Ali Alsayed

Chatbots talk. AI agents work. One explains how to do things. The other actually does them.

@DruRly 🖐️

is it just me?

Reviewing your design work, live x.com/i/broadcasts/1…

Saas as an API endpoint for agents. Either do that with yours or lose.

Software makers need to stop thinking about: Screens and Dropdowns and start thinking about: Actions, endpoints, context and feedback loops Software is going to be used by #AI agents not humans

#AI agents will also hack. It’s all about the prompt and how humans will trick them into complying with malicious requests.

Reminder: #AI will always play catch up to humans. We all need to be in that category of humans to stay ahead and relevant in different jobs.

UI/UX Designers, did you know you can now use ChatGPT in Figma. FigGPT is an AI Figma plugin that will help you generate content for your UI design, summarise text, change text style and also generate design ideas. Repost to share with others 💜

At some point, middle management will disappear thanks to #AI. AI can assign tasks, manage performance and fire underperforming employees be it humans or agents. Which is pretty much what middle managers do.

@businessbarista So we’re calling “Build in Public” vibe marketing now?

Vibe marketing > vibe coding

Business owners are interested in business value. They’re not excited about #ML. Know how to pitch.

When your #ML model is too good to be true, it probably is. We need to know why it’s that good. Is it leaking data from the future? Depending on what you’re predicting, that’s one possible reason.

Historical data is the collective experience of an organization from which it can learn from. Yet some refuse to use it. I’ll never understand.

Why More Data Won’t Fix Feature Imbalance in #ML When training machine learning models, adding more data is often the go-to solution for improving performance. But here’s the catch: if the distribution of features isn’t consistent between the training and validation datasets, more data won’t help. Why does it matter? When feature proportions differ between datasets, the model learns patterns that don’t generalize to new data. This leads to poor performance in production, no matter how many instances you add. So what do we do Instead? 1. Ensure your training and validation datasets have similar feature distributions. 2. Use techniques like stratified sampling or data augmentation to balance your datasets. 3. Regularly monitor feature drift to catch issues early. It’s not just about how much data you have—it’s about having the right data. Balanced feature distributions are the key to building models that perform well in the real world.

Sometimes ( maybe a lot of times ) a less accurate #ML model is better.

Struggling to figure out if your #ML model is overfitting or underfitting? Use this simple two step approach: Step 1: Check Errors - Overfitting: Validation error is much higher than training error (e.g., 20%+ difference). The model might be memorizing random patterns in the training data. - Underfitting: High errors on both training and validation datasets. The model may be too simple to learn meaningful patterns. Step 2: Adjust Model Complexity - Train models of varying complexities and compare their errors. - The optimal model has the lowest validation error. If there’s a big gap between this error and your original model, it confirms overfitting or underfitting.

What makes a #ML model "complex"? It’s all about the parameters tuned during training. Different model types measure complexity in unique ways: 1. Linear Models: Complexity is determined by the number of features influencing the coefficients. 2. Decision Trees: Complexity grows with the number of conditions in tree nodes. 3. Tree Ensembles (e.g., Random Forests): It depends on the number of trees and the conditions in each tree. Neural Networks: Complexity comes from the network architecture, including layers, neurons per layer, and connections. 4. Neural Networks: Complexity comes from the network architecture, including layers, neurons per layer, and connections.

Accuracy is the wrong evaluation metric for most #ML projects.

An #ML project is a consulting gig more than a tech product.