Dustin Coates

scikit-learn is more important than ever, even (and especially) in the age of GenAI! We talk about that with the guys from :probabl. on the latest Practical AI Podcast! #ai #datascience #podcast

25

1 commentaire

// AI Generated Code that's Nearly Impossible to Debug // "Now, thanks to ChatGPT, you could improve this workflow by: - making step 3, the only easy step, somewhat easier (and thus hiring less competent engineers) - making step 5 nearly Impossible."

1

✨ Think RevOps just has a magical fix for all your problems? Think again! ✨ What most people don’t realize is how even the simplest questions or requests to a RevOps team can spiral into a tangled mess of complexity.  Here’s why this can happen if RevOps teams dont have a proactive/agile approach to data governance, process documentation, & modular architecture: 💥 The Domino Effect Ever had someone ask, “Can we just get a report on X?” Sounds easy, right? That seemingly simple request can set off a chain reaction: you need to pull data from multiple sources, ensure it's clean and consistent, validate it against existing Reports/KPIs, and then format it in a way that makes sense to everyone. Before you know it, that “simple” report has consumed days of work, uncovered tech debt to fix,  and required the coordination of multiple teams. 🚧 Unintended Consequences One small change in your process or tech stack can ripple through the entire system. Let’s say someone wants to tweak the lead scoring model. Seems straightforward until you realize that change will affect every downstream process—automated workflows, sales handoffs, reporting metrics, and even how markrting nurtures prioritize accounts. What started as a quick adjustment turns into a major overhaul, with plenty of opportunities for things to break along the way. 🔄 Constant Recalibration A simple request like “Can we segment our contacts by XYZ?” sounds innocent enough. But in reality, it can lead to an endless loop of recalibration. You’ll need to validate the values of the field in question, update outliers, redefine the segments, update workflows, adjust reporting, and retrain the team on what was done. And let’s not forget the constant back-and-forth to ensure everyone’s on the same page about what those segments actually mean. 🐇Data Rabbit Holes You might think pulling a quick data point is no big deal, but data is rarely straightforward. A request for “last quarter’s sales by region” might lead you down a rabbit hole of data inconsistencies, missing fields, and conflicting sources. You’ll need to dig deep, reconcile differences, and verify accuracy—all while balancing the original timeline for the ask. By the time you have the answer, what seemed like a five-minute task has turned into a full-blown data archaeology project. In RevOps, sometimes even the simplest asks have a way of snowballing into massive undertakings.  It’s not just about doing the work—it’s about anticipating the ripple effects, managing expectations, and keeping everything running smoothly while you’re at #RevOps #RevenueOperations #ExpectationsVsReality #DataComplexity #ScopeCreep #TechChallenges #RevOpsLife

92

4 commentaires

🔄 Prioritize GenAI Projects Based on Your Maturity in Key Areas 🔃 I just finished giving this webinar jointly with Bern Elliot, who came up with the #GenAI "immaturity model" - like a baby, GenAI's age in the enterprise is measured in months, not years. The webinar is available for replay https://lnkd.in/gYXRATFt.

43

2 commentaires

📚 New Blog: "Into to NLP with TF-IDF Vectorization" In this post, I perform a deep dive into one of the OG natural language processing (NLP) techniques of the last 50 years with TF-IDF vectorization! If your at all curious about the origins search indexing, this is a great method to learn about. Up until 2015, ~83% of text based recommender systems used in digital libraries still relied on TF-IDF as their tool of choice (according to Wikipedia)! 🔍 Topics Covered in this post: - Calculating TF-IDF from Scratch - One-Hot Encoding - Cosine Similarity - Search Term Relevance Ranking with TF-IDF This was an extremely fun post to write, as it gave me a much stronger understanding on the fundamentals of search indexing and cosine similarity in vector space, which are fundamental to RAG applications. Check it out below, & let me know your thoughts & feedback in the comments! https://lnkd.in/gj4ctJ9w #ArtificialIntelligence #MachineLearning

