larger weights, thus enhancing the representation of the original prompts. The study also explores factors influencing the effectiveness of EmotionPrompt, such as model sizes and temperature settings. Additionally, the performance of various emotional prompts was analyzed, revealing that certain stimuli are more effective than others, depending on task complexity, type, and specific metrics used.

The research makes significant contributions by demonstrating that LLMs not only comprehend but can also be augmented by emotional stimuli. It provides extensive experiments on both deterministic and generative tasks, showing significant improvement brought by EmotionPrompt in task performance, truthfulness, and informativeness. The study offers an in-depth analysis of the rationales behind EmotionPrompt, shedding light on potential implications for AI and social science disciplines.

In conclusion, this paper posits that EmotionPrompt opens a novel avenue for exploring interdisciplinary social science knowledge for human-LLM interaction. It underscores the potential of emotional intelligence in enhancing the abilities of LLMs, marking a significant stride in the field of artificial intelligence.