They made it completely useless

🦔Found a supercut of Jensen Huang saying "AI" 121 times during the CES keynote. That's about once every minute for two hours. Meanwhile actual business adoption fell to 12%, Dell stopped AI-first marketing because consumers don't care, and Microsoft quietly cut Copilot sales targets in half. The hype machine is working overtime but the receipts aren't matching. I figured we could all use a smile! Hedgie🤗

"We expect adversaries to keep adapting." Yes. That's what adversaries do. This is not insight. This is the definition of the word. "Prompt injection is unlikely to ever be fully solved." We knew this before we shipped. We shipped anyway. That decision made someone a billionaire. Several someones. None of them are in the rapid response loop. "But we're optimistic." Optimism is what you have when you don't have a solution. We have a lot of optimism. It's in the S-1. "A proactive, highly responsive rapid response loop." That's seven words to describe reacting after something breaks. "Proactive" means we'll watch it break faster. "Highly responsive" means we'll tweet about it within the hour. "Rapid response loop" means we'll do this again next week. And the week after. Forever. We've productized whack-a-mole and called it strategy. "Materially reduce real-world risk over time." Not eliminate. Reduce. Materially. Over time. These are the words of someone who has talked to Legal. Legal said we can't promise safety. So we promised trajectory. Trajectory is a direction, not a destination. You can walk toward a cliff your entire life and never arrive. As long as the graph shows downward risk, the board is satisfied. The board doesn't use the product. "Combining automated attack discovery with adversarial training and system-level safeguards." I've read this sentence four times. It means we'll use AI to find the holes in our AI. Then we'll train the AI on the holes. Then new holes will appear. Then we'll find those. This is not security. This is an arms race we're running against ourselves. We are the adversary. We just don't bill ourselves that way. "Identify new attack patterns earlier." Earlier than what. Earlier than the last time we didn't identify them. The bar is on the floor. We're proud to step over it slightly faster. "Close gaps faster." The gaps exist because we shipped before closing them. Now we close them in production. On your data. With your prompts. Thank you for your participation in our security program. It's mandatory. You agreed to it. It was on page 47 of the terms of service. "Continuously raise the cost of exploitation." Not prevent. Raise the cost. Exploitation is still possible. It's just more expensive. Like healthcare. If you're sophisticated enough, you can still do it. We've made prompt injection a luxury good. Nation-states can still afford it. Script kiddies cannot. This is progress. We've written a blog post about it. The blog post has a graph. The graph goes up and to the right. It measures "adversarial friction." I don't know what that means. Neither does the person who made the graph. But it goes up. And up is good. We're optimistic.

Hegseth: "We mean to increase acquisition risk in order to decrease operational risk. Let me say that again. We need to increase acquisition risk in order to decrease operational risk ... An 85% solution in the hands of our armed forces today is infinitely better than an unachievable 100% solution endlessly undergoing testing."

I'm glad to see Sachs finally getting roasted, and I hope that @Columbia takes note of how Russia's propaganda services have been using his tenure at their institution to launder their delusions into Western media. It's gotta stop, y'all.

Is that good

Brian Eno Is Practicing Ambient Music the Right Way