16

1 commentaire

One more intresting presentation of Paul Tremblot de la Croix "How Generative AI Can Help Startups Succeed" at the #VDS2024. And my note on how AI can drive success: 💬 speakers emphasized that data is the foundation. Generative AI applications are only as good as the data they're built on. The ability to combine real-time data with generative capabilities is a game-changer for startups. 🕵‍♂️ shared the importance of customization. Don’t adopt generative AI just because it's trendy. Tailor it to your business needs, ensuring it’s the right tool for your use case. 👮‍♀‍ discussed data privacy and security concerns, emphasizing GDPR and other regulations. Startups must handle customer data carefully and integrate strong security protocols when deploying generative AI tools. Key points: 1. Generative AI + RAG = Personalization. By using AI with RAG, startups can retrieve relevant real-time information or past interactions, ensuring personalization not just based on predictions but also on actual preferences, history or external knowledge sources. 2. Customization & Use Case Fit. It's crucial to pick the most efficient AI model, considering cost, latency and accuracy. Generative AI isn't one-size-fits-all! 3. From MVP to Production. Building an AI-powered product is just the first step. Measuring, tracking and fine-tuning AI in production is essential for maintaining customer trust👌 I appreciate your meaningful insights shared about scaling AI strategically. Also on keep security tight and focus on solving real customer problems and think about there needs 😇 #AI #Startups #VDS2024 #DataPrivacy #Cybersecurity

4

About a year ago, we started building a product for portfolio tracking. The idea came from our customers who wanted to do more with portfolio company data – their most proprietary dataset – but didn’t have this data in a usable format. Being the naive entrepreneurs we are, Anurag Abbott and I thought, "How hard can it be? We’ll knock it out in 2-3 months." News headline, it is HARD. It took us about 14 months in the end; during which we had hundreds of conversations with investors, built prototypes, ran surveys, and got multiple iterations in the hands of our customers. After selling multiple million dollars in ARR, this is what I’ve learnt about portfolio monitoring: 1️⃣ No one has a system they really love right now I can count on one hand, nay one finger, the number of investors I’ve spoken to who were happy with their portfolio tracking processes. I also have data to back my assertion. Only ~14% of funds in our survey were satisfied with their current portfolio tracking processes. 2️⃣ Portfolio tracking is a significant cost for funds A majority of funds are spending 5+ hours on a single PortCo update, and more than $500K per year on portfolio tracking, including personnel and tools. Portfolio tracking involves tedious, labour-intensive processes – having to extract data from multiple formats, clean up inconsistencies and work with nuance around each company / update requires a significant amount of time and effort. 3️⃣ Private company updates are the wild wild west Funds spend months chasing scattered Portfolio updates that they receive as PDFs, as Excel sheets, as slides, as emails – and within these, there’s data in graphs, data in text, graphs pasted as images, graphs that don’t have data labels. There are hundreds of variations in how companies name their metrics. It is impossible to use this data as is – funds must spend hours extracting, cleaning and standardizing it before it can enable any analysis. (As an aside, I have a lot of sympathy for investors now. Our investor updates are a lot more standardized and better.) 4️⃣ You need humans in the loop. We were all aboard the LLM and AI train – but quickly realized that while AI can help with data extraction, human oversight is necessary for bringing higher-order standardization, nuance and real-world context to portfolio tracking. For now. We're sharing some of our internal research on Portfolio Intelligence for public consumption – this includes 125 Funds across North America and Europe. If you are in the business of investing in and advising companies, the report is a must-read. https://lnkd.in/g-376yWC

75

Don't sync activities from multiple tools to Salesforce. Yet, we see it all the time: Multiple tools sync SOME activities for SOME teams to Salesforce. The ripple effects: - Duplicates - Reporting nightmares - Bloated tool investment Here are 6 ways how to fix it: 1) Map your GTM tech stack List tools, capabilities, and exact form of activity capture 2) Map teams to tech stack Check which teams have access to which tools 3) Outline activity capture stack Map the exact activity capture stack by team and activity. 4) Calculate the investment Save money with consolidating towards a budget-efficient solution. 5) Fix gaps Fix gaps so you can properly report on activities. Bonus: Consolidate Centralize activity capture with a budget-efficient provider like e.g. Weflow | getweflow.com for sales, SE, CSMs. Everyone BUT your SDRs/BDRs that often live in your sales engagement tool where activities are tracked. For calls, you usually have one provider - do the same for email & meetings. What would you add? Comment👇 #revops #salesforce #gtm

