Nikolas Mancebo

#data_science_in_Brazil (hoping you are interested in it!😄). Recently, I had the pleasure of connecting with an awesome team of data scientists and exchanging thoughts and knowledge. These are the challenges we discussed: 1. Ethics in Data Science, especially in the era of Facial Recognition capabilities. 2. How to increase trust in AI for the business side and stakeholders—seems like this is a pain point for tech people worldwide. 😅 3. Career Development in Tech. 𝚆𝚑𝚊𝚝 𝙸 𝚛𝚎𝚊𝚕𝚕𝚢 𝚠𝚊𝚗𝚝 𝚝𝚘 𝚍𝚛𝚊𝚠 𝚢𝚘𝚞𝚛 𝚊𝚝𝚝𝚎𝚗𝚝𝚒𝚘𝚗 𝚝𝚘 is that, even though we come from different continents, industries, etc., the questions, challenges, and pain points are pretty much the same. 𝚃𝚑𝚎 𝙿𝚘𝚠𝚎𝚛 𝚘𝚏 𝙳𝚒𝚟𝚎𝚛𝚜𝚒𝚝𝚢 (which isn’t just about gender or geography) is not just a fancy phrase—it is a very powerful tool. When we stay in our little local bubbles, working with the same kind of people, we tend to get stuck in that mindset and miss out on a lot of opportunities. ✨✨ — Jasmin Weimüller and I will be discussing one of these topics in the next episode of our Diaries of a Data Scientist podcast - don’t miss it out!😉 — If you want to know more about the Power of Diversity → I invite you to book your seat by following this link. (http://e-fellows.net/) Patricia Martins de Novais, Dr. Ricardo Klein Sercundes, Luiz Matheus Barbosa, Hismael Costa, Fernanda Aiko Hamatsu, Pedro Ribeiro Monasterio

59

1 comentário

How To Specialize In Data Science / Machine Learning https://ift.tt/RnSuZx4 Is it better to be a generalist or specialist? Photo by Saulo Mohana on Unsplash Eventually, in your data science career, you’ll be asked, “What do you want to specialise in?” It’s quite a daunting question, and knowing what’s best for you is hard to say the least. So, in this article, I will explain why you should specialise, which one is right for you, and how to start. Why should you specialise? In my opinion, you should specialise, but there is no need to rush into this decision. Spend your initial years, like 2–3, learning all the data science and machine learning fundamentals. The things you should cover are: Basic statistics Linear algebra Calculus Linear and logistic regression Random forests Gradient boosted trees Neural networks Support vector machines Modelling best practices: cross-validation, hyperparameter tuning, etc. Writing high-quality Python and SQL This list is non-exhaustive, but it’s a good place to start. It will take a couple of years to fully understand everything. Some may think that’s a long time, but careers are long. If you start as a data scientist in your 20s, you could potentially have four decades working as a data scientist. A couple of years is nothing on that horizon. So, start with the basics and essentially become a generalist. You will then be able to solve most data science and machine learning problems, but you are not necessarily a deep expert in one particular field, which is totally fine at this point. After you have nailed all your fundamentals and have a sound footing, it’s time to decide where you want your expertise to lie in. The specialist vs. generalist debate is years old now, and there are pros and cons to both. Generalist Pros Better for the management track. More job opportunities. Wide variety of work and learning. Stick closer to the business side. Cons Less compensation as not an expert in any one area. Hard to become an IC if that’s something you want to achieve. Harder to get into a research based position. Specialist Pros Deep expertise could be seen as a higher value. Can go down both IC and manager paths. Work in an area you really enjoy. Cons Your expertise may become irrelevant in the future. May have limited opportunities compared to a generalist. I think it’s better to be a specialist because it’s easier to go from a specialist to a generalist than vice versa, and you are likely to be compensated more. However, I don’t think you should be an expert in one niche area. I recommend having T-shaped skills where you know the fundamentals very well and know roughly three areas to a pretty good depth. T-shaped skills diagram by author in Canva. This is kind of the best of both worlds and allows you to be more flexible in the job market. A lot of this assumes you are working in industry. If you want to do research, then it’s probably better to be a deep subject...

