Dean 🐶

Nice @ChatGPT use case for travelling runners. I give it my location and the run profile and ask it what’s the nicest location/route to do it. It’s been getting it right so far, and I’m discovering great places!

Just dropped a new episode of the MLOps podcast! I sat down with Liron Itzhakhi Allerhand, PhD in AI, ex-Microsoft, and someone who’s actually shipped LLMs to production — multiple times. 🎧 Listen here → https:// dagshub . com /blog/practical-building-llm-applications

Experiment tracking is often messy. Comparing runs, spotting key differences, and reproducing results takes way too much time. We just upgraded our Experiment UI to show diffs, zoom on charts, and reproduce everything with 1 click Check it out and let me know what you think.

X @premium support is so annoying. I tried to upgrade, it won't let me pay, and the support keeps telling me it's escalating me to a human, only to yet again send me a link for submitting a refund. Why is it so hard to pay you? @elonmusk

Active learning and auto-labeling automate improving data quality. Using models to annotate data, minimizes the need for manual labeling and enables teams to focus on reviewing only lower-quality samples.

When we think of machine learning, our minds often go to the models, algorithms, or predictions that seem almost magical. However, as someone with experience in the field, I can confidently say this: no magic happens without the foundation of quality data labeling. Over the years, I’ve seen firsthand how the success of a machine-learning model hinges on the clarity and accuracy of its labeled data. It’s not just about feeding the model data; it’s about teaching it to see patterns, recognize objects, and make sense of the world—something only well-labeled data can achieve. Data labeling might seem straightforward at first glance, but it’s far from simple. It takes thoughtful planning, attention to detail, and consistent quality checks. From choosing what to label to maintaining accuracy across thousands of samples, every step is crucial to success. Because at the heart of every great machine learning project lies one truth: better data, better results.

Building a high-performing model is only half the battle—making it useful and cost-effective in the real world is where the real challenge lies. I just had a great conversation with Natanel Davidovits, an AI/ML veteran who has led data science teams at companies like @Nike and @DoubleVerify. Our chat was packed with practical insights for going beyond academic benchmarks to deliver genuine business value. My 3 Takeaways from the Episode: 1. Start lean and scale: Don’t jump straight to massive models. Begin with a minimal viable model and only add complexity if the baseline demands it. This approach can save time, resources, and headaches. 2. Make business your bottom line: Accuracy and F1 scores are great, but money talks. Collaborate closely with product and business teams to translate ML metrics into revenue-driven KPIs. 3. Specialize to stay relevant: With LLMs and automated tools rapidly commoditizing general tasks, domain expertise, and product thinking are your secret weapons. Focus on understanding the problem space deeply—your career will thank you. Natanel’s insights will help you rethink your approach to model-building and long-term career growth. Tune in, it’s worth your time!

Excited for my end of year US tour 🇺🇸! Starting with @_ODSC West in SF—I'll be giving a talk about 'Active Learning & Auto Labeling with YOLO'. Then heading to Austin for @MLOpsWorld & NYC. If you're around, let's grab a coffee! ☕ #ODSCWest #MLOpsWorld