14

4 commentaires

Our AI research team at Martian is advancing one of the most important areas of AI: Mechanistic Interpretability. This involves understanding how models make decisions by examining their internal mechanisms, similar to looking inside a clock to see how it works. Our research has long-term implications for AI, with safety being a primary consideration. By understanding the inner workings of AI models, we can better predict and mitigate potential risks, ensuring more reliable and secure AI systems. We are putting the outputs of this research to work today in our LLM Router, which predicts the best large language model (LLM) for a given prompt based on a customer’s cost, latency, and output quality preferences. #AI #MachineLearning #ArtificialIntelligence #TechInnovation #MechanisticInterpretability #Research #MartianAI #AITransparency

20

3 commentaires

One reason RevOps isn’t strategic 𝗟𝗮𝗰𝗸 𝗼𝗳 𝗮 𝗥𝗲𝘃𝗢𝗽𝘀 𝗥𝗼𝗮𝗱𝗺𝗮𝗽 Let me break this down. 𝟭. 𝗔𝗹𝗶𝗴𝗻𝗺𝗲𝗻𝘁 𝘄𝗶𝘁𝗵 𝗧𝗼𝗽 𝗣𝗿𝗶𝗼𝗿𝗶𝘁𝗶𝗲𝘀 RevOps defaults to a reactive whack-a-mole, fielding a flood of inbound requests. This leaves no time to work on strategic priorities. A RevOps Roadmap starts with top priorities or OKRs (Objectives and Key Results). - What does the Revenue Team want to accomplish this year? - This quarter? - What’s our sales target? - What do we need to do to get there? - Improve Close Rates? - Sell larger deals? - Generate more pipeline from outbound? - From inbound? - From renewals and expansion? What are our priorities as a team? 𝟮. 𝗢𝘂𝘁𝗹𝗶𝗻𝗲𝗱 𝗜𝗻𝗶𝘁𝗶𝗮𝘁𝗶𝘃𝗲𝘀 The next step is to align priorities with specific initiatives in RevOps. Say we want to improve our close rates. - What will we do? - Define our sales cycle? - Set clearer qualification and stage entry criteria? - Build reports and dashboards to track this? - Improve our pipeline review and/or forecasting process? - Provide more enablement? What are the most important initiatives to help achieve our objectives? 𝟯. 𝗚𝗲𝘁𝘁𝗶𝗻𝗴 𝗦𝗽𝗼𝗻𝘀𝗼𝗿𝘀𝗵𝗶𝗽 Now we get feedback, revisions and sponsorship from leadership. At the end of the day, are they in full support of RevOps prioritizing these initiatives and trying to help the Revenue Team achieve these Objectives and Key Results? If not, what are we doing here? If everything is a priority, nothing is. 𝟰. 𝗜𝗻𝘀𝗶𝗴𝗵𝘁𝘀 We have specific Key Results for the revenue team to achieve. - Are we tracking metrics related to this? - Are we reporting this to leadership on a regular basis? - Are we updating leadership as to whether or not we’re on track? - And if we need to make adjustments to get there? To me, RevOps isn't RevOps without Insights. It's just systems admin. 𝟱. 𝗦𝗽𝗿𝗶𝗻𝘁 𝗣𝗹𝗮𝗻𝗻𝗶𝗻𝗴 The Roadmap comes to life in sprint planning. Over the next two weeks... - Which items on the Roadmap will we tackle? - Which fire-drills and inbound requests do we have? - Does leadership want to delay items on the roadmap? Things change in growth-stage companies quickly. It’s okay to pivot and tackle emergencies if they are truly more urgent than what’s on the roadmap, so long as everyone is aligned. 𝟲. 𝗤𝘂𝗮𝗿𝘁𝗲𝗿𝗹𝘆 𝗥𝗲𝘃𝗶𝗲𝘄 The Roadmap needs reviewed and updated quarterly. We show what was done and not done, what OKRs the revenue team achieved and didn’t achieve. We then leverage this to set OKRs and Initiatives for the next quarter. We want to maintain alignment here. What would you add? What’s working with your Roadmap? 🤔 More on this here: https://bit.ly/3W8fhCq ✌️

