Closing out @MetLife’s TTX today with the lovely & brilliant Tamar Shapiro, who is also a Meta alum! We talked about my journey from machine learning engineer into AI education, how to handle AI overwhelm, using AI to build code that self-sustains, and did a demo building a global mortality rates dashboard and dating for further work. Thank you very much for having me, MetLife!

Prada’s latest campaign is a good reminder that the best marketing teams are curators of creative talent. They hired legendary animator John Pomeroy who worked on Charlotte’s Web, Beauty and the Beast, dozens of the classics. It feels familiar and nostalgic because it is.

Mixture of Experts (MoE) Model Architecture ExplAIned Meta’s new open source model Muse Spark Open just dropped, and it uses this foundational architecture called the Mixture of Experts architecture. Pretty much every AI model released today uses it, and it’s founded on this premise: a huge AI model doesn’t have to activate all its parameters for every input. With mixture of experts, or MoE, a router decides which expert networks to activate for each token. Instead of using the entire model every time, it selects specific experts, saving compute and processing time. The total parameter count can be huge, but only a fraction is active at any given time: the parameters of the selected experts, plus shared parameters used for every input. That’s what makes MoEs so powerful: the learning capacity of much larger models, while using only what they need for each input.

Google and AI are completely different, and understanding how they work changes how you use them. *To be specific: by “AI” here I mean Large Language Models, or LLMs. Google has technically always been AI – its PageRank algorithm, which decides which links to return to you, has always been a form of AI. And yes, Google is now augmented with LLMs. What we’re comparing in this post is traditional Google (using PageRank) and LLMs.* Lets break down the differences in plain English & answer some of the most commonly asked questions below (so you know when to use which tool): Q: What’s the biggest difference between Google and LLMs like ChatGPT? Google searches. LLMs predict. Google scans the internet and returns the most relevant websites using a ranking system. LLMs don’t search. They learn from massive amounts of text — and then predict what word comes next based on probability. Q: So are LLMs just looking things up? No. They’re word generators. They learn patterns in language, and then generate text that follows those patterns. That’s why they can sound confident — even when they’re wrong. Many - crucially, not all!!! - models today have been augmented with the ability to pull some data from the internet and insert it into its context window (i.e. the input into the LLM, the prediction machine) which reduces the risk of “hallucination,” (i.e. making things up) but even so, the model is still a predictor. It can (and often does) still mess up EVEN with web search added. Q: What’s the biggest tradeoff to watch out for? Hallucinations. LLMs are generators (& not a fact checkers). So they can make things up. That’s why tools like web-browsing and retrieval-augmented generation (RAG) are so important. They anchor generation to real data. But they’re not foolproof. Bottom line: If Google is a librarian, LLMs are improv actors who’ve read the entire library. Sometimes brilliant. Sometimes also wrong. But when you understand *how* they think — you can use them way more effectively. Leave your questions or comments in the section below 🥰👇 & share to spread the word!! This concept is simple and yet so important. You guys rock, sending love to you all!

AI watermarks act like invisible labels that help people tell if something was created by artificial intelligence. This new tech by Google DeepMind adds these watermarks to AI-generated content, but it has several issues that limit its effectiveness as a complete solution. What do you think about this tech? How would you propose we implement AI watermarking? Hi, I’m Harper, AI educator, engineer and advisor. ••• My goal is to democratize AI, so everyone can understand it, use it, and make informed decisions about it. Happy you’re here. ••• I have two degrees in Computer Science specializing in Artificial Intelligence from @Stanford, experience teaching PhD-level AI courses there, and 4+ years building AI at @Facebook / @Meta Follow @HarperSCarroll if you want to learn about AI, and let me know in the comments if there’s anything specific you want to learn about.

Effective Immediately, all organizations using PHP for their systems are advised to be prepared for an update addressing two critical vulnerabilities. PHP is widely used for web applications. These vulnerabilities often provide a direct path to the server's file system or database. If an immediate update is not feasible, it is recommend that organizations take the system offline until it can be fully mitigated. NSA will follow up shortly with more specific guidance.

In honor of @NVIDIA GTC in Berlin, here’s a video of me riding around San Francisco in the first production car to ship with @nvidiaai’s full self-driving stack, a @mercedesbenz ! I’ve been a huge fan of autonomous vehicles because they are statistically much safer than regular cars. Just noticing the graphic says the software + hardware spans through Level 5 (driverless anywhere a human could, in all driving conditions) but it’s actually Level 4 (driverless within defined areas or conditions). This video is from December 2025 :) thank you for having me, NVIDIA!!

NEARLY $2B NEWLY COMMITTED TO AI + BIOLOGY Biohub, the U.S. Department of Energy, and the NIH announced a $1.8B international push to generate the foundational data behind AI models that can predict — and eventually treat — human disease. It's the largest commitment to AI-ready biological data to date. Who's putting up what: @biohub, $500M: the founding commitment. $400M for new measurement tech (cryo-electron tomography at near-atomic detail, microscopy imaging millions-to-billions of cells in living tissue, tools to build and perturb biology at every level), $100M for research outside Biohub. @ENERGY, $500M+ over five years: through its Genesis Mission: cell measurement, AI analytics, modeling, and exascale supercomputing across the National Labs. @NIH, coordinating datasets: built through $500M+ in prior federal investment, and working with Biohub to standardize them for AI model training (the Bio Genesis Mission). @GoogleDeepMind, @IsomorphicLabs, @Meta, $300M collectively: into the Virtual Biology Initiative for the technologies and multi-modal datasets needed to build predictive models of life. Plus: the @AllenInstitute, @broadinstitute, @GladstoneInst, Human Cell Atlas, Human Protein Atlas, and Wellcome Sanger Institute aligning the science; @nvidia on accelerated computing; @RenPhilanthropy expanding data-generation funding. What it's all for: an open data commons to train a "virtual cell" letting researchers ask, predict, and answer biological questions *digitally* instead of running lab experiments first. In Biohub head of science @alexrives's words: a predictive model of biology could "dramatically accelerate scientific discovery by enabling scientists to perform experiments digitally." I sat down with him to talk about what Biohub is building & what the current roadblocks are. That interview is coming soon.

What are the benefits of open-source and open-weight models? Check out last week’s episode for the difference between “open source” and “open weight.” Welcome to the next episode of the Open Model Series with @mervenoyann, @huggingface AI/ML engineer & open-sourceress! Every week we will discuss open-source & open-weight AI: what it is, how it works, and how you can run them on your own compute. Episode four out next week. Please leave all your questions about open-source & open-weight AI in the comments! This is a super important topic that is novel to most people, and it can be challenging to grasp - so rest assured your questions are shared with many others!

What does it mean when an AI model is “open weights” and how is it different than “open source”? Welcome to the next episode of the Open Model Series with @mervenoyann, @huggingface AI/ML engineer & open-sourceress! Every week we will discuss open-source & open-weight AI: what it is, how it works, and how you can run them on your own compute. Episode 3 out next week; episode 1 linked below. Please leave all your questions about open-source & open-weight AI in the comments. This is a super important topic that is novel to most people, and it can be challenging to grasp - so your questions are shared with many others.