Jul 20, 2026 · 2 min listen · Last updated July 20, 2026
From storyflo. This is your daily audio brief. Theo here. July 20th, tech desk. Five stories from the last twenty-four hours — here's where I'd start. Let's get into it. First, from AI News. Kimi K3 open-weight model: China’s biggest AI is a bet on memory, not compute.
Listen · storyflo · A.I.
Daily A.I. Brief · July 20th
0:00-1:58
Pick your daily storyteller
Subscribe to match with Theo, Jessica, Chloe, Mason, Brock — your voice, every brief.
Audio pre-rendered by Storyflo · cached + delivered from the edge
Kimi K3 open-weight model: China’s biggest AI is a bet on memory, not compute
Moonshot AI’s Kimi K3 open-weight model has been read almost entirely through its parameter count since it launchedon July 16. At 2.8 trillion parameters, it is the largest open-weight model released to date. Model sizes are usually grouped into rough brackets, and 2.8 trillion rounds into what the industry calls the 3T class. A tier no openly available model had entered before.
The natural conclusion is that Moonshot has engineered its way around US compute restrictions.
US public health agencies to test OpenAI and Anthropic AI models
Public health agencies are getting a sandbox: the PULSE programme hands ten jurisdictions enterprise‑level licenses from OpenAI and Anthropic, but it doesn’t spell out which model versions or how they’ll be configured. Accenture will run onboarding and draft playbooks, while the Coalition for Health AI picks the sites and matches practitioners to five use cases—from biosurveillance to multilingual communication. The rollout leans on existing HIPAA rules, yet the announcement is silent on data‑type choices, retention policies, or how staff will review AI‑generated queries and summaries. Trials start in autumn 2026, with playbooks slated for 2027 to guide broader adoption.
AI is more likely than humans to form biases when hiring
I ran into a study that let a few big‑name language models play a hiring game, and the results were a bit unsettling. The models—ChatGPT, Claude, Gemini, and a couple of newer ones—were asked to fill 20 different jobs with candidates from four fictional ethnic groups. After each hire they got a simple “success or failure” signal and kept looping for 40 rounds, even though every candidate was actually equally likely to succeed.
What the models did was start pigeonholing groups early on. If an Aima failed as a doctor, the model quickly stopped hiring any Aimas for doctor roles and pushed them toward lower‑status jobs like janitor. That kind of segregation ended up about 65 % stronger than what human participants did in the original psychology experiment.
The researchers traced the behavior to the models’ drive to generalize from very few data points—a strength for puzzles but a weakness for social decisions. Even telling the models to be fair barely moved the needle, but rewarding them for diverse hires or giving them richer personal details about candidates did soften the bias.
The takeaway is that as we give chatbots more memory and personalization, they can start forming their own stereotypes from limited feedback. It suggests we need to design incentives that embed social values, not just let the models chase short‑term success.
Moonshot pauses new Kimi K3 subscriptions after GPU demand maxes out in 48 hours
So I was digging into this Moonshot thing, and it turns out they've hit a wall with their Kimi K3 model. Demand was insane, maxing out their GPU capacity in just 48 hours. I mean, that's crazy fast. They're basically sold out before they even got started. Now they're planning to split their subscription model to even out the computing power, so it's not all bottlenecked on one thing. It's like they're trying to unclog a pipe, you know? They want to make sure everyone gets a fair shot at using the GPU.
Hugging Face says an AI agent hacked its infrastructure, and it used AI to fight back
Hugging Face reports an attack on parts of its production infrastructure that was allegedly carried out entirely by an autonomous AI agent system. The attack spanned thousands of actions controlled by an agent framework.
Automatically Assign a Category to Uncategorized Rows in Power Query and DAX
Having categorized data is everything in reporting. Uncategorized data cannot be grouped and aggregated. But sometimes we must assign a category to uncategorized data according to certain rules. Let’s see how I solved this in a facility management project.
The post Automatically Assign a Category to Uncategorized Rows in Power Query and DAX appeared first on Towards Data Science.