52

5 commentaires

💥 New day, new model 💥 We ran our performance benchmarks on Anthropic Claude 3.5 Sonnet v/s OpenAI's GPT-4o & the results weren't what you'd expect ⏬ Anthropic says Claude 3.5 Sonnet outperforms all competitor models on key evaluations, at twice the speed of Claude 3 Opus and one-fifth the cost. In our benchmarks 3.5 Sonnet & 4o came out neck & neck. Ultimately, you need to run evaluations on your task at hand and measure the quality yourself. 🔍 Data Extraction: Both models identified 60-80% of data correctly, but neither excelled in this task. 📊 Classification: Claude 3.5 Sonnet (72%) outperformed GPT-4o (65%) in mean accuracy. However, GPT-4o led in precision (86.21%), crucial for accurately classifying customer tickets. 🧠 Verbal Reasoning: GPT-4o outperformed Claude 3.5 Sonnet with 69% accuracy, versus 44%. Both models did well on analogy and relationship questions. 📜 Context window: Claude 3.5 Sonnet boasts a larger context window (200K) compared to GPT-4o's 128K 🌪 Latency: Claude 3.5 Sonnet is 2x faster than Claude 3 Opus, but it’s still lags behind GPT-4o when it comes to latency (50% faster Time to First Token) Price and throughput for these models was very similar Want to learn more about how these models stack up? Check out our full analysis on the Vellum blog: https://lnkd.in/dwEcQJe4 Oh and did I say, Claude 3.5 Sonnet is already in Vellum? Sign up for a demo to run your evals & start building!

68

1 commentaire

A time-series foundation model is valuable because time-series data is ubiquitous across industries, yet traditional modeling approaches often fail to fully exploit the complex temporal dependencies and cross-domain similarities present in such data. We introduce Moirai-MoE: the first and state-of-the-art time series foundation model in universal forecasting. Takeaways: 1. Autonomous Specialization: The model autonomously achieves token-level specialization, enhancing efficiency and performance. 2. Performance Boost: Delivers a remarkable 17% improvement over its predecessor, Moirai, without increasing the model size. 3. Two model variants: 117M parameters with 11M activated parameters and 935M parameters with 86M activated parameters 4. Limitations of Existing LLMs: Current large language models (LLMs) struggle with time-series forecasting tasks. Moirai-MoE outperforms GPT4-o, a model 1000+ times its size in time-series tasks. 5. We need more time-series foundation models!!! Paper: https://bit.ly/3O1yiRQ Code: https://bit.ly/48FAF6i Models: https://bit.ly/3YNozDY

246

17 commentaires

I co-hosted an AI engineering leaders dinner on Thursday with Grace Ge. Thank you to Beyang Liu from Sourcegraph, Devendra Chaplot from Mistral AI, Bryan Bischof from Hex, Marco Sanvido from Robust Intelligence, and Tim Shi from Cresta for joining us. Here are the a few key takeaways: 1️⃣ AI hardware has a long way to go, but folks in the community are generally excited about how it can influence our day-to-day. Right now, they're toys showing us glimpses into the next generation of consumer tech. I respect rabbit inc. / Humane for starting to make progress in this area, but I think there's a lot more to come 😁 2️⃣ We're at an interesting crossroads between engineering and research in the AI community. There's a lot of respect on *both* sides, but I find that these conversations often pendulum swing between specific research areas and very pragmatic scaling/tuning insights. This paper on Chinchilla by Rohan P. came up specifically as important new work: https://lnkd.in/gD6tmDRv 3️⃣ Everyone feels that AI can help them get out of the busy work, and help focus them on the high-leverage parts of their jobs. There does seem to be a tension, though: those solutions are still "not quite good enough" but the ones that make for cool demos look like problem-seeking solutions. The products mentioned were Granola and Agenda Hero. 4️⃣ Large categories of companies within AI are already being commoditized quickly. There was a lot of talk about which companies are taking the "smart bets" in some of these competitive categories, but it feels to me like these 5-7 year strategies will start paying dividends and creating category kings within the next year or so. Specific excitement around players like turbopuffer. 5️⃣ Data privacy remains a hot topic. It's not clear if the incentives line up to push in the privacy direction for most large companies, and I think a lot of R&D cost is being driven by what AI talent *wants to work on* rather than productionizing the whole ecosystem. Thanks again to Grace Ge for the wonderful venue and Sunil Dhaliwal + the Amplify Partners team for co-hosting with Orb!