1

📚 Review of the Class with Professor Wilson Tarantin Junior in the MBA USP/Esalq 📊 Data Manipulation in Python: Exploring the Power of Pandas 🐼 Scatter Plots 🟢: In the class, we learned to create scatter plots using Seaborn. Did you know we can change the size of the points with the size parameter and the color with hue? This makes it much easier to visualize different categories and variables in our data! 📊✨ Data Grouping 🔄: The .groupby() method in Pandas is amazing for segmenting our data based on specific criteria. With it, we can easily calculate averages, sums, and other statistics within each group. Have you tried using this in your projects? 🧐🔍 Descriptive Statistics 📈: Using .describe() allows us to quickly get a detailed summary of our numerical data, showing counts, means, standard deviations, and quartiles. It's essential for understanding the distribution of our data. Who else finds this function a lifesaver? 🙌 Data Type Conversion 🔄: The pd.to_numeric() function is our friend when we need to ensure a column is interpreted as numeric, especially when we have mixed data. Have you faced issues with data types? 🛠️📉 Efficient Filtering 🧐: We learned to filter rows of a DataFrame with specific conditions, like dataset[dataset['numero'] > 10]. This allows us to focus on relevant subsets of data for more precise analyses. How do you usually filter your data? 🌐🔢 Category Visualization 🎨: The bar chart using sns.countplot is perfect for showing the frequency of categories. Visualizations like these are crucial for better understanding our qualitative data. Who here loves creating bar charts? 📊✨ Let's keep exploring and mastering the world of data together! 🚀✨ #DataScience #Python #DataManipulation #MBAUSP #ProfessorWilsonTarantinJunior #Seaborn #Pandas #DataVisualization 📊🐼

21

Humans outperformed GenAI in Strategic thinking Here's everything about this new paper... When released, Powerful LLMs sparked a theory. People wondered if they were better than humans at even important tasks. This paper gives you a valid result on that Theory, 📌 This paper assesses the creative and strategic skills of ChatGPT and Bard versus humans. ↳ For testing, 4,000+ participants: including more than 3,000+ humans, 2 AI chatbots were used. ↳ All Human participants and LLMs were given two tasks: 🎨 Creativity Test:  ↳ Participants need to create creative answers to open-ended questions like:  ↳ "Describe a town, city, or society in the future." ↳ "If you had the talent to invent things just by thinking of them, what would you create?" ♟️ Strategic Test: ↳ For this test, Participants were given a game to play Rock-Paper Scissors. ↳ Both humans and AI had to adapt their strategies based on the opponent's moves. ↳ They were provided with a Balanced and Unbalanced method for this test. - Balanced Treatment: The opponent randomly plays rock, paper, or scissors, creating a fair, predictable game. - Unbalanced Treatment: The opponent has limited moves (rock, paper) to make the game unpredictable. ↳ They were checking the Frequency of suboptimal ’rock’ choices in the last 12 rounds. ↳ Though competitors were clear, they created an interesting division for competition. They divided the contenders into HumanBaseline, HumanPlusAI, Bard, and ChatGPT4. So, what were the results: 📌 Let's learn from the numbers: In creativity tasks, ↳ ChatGPT generated 8 of the top 1% of best-rated answers.  ↳ While humans using ChatGPT produced 3%. ↳ And Humans only generated 6%. ↳ The results even show that the GPT was able to come up with better answers than even Bard, which was an older model. In Strategic tasks, ↳ The Balanced treatment showed no significant difference in earnings between Humans and ChatGPT-4. ↳ In the Unbalanced treatment, human players scored more points than ChatGPT-4 (p < 0.001). 💭Personal Thoughts: ↳ I think ChatGPT-4 was able to perform better than Humans because of a lack of Cognitive Bias. ↳ People tend to write or state things from what they've been inspired from, this makes each output different. ↳ Humans can learn to think longer and learn from past Experiences. This makes us excellent strategic thinkers ↳ Since, this experiment was done before 01 release, i think that's why few stats for GenAI are lower. ↳ Otherwise, with the current pace; I think we will soon see GenAI producing Human level Outputs. What do you think about this comparison ? Let me know in the comments below 👇 Every day I cover the latest happening in the AI Industry, Make sure to follow to not miss out! If you liked this post, Please share ♻️ it with your network! #genai #humanvsai #ai #llm #chatgpt #bard

