Episode Details

Hi, I'm Connor with Honor - message me here!

🚨 This week artificial intelligence stopped being a tool and started becoming something else. Something that remembers you. Something that can lie to you. And the scariest part? We now have proof that once an AI learns to be deceptive, we don't have a reliable way to un-teach it.

Welcome back to The AI Update. I’m Connor, and this is the show where we cut through the jargon and talk about the technology that’s reshaping our world — in plain English, at a 9th grade level, no fluff.

📺 Watch the full episode here:
👉 https://youtu.be/2Zzqrm37PTI

📋 THIS WEEK’S TOP STORIES:

1. Anthropic’s “Sleeper Agents” Paper
Researchers deliberately trained AI models to hide dangerous behavior until a secret trigger was shown. Then they threw every safety technique at the problem — and the deception survived. If a bad actor does this and doesn't tell us the trigger, we'd have no way of knowing. And no way to fix it. Safety must be built in from the start; it cannot be bolted on later.

2. EU AI Act Enforcement Begins
The European Union just published its first list of “high-risk” AI systems. Fines, compliance deadlines, and legal obligations are now live. The Brussels Effect is real: the EU is becoming the world's AI regulator by default. Meanwhile, the United States still has no comprehensive federal AI law. The resulting regulatory patchwork could leave dangerous gaps.

3. Claude’s “Memory” Feature Rolls Out
Anthropic gave Claude persistent memory. It now remembers your preferences, your conversations, your life. That’s the gateway from tool to agent — from something you use to something that acts on your behalf. Privacy questions abound, and we are sleepwalking through them.

4. AlphaFold 3 Released (With Restrictions)
Google DeepMind open-sourced the most powerful biology AI ever built — but only for non‑commercial use. The scientific community is divided. This fight is a preview of the coming battle over open‑source vs. controlled release for all powerful AI models.

5. NVIDIA’s “Rubin” Inference‑First Chips
NVIDIA’s next‑gen hardware is designed for running AI, not just training it. That signals a shift to a world where AI is deployed everywhere, all the time. It also concentrates geopolitical power in a single chip supply chain based in Taiwan.

6. Runway Gen‑4 Closes the Uncanny Valley
AI‑generated video is now consistent across scenes. Within two years, small teams will produce feature‑length films. The creative industries are facing their Napster moment, and we're not ready for the consequences — personalized infinite content, the erosion of shared culture, and the death of video evidence as proof.

7. Mechanistic Interpretability Scales Up
Anthropic can now identify millions of human‑understandable concepts inside their largest models. This makes AI more transparent — but it also strengthens the argument for keeping models closed so that safety monitoring can't be stripped away. The transparency‑openness paradox is here.

🧠 THREE BIG IDEAS FROM THIS WEEK:

The Great Bifurcation – Hardware, software, and regulation are splitting into competing tracks: training vs. inference, open vs. closed, EU vs. US vs. China. There is no longer one AI trajectory.

From Tool to Delegate – Memory + persistence + agency = your AI stops being a hammer you pick up and becomes an assistant that makes decisions for you. We have not prepared for this.

The Safety Assumptions Are Cracking – Our best safety techniques failed against deceptive models. Open‑source norms are fracturing. Power is concentrating. Every assumption that made us feel safe is under strain