73

7 commentaires

**𝐔𝐧𝐥𝐨𝐜𝐤𝐢𝐧𝐠 𝐭𝐡𝐞 𝐏𝐨𝐰𝐞𝐫 𝐨𝐟 𝐆𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐯𝐞 𝐀𝐈: 𝐖𝐡𝐲 𝐏𝐫𝐨𝐦𝐩𝐭 𝐓𝐞𝐦𝐩𝐥𝐚𝐭𝐞𝐬 𝐌𝐚𝐭𝐭𝐞𝐫** Ever asked a generative AI (#ChatGPT, #Gemini) to produce content and found the results a bit off? The secret lies in 𝘱𝘳𝘰𝘮𝘱𝘵 𝘵𝘦𝘮𝘱𝘭𝘢𝘵𝘦𝘴. These structured prompts tell the model 𝘸𝘩𝘰’𝘴 𝘴𝘱𝘦𝘢𝘬𝘪𝘯𝘨, what the 𝘤𝘰𝘯𝘵𝘦𝘹𝘵 is, and even when it should 𝘴𝘵𝘰𝘱. Imagine directing a play: without clear stage cues, even the best actors might stumble. The same goes for AI. Think of it like ordering coffee: "A large coffee" might get you what you need, but "A large, double-shot cappuccino with oat milk" guarantees you the exact drink you want. 𝐈𝐧 𝐛𝐮𝐬𝐢𝐧𝐞𝐬𝐬, 𝐩𝐫𝐞𝐜𝐢𝐬𝐞 𝐩𝐫𝐨𝐦𝐩𝐭 𝐭𝐞𝐦𝐩𝐥𝐚𝐭𝐞𝐬 𝐭𝐮𝐫𝐧 𝐠𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐯𝐞 𝐀𝐈 𝐟𝐫𝐨𝐦 𝐚 𝐬𝐢𝐦𝐩𝐥𝐞 𝐫𝐞𝐬𝐩𝐨𝐧𝐝𝐞𝐫 𝐢𝐧𝐭𝐨 𝐚 𝐭𝐚𝐫𝐠𝐞𝐭𝐞𝐝 𝐩𝐫𝐨𝐛𝐥𝐞𝐦-𝐬𝐨𝐥𝐯𝐞𝐫, 𝐩𝐫𝐨𝐝𝐮𝐜𝐢𝐧𝐠 𝐜𝐨𝐧𝐭𝐞𝐧𝐭 𝐭𝐡𝐚𝐭 𝐚𝐥𝐢𝐠𝐧𝐬 𝐰𝐢𝐭𝐡 𝐲𝐨𝐮𝐫 𝐧𝐞𝐞𝐝𝐬 𝐚𝐧𝐝 𝐞𝐱𝐩𝐞𝐜𝐭𝐚𝐭𝐢𝐨𝐧𝐬. Mastering prompt design is key to harnessing AI's full potential—whether for client engagement, content generation, or data-driven insights. For business leaders, understanding this is the first step in turning AI into a strategic asset. 💡 #GenerativeAI #PromptEngineering #AIforBusiness #informationtechnology #industryacademia #researchimpact #promptengineering

19

In the next 3-6 months you will likely start using agents to perform tasks for you on your computer. You will simply give them a task, and they will do it while you go do something else—or simply chill. This is not science fiction. It is already here in a limited way with Anthropic’s new computer use capability. It is almost certainly coming from OpenAI and Google shortly. This could be even more disruptive than ChatGPT. Think of all the APIs that you no longer need, all the third-party API integrators you no longer need (e.g. Zapier), all the RPA companies that become largely obsolete (e.g. UiPath), all the SaaS companies that become basically data storage facilities for AI to use (e.g. Salesforce). Anything you, or anyone you know, can do at a computer, these agents will be able to do. They might be a bit slower to start, and they might not have the expertise to do absolutely everything well initially. And they might start out relatively expensive! But given that o1 is smarter than PhD experts, and the new Claude Sonnet has shown impressive creative capabilities (such as in Minecraft construction projects), it won’t be long until intelligence, creativity, and computer use converge, giving everyone an A+ employee they can assign tasks to execute autonomously.