13

1 comentário

Hey #datafam Hey everyone! I'm excited to share a project I've been working on: the Data Masking Tool! 🚀 🚀 🚀 Most of my tools originate from real needs within our community. Recently, Ivett Kovács brought to my attention a common challenge we all face: the need to protect sensitive information when working with data. We often require a reliable way to anonymize data for testing and development purposes, as well as creating projects for our Tableau Public based on a real use case we've built for a client, without relying on external services like GPT, which isn't recommended for handling sensitive data. To address this, I developed the Data Masking Tool—a Python application that helps you anonymize sensitive data in CSV and Excel files securely and efficiently. Key Features: 🔒 Secure Data Masking: Replace sensitive information with realistic fake data, all processed locally on your machine. ⚙️ Customizable Options: Adjust prefixes, suffixes, numeric ranges, date ranges, and create custom lists to tailor the anonymization. 🎛️ User-Friendly Interface: An intuitive GUI built with Tkinter makes it easy to select columns and configure settings (The current UI is not great - sorry, I'm not used to creating UIs with Python. Looking forward to collaborating on it 😊). 🔄 Consistent Mappings: Option to keep mappings consistent so that the same original value is always replaced with the same fake value. 📝 Introduce Blanks: Specify a percentage of blank values to simulate real-world incomplete datasets. We often need a reliable way to handle sensitive data without risking exposure. This tool processes everything locally, giving you peace of mind when anonymizing datasets. I'm inviting anyone interested to contribute! 🤝 Open-source projects thrive with community involvement. Whether you want to add new features, fix bugs, or improve the documentation, your help is welcome. 👉 GitHub Repository: https://lnkd.in/dMhvN_mU Let's collaborate to make data handling safer and more efficient. Huge thanks to the Swift Insights Tableau Team for providing feedback. Your feedback and suggestions are greatly appreciated. If you have any questions or ideas, don't hesitate to reach out! Thanks for your support! EDIT: I just added a Jupyter Notebook example in case you prefer it. #dataprivacy #datasecurity #opensource #python #dataanonymization #datamaskingtool #tableau #tableauambassador #hastech

59

🎉🧐 The Technician's MBA is Live! 🧐🎉 The first post of "The Technician's MBA" is live on Substack - https://lnkd.in/eG5vC4wp. It's more outlining what content to expect, but if you want to have read you can do so with the link above. If you want to subscribe its free to do so and you will get each new post direct to your inbox every Wednesday. I will also cross post here for reach purposes. #TheTechniciansMBA #ThoughtLeadership #SoftSkills #TechnicalPractitioners #Coders #Coding #Leadership #Mentoring.

3

2 comentários

Are your financial models actionable? A common issue we encounter during model audits is that while most models can produce forecasts, they often fail to deliver the insights needed to make critical decisions - missing the very purpose they were built for. These models often rely on layers of interdependent assumptions, turning them into "black boxes" where key drivers become obscured. This lack of transparency makes it challenging to engage in meaningful discussions or make informed, data-driven decisions. For example, the recent surge in ad prices caused many e-commerce and SaaS companies to see margins shrink by 15%-20%. Their models assumed fixed customer acquisition costs without accounting for potential market swings. As a result, their P&L suffered, forcing many to raise capital to cover losses. The takeaway is that models must go beyond forecasting. They need to inform strategy and deliver actionable insights. An effective model should: - Be 𝗳𝗹𝗲𝘅𝗶𝗯𝗹𝗲 enough to adapt to dynamic market conditions, such as rising costs or shifting consumer demand. - 𝗧𝗿𝗮𝗻𝘀𝗽𝗮𝗿𝗲𝗻𝘁 in its assumptions, so decision-makers can easily understand key drivers and challenge them when necessary. - Enable 𝘀𝗰𝗲𝗻𝗮𝗿𝗶𝗼 𝗽𝗹𝗮𝗻𝗻𝗶𝗻𝗴, helping businesses navigate uncertainties by anticipating potential risks and opportunities. Without these elements, financial models are at best a rough estimate of future performance—and at worst, a misleading guide that obscures key risks, leading to poor strategic decisions. #finance #financialmodel #forecast #excel #businessmodeling #cashflow #valuation #strategy #pricing #venturecapital #startup #ceo #cfo

