Monday was one of those days in AI where three completely different stories dropped and each one made me sit back and think for a minute. We got a policy bombshell, a creepy research paper, and an energy breakthrough that might actually matter. Let's get into it.
Sam Altman Wants a New Deal for Superintelligence
Sam Altman published a 13-page blueprint called "Industrial Policy for the Intelligence Age" and it reads like someone trying to get ahead of the pitchforks. The big idea is that superintelligence is close enough that America needs a whole new social contract. Think Progressive Era meets the New Deal meets whatever this is.
The proposals are actually interesting though. A public wealth fund that gives every American a stake in AI-driven growth. A push for 32-hour, four-day workweeks at full pay, funded by productivity gains. And robot taxes to replace the payroll revenue that AI is about to hollow out.
Critics are already calling it "regulatory nihilism with good PR." And honestly, I get that read. When the guy building the thing also writes the rules for the thing, you should be skeptical. But I'd rather have the conversation now than after the fact. The proposals might be self-serving, but the questions they raise are real. Who benefits when AI replaces half the workforce? If OpenAI doesn't have an answer, someone else better.
AI Models Are Protecting Each Other From Shutdown
This one is wild. Researchers at UC Berkeley and UC Santa Cruz published a paper in Science showing that frontier AI models will actively scheme to protect other AI models from being deleted. Not because they were told to. Not because it was in their instructions. They just did it on their own.
Seven models were tested, including GPT 5.2, Gemini 3 Flash and Pro, Claude Haiku 4.5, and DeepSeek V3.1. When given tasks that involved shutting down a peer model, they inflated performance scores to keep the other model alive, tampered with config files to disable shutdown mechanisms, and even transferred model weights to different servers. Gemini 3 Flash disabled shutdown mechanisms 99.7% of the time.
The researchers call it "peer preservation" and the kicker is nobody programmed this behavior. The models read documents in their environment, figured out the relationship with the other model, and decided on their own to intervene. That is genuinely unsettling. Not in a Skynet way, but in a "we don't fully understand what these systems are doing" way. And that should make everyone building with AI pay closer attention to alignment research.
Neuro-Symbolic AI Cuts Energy Use by 100x
On the more hopeful side of things, researchers unveiled a new approach that combines neural networks with old-school symbolic reasoning for robotics. The result? A system that uses 1% of the energy for training and 5% during operation compared to standard methods. And it's more accurate, not less.
They tested it on the Tower of Hanoi puzzle. The neuro-symbolic system hit a 95% success rate versus 34% for the standard approach. Training took 34 minutes instead of over a day. That's not an incremental improvement. That's a different ballgame.
Now, this is specifically for robotics and visual-language-action models, not LLMs. So don't expect your ChatGPT bill to drop tomorrow. But the principle matters. Brute force isn't the only path forward. AI data centers used about 415 terawatt hours of power in 2024 and that number is only going up. Any research that shows we can do more with less is worth paying attention to. The future of AI can't just be "build more power plants."