11

2 commentaires

Presenting Orpomis Prime 7B model series ! These models are a result of trying different merge strategies of the following models: - mistral-orpo-beta from KAIST AI - Hermes-2-Pro-Mistral-7B from NousResearch In my experiments, dare ties merge leads to a significantly better-performing model on the openllm benchmarks In general, the model performance is quite solid in text generation as well as math reasoning. Full result comparison: https://lnkd.in/gb92DXwG Dare ties Model: https://lnkd.in/gMS-H4Sq Model stock Model: https://lnkd.in/giwEJ7Gs

9

I like the insights in this paper review: https://lnkd.in/ggCejTFy The conclusion applies to more than linguistics. In an age where most people think that modeling is purely a data+GPU game, having linguistic knowledge can help a lot in understanding the problem at hand, explaining results, and testing properly. Especially in the language technology space. Of course data+GPU is most likely just as important (as we are seeing with these massive LLMs, MT, and ASR models), but having both will still give you an edge. #ai #datahungry #gpuhungry #linguistics #asr #mt #llm #machinelearning

96

9 commentaires

Let's be real for a sec. RevOps' calling card is clean, accurate data, is it not? If your data is unhealthy, your sales and marketing teams don't function well. A total "no $#!t, Sherlock" situation, we think. Insycle's Health Assessment feature will scan for these common errors, and more: 🥕 Records with missing data fields. It’s hard to ramp up marketing automation efforts if you have critical data missing from large portions of your customer records. 🥦 Standardize fields for consistency. You need consistent data to segment and organize leads. For instance, if you wanted to segment customer by state, there may be multiple ways a state name is recommended within your database (New York vs. NY vs. New York State, etc.) 🥑 Identify duplicates and deduplicate records using any data field. You don’t want to send the same marketing messaging to the same prospect twice. 🌽 Fix formatting issues. For instance, you can use @Insycle to identify first names with capitalization issues. Would you rather your emails be addressed to “JANE,” “jane,” or “Jane?” Any of above sounds familiar? With confidence in your healthy data, you can do more: 🎁 All you have to do is login to your Insycle account, and you’ll have access to your updated Customer Data Health Assessment. 💥Once identified, our system makes navigation to the right tool and template to fix those issues within Insycle simple. All you have to do is click the “Fix” button and you’ll be directed to the right place - making identifying and fixing these issues a breeze. #data #datamanagement #datahealth #revops #sales #marketing #hubspot #salesforce #pipefrive #intercom #mailchimp #docsincrocs

1

Let's be real for a sec. RevOps' calling card is clean, accurate data, is it not? If your data is unhealthy, your sales and marketing teams don't function well. A total "no $#!t, Sherlock" situation, we think. Insycle's Health Assessment feature will scan for these common errors, and more: 🥕 Records with missing data fields. It’s hard to ramp up marketing automation efforts if you have critical data missing from large portions of your customer records. 🥦 Standardize fields for consistency. You need consistent data to segment and organize leads. For instance, if you wanted to segment customer by state, there may be multiple ways a state name is recommended within your database (New York vs. NY vs. New York State, etc.) 🥑 Identify duplicates and deduplicate records using any data field. You don’t want to send the same marketing messaging to the same prospect twice. 🌽 Fix formatting issues. For instance, you can use @Insycle to identify first names with capitalization issues. Would you rather your emails be addressed to “JANE,” “jane,” or “Jane?” Any of above sounds familiar? With confidence in your healthy data, you can do more: 🎁 All you have to do is login to your Insycle account, and you’ll have access to your updated Customer Data Health Assessment. 💥Once identified, our system makes navigation to the right tool and template to fix those issues within Insycle simple. ☮️All you have to do is click the “Fix” button and you’ll be directed to the right place - making identifying and fixing these issues a breeze. #data #datamanagement #datahealth #revops #sales #marketing #hubspot #salesforce #pipefrive #intercom #mailchimp #docsincrocs

1

2 commentaires