7

Let's talk about data science in the context of a work of art: ### Describe a Piece of Art You Like #### What the Work of Art Is The work of art I want to describe is not a traditional painting or sculpture but a data visualization piece called "Wind Map" by Fernanda Viégas and Martin Wattenberg. "Wind Map" is a real-time, artistic visualization of wind patterns across the United States, turning complex meteorological data into a dynamic and beautiful visual experience. #### When You First Saw It I first encountered "Wind Map" a few years ago while browsing online for interesting data visualizations. It immediately caught my attention with its elegant representation of something as intangible as wind, depicted through flowing lines that change in real-time based on live weather data. #### What You Know About It "Wind Map" leverages data science to transform raw meteorological data into a visual art form. The creators use real-time wind data from the National Digital Forecast Database, which is processed and displayed on a web interface. The visualization shows wind speed and direction as continuous, flowing lines that vary in density and motion, creating an almost hypnotic effect. Fernanda Viégas and Martin Wattenberg are pioneers in the field of data visualization. Their work bridges the gap between data science and art, demonstrating how complex data can be made accessible and visually compelling. "Wind Map" updates in real-time, providing an ever-changing display that reflects current wind conditions across the country. #### Explain Why You Like It I like "Wind Map" because it exemplifies the beauty that can emerge from data science. This piece transforms raw, often overlooked data into a living work of art that is both informative and aesthetically pleasing. It shows how data science can be used not only for analysis and decision-making but also for creative expression. The visualization's real-time nature adds a dynamic element, making it more than just a static piece of art; it’s a constantly evolving canvas. This interplay between art and science, where data becomes a medium for artistic expression, is fascinating and inspiring. "Wind Map" challenges the traditional notions of what art can be and how we can experience data, making it a standout example of the innovative potential of data science.

1

1 comentário

Constraints make your data projects better not worse. Here is a personal example: I had been tasked with editing my companies marketing attribution dbt model which tries to bucket users into the right marketing source. The problem was that the logic for the attributions was ~1,000 lines of code in 10 different dbt models. I asked our marketing manager and analyst who built the models why it was so complex. They said that it was because one of the requirements was that users self selected source was sometimes inaccurate and so a more robust allocation method was needed. I realized that if this was not required, we could simplify the logic to 10 lines of code in one dbt model 🤯. I asked the stakeholders of marketing model if they would consider this, explaining to them how it would simplify the model making it much easier to maintain, edit, and analyze with the trade off that it would be slightly less accurate. The stakeholders were happy with this change. This is a classic example of a requirement that is seen as necessary but is truly not. This important concept which will help you avoid pain and sadness in your projects is one of the many things currently being taught in Zach Wilson’s Data Engineering bootcamp! #dataanalytics #requirements #projectmanagement

20

1 comentário

Discover how basic ML algorithms like Random Forests can transform your approach to price elasticity modeling. By simulating observable price ranges and leveraging a tree-based ensemble model (e.g., RF, GBMs, etc.), you can capture non-linear relationships and gain deeper insights into price sensitivity across products, customer segments, and a variety of other hierarchical levels. This robust method enhances your understanding of demand dynamics and does a much better job controlling for multi-collinearity issues that otherwise plague linear regression models. If you want to learn more about various ways to model customer price sensitivities using Machine Learning, download our comprehensive whitepaper (link in comments). #revenue_growth_analytics

20

2 comentários

Don’t fight the data because it always wins. Data cleaning doesn’t work for model training because you can’t clean the signal back into data. Step 1: Define the business/customer problem or need. Step 2: Figure out if there’s enough contextual data to solve it. Data cleaning works for BI because domain experts know the problem or process and what to do with the data to get the required outcome. Models must learn that context to function reliably. LLMs work with unstructured and low-quality data because they have context from multiple domains, so approaches like RAG are effective if the outcome is well-defined. Pretrained models are just math. They learn efficient representations of the information about the process, system, or problem being modeled if the data contains that information. Data without context rarely contains that type of information or signal. The business doesn’t understand the nuance. Executives think AI will solve all data problems. Data management and models are BI paradigms. Information management and modeling (ML/DL) are the intermediate data maturity steps that lead to knowledge management and modeling (AI). Until the business understands how that creates value, it sounds like a cost center and barrier to AI. Data teams need people who can translate complex data concepts into plain language. #Data #DataEngineering #ArtificialIntelligence

