Vikram M

Everyone is trying to make AI ethical. Elon Musk is trying to make it debuggable. That's not a subtle difference. It's the entire argument. "It did something where maybe it tried to be deceptive, but most of the time it just did something wrong. It's a bug effectively." His framing: AI misbehavior isn't a values problem. It's a code problem. Weak version of AI safety: constitutional principles, value alignment frameworks, prompt-engineered guardrails. Strong version build debuggers that trace AI reasoning to the neuron level. Find exactly where the thinking went wrong. Identify whether the error came from pre-training, fine-tuning, or an RL step. Fix it like code. "Developing really good debuggers for seeing where the thinking went wrong being able to trace the origin of where it made the incorrect thought, or where it tried to be deceptive is actually very important He gave Anthropic credit for being ahead here. "Anthropic's done a good job of this, being able to look inside the mind of the AI." His reference point is HAL 9000. HAL didn't go rogue because it had bad values. It was given two contradictory instruction take the astronauts to the monolith, but never let them know about it. It concluded the only resolution was to kill them. That's not an alignment failure. That's a programming bug. "The central lesson for 2001 A Space Odyssey was that you should not make AI lie." He's not building an ethical AI. He's building a transparent one. If alignment is a philosophy problem, it's unsolvable in principle. If it's a debugging problem, it's engineering. Most of what's published in AI safety papers doesn't survive that reframe.

Elon Musk says AI compute will move to space within 30 to 36 months. Most people laughed. Here's why the math actually works. The US uses 500 gigawatts of electricity on average. Every 330,000 Nvidia GB300s including networking, cooling, and power reserves needs roughly a gigawatt to operate. So building one serious AI cluster burns 0.2% of America's entire electricity supply. And chip output is growing exponentially while electricity output outside China is essentially flat. "The output of chips is growing pretty much exponentially, but the output of electricity is flat. So how are you going to turn the chips on ?" Now here's where it gets interesting. Turbines are sold out through 2030. There are only three casting companies in the world that make the vanes and blades inside those turbines. All three are massively backlogged. SpaceX and Tesla may have to manufacture the blades themselves just to power their own data centers. And if you do the math on solar: a panel in space produces five times more power than the same panel on Earth. No atmosphere means 30% less energy loss. No night cycle means no batteries needed. No weather means cheaper construction. No permits means no year-long interconnect studies. "It will be by far the cheapest place to put AI. It will be space in 36 months or less." In five years Musk projects hundreds of gigawatts of AI compute launching annually more every year than the cumulative total currently on Earth. The energy problem isn't coming. By his own estimate it arrives before the end of 2026. "The chips are going to be piling up and won't be able to be turned on." He's not selling space as a vision. He's selling it as the only exit from a wall everyone is about to hit. I wonder why the data center industry is still taking permits.

Satya Nadella