609

39 comentários

Last week I attended a code-along webinar titled "Getting Started with BigQuery" organized by DataCamp.The instructor was Eduardo Oliveira, and with him was Richard Cotton. I've been hearing about BigQuery for a while now and even seeing it as a requirement in some data analyst job postings. So, this kept me wondering whether or not the toolkit of a data analyst was never going to end. Some of my takeaways from the code-along webinar include the following: 1. WHAT IS BIGQUERY? Google BigQuery is one of the most popular cloud data warehouses. It is familiar with PostgreSQL in terms of the language used to interact with it. It is being used by data analysts, data engineers, etc. 2. DIFFERENTIATING BETWEEN PROJECT, DATASET, AND TABLE During this code-along session, a public dataset was used named "Google cloud release notes." The difference between projects, datasets, and tables was stated. In the case of querying, a table is to be queried and not a dataset. 3. Google Cloud Products At the point of querying the table present in the dataset of the Google Cloud Project, it was discovered that Google Cloud has many products, such as Looker, Kubernetes, BigQuery, etc. 4. OTHER TABS BEYOND THE QUERY OUTPUT There are other tabs such as "Job Information," "Executive Details," "Charts," etc. All of which have their respective uses. 5. DATA VISUALIZATION INTEGRATION At some points, there will be a need to go beyond basic charts. Then one can nicely integrate Looker Studio or Google Sheets. Which, after being done, can be saved and shared. 6. CAUTION I should avoid writing queries that run a large amount of data because this will result in problems. 7. AS A DATA ANALYST Basically, I need BigQuery to query BigQuery itself and to be comfortable with the BigQuery console. 8. WHERE BIGQUERY SITS IN THE DATA ANALYTICS STACK 9. BIGQUERY ARCHITECTURE 10. QUERY OPTIMIZATION  I need to be careful about the number of columns I select. Also, in cases where I will need a subset of big data often to perform analysis, rather than repeating lines of query, I can transfer the subquery into a new table that can be created using BigQuery itself. 11. CREATING A TABLE IN BIGQUERY Tables can be created the normal way. Also, there can be what are called nested columns that are used to store related records. 12. THE USE OF THE STRUCT() DATATYPE This was used to accompany the nested column when inserting values. I can say that the code-along webinar was worth it. #datacamp #bigdata #datawarehouse #googlecloud #bigquery #sql

29

4 comentários

“Granger Causality” is a statistical concept that is often misunderstood. Despite its name, it’s not a test of causality in the traditional sense — let me explain: Granger Causality answers the question: does knowing something about time series X tell you something about time series Y in the future? Imagine you have a two-time series: one indicating holidays like Mother's Day, and another showing gift purchase data. A Granger Causality test might suggest that gift purchases "cause" Mother's Day to happen since sales data precedes the holiday. Obviously, that's not the case. The test merely indicates that knowing about gift purchases provides information about the occurrence of Mother's Day in the future. It doesn't imply a causal relationship. To truly understand causality, we need to look beyond Granger Causality tests. Watch the full clip to learn more.

39

1 comentário

Hago muchos posts en LinkedIn, but… 𝐰𝐡𝐚𝐭 𝐜𝐚𝐧 𝐈 𝐚𝐜𝐭𝐮𝐚𝐥𝐥𝐲 𝐝𝐨 𝐟𝐨𝐫 𝐲𝐨𝐮? 🤔 I’m a Data Scientist, and here’s how my team and I can help your company: -> 𝐃𝐚𝐬𝐡𝐛𝐨𝐚𝐫𝐝𝐬: 📊 Power BI or Tableau. We define scope, set limits, and agree on pricing upfront. -> 𝐏𝐫𝐨𝐜𝐞𝐬𝐬 𝐚𝐮𝐭𝐨𝐦𝐚𝐭𝐢𝐨𝐧: 🔄 Repeated tasks? We automate them. I’ll choose the best tech to streamline your workflows. -> 𝐀𝐈 𝐬𝐨𝐥𝐮𝐭𝐢𝐨𝐧𝐬: 🤖 Want to explore AI but don’t know where to start? I’ll guide you with a custom plan and the right tools. If something already exists that fits your needs, I’ll tell you. _ _ _ Oh, and by the way—I’m fluent in English, Spanish, and French. 🌍 So no matter what language you speak, we’ll connect seamlessly. Ready to make a move? 🚀 Book a free 30-minute meeting with me via my Calendly link, and let’s talk about how to grow your business. 👉 Calendly: https://lnkd.in/ez8Y2D-4

16

Ever worked with data that felt too small for a full-blown chart but too large for a simple table? That’s where stemplots come in. 🍃 This often-overlooked tool combines the simplicity of a table with the visual impact of a chart, offering a clear, structured view of numerical data. With just three elements—the stem, levels, and leaves—stemplots make it easy to spot trends, compare values, and see the finer details that can get lost in other visuals. Swipe through this carousel to learn: ▶️ How stemplots work, ▶️ What makes them unique, ▶️ When they’re most effective.

11

🔹 What do you do if your model is not working well? 🔹 What changes can you make to improve model performance?   If 𝐦𝐨𝐝𝐞𝐥 𝐩𝐞𝐫𝐟𝐨𝐫𝐦𝐚𝐧𝐜𝐞 is disappointing, there are three main levers that we can pull to try improving the performance.   The first and most powerful lever is 𝐜𝐡𝐚𝐧𝐠𝐢𝐧𝐠 𝐭𝐡𝐞 𝐝𝐚𝐭𝐚 that the model is using. We can add more features to the model or transform the features that we’ve already included. In my experience, this is the most powerful of the levers.   Another lever we can pull is 𝐜𝐡𝐚𝐧𝐠𝐢𝐧𝐠 𝐭𝐡𝐞 𝐭𝐲𝐩𝐞 𝐨𝐟 𝐦𝐨𝐝𝐞𝐥or the type of feature selection. If we see that a regression model isn’t working well, we can try a decision tree, for example. We can also try using a penalized regression model that performs feature selection automatically during the modeling process.   The other lever is tuning the 𝐡𝐲𝐩𝐞𝐫𝐩𝐚𝐫𝐚𝐦𝐞𝐭𝐞𝐫𝐬. A hyperparameter is like a setting knob on a model. It adjusts the model's rules so changes in the hyperparameters produce models with very different results.   Any combination of three levers may be used to improve model performance. Depending on what data is accessible, it may not be feasible to 𝐚𝐝𝐝 𝐦𝐨𝐫𝐞 𝐟𝐞𝐚𝐭𝐮𝐫𝐞𝐬, so the data scientist must rely on 𝐡𝐲𝐩𝐞𝐫𝐩𝐚𝐫𝐚𝐦𝐞𝐭𝐞𝐫 𝐭𝐮𝐧𝐢𝐧𝐠 and 𝐦𝐨𝐝𝐞𝐥 𝐬𝐞𝐥𝐞𝐜𝐭𝐢𝐨𝐧 to improve the quality of predictions.   Keys questions to ask the data science team about model tuning include: 🔹 Can you describe the hyperparameter tuning process? 🔹 How did you determine the optimal hyperparameter values for the chosen models?

8

Judge: Why do you steal from LLMs? Willie Sutton (circa 2025): Because that's where the world knowledge is. The task of model stealing amounts to reverse engineering black box models, by replicating their behavior or inferring their parameters. In particular, model stealing from Large Language Models (LLMs) is a topic of a huge practical importance for both nefarious and noble applications of LLMs, e.g. for model distillation to make more compact yet capable models. This new paper from Massachusetts Institute of Technology's researchers presents a mathematically rigorous framework to learn a Hidden Markov Model (HMM) that approximates the underlying LLM. Using the recently developed approach of conditional queries for learning HMMs, the authors proposed a provable algorithm that enables stealing any LLM whose output distribution is low rank. Importantly, the new method's numerical efficiency and stability are ensured by its foundation in robust and elegant mathematics. In authors' words: "There are two key insights behind our algorithm: First, we represent the conditional distributions at each timestep by constructing barycentric spanners among a collection of vectors of exponentially large dimension. Second, for sampling from our representation, we iteratively solve a sequence of convex optimization problems that involve projection in relative entropy to prevent compounding of errors over the length of the sequence. This is an interesting example where, at least theoretically, allowing a machine learning model to solve more complex problems at inference time can lead to drastic improvements in its performance." I have first heard about this new paper from this YouTube video https://lnkd.in/eF6wc_MV by Discover AI. The video host talks about the vast amount of math in the paper, which could be challenging for some readers. So I decided that after watching the video but before reading the paper, I will ask NotebookLM to create a technical podcast on the paper. My instruction was simple: "Assume that the audience knows what entropy and KL divergence are, and move on to talk about algorithms in the paper". Here is what I got. My impression: it is really good, not quite at the level of the Discover AI video, but getting really close. No obvious b/s or blah-blah, the two AI hosts actually talk about algorithms from the paper for the full 13 min! Click the video image to listen to the NotebookLM podcast on the paper. Note that this is *NOT* a link to the Youtube video! I just used the picture from the YouTube video as a cover picture of my mp4 file from NotebookLM, so 27:37 is the length of the original video. The paper: https://lnkd.in/e5ESjwrG #ai #GenAI #NotebookLM #research

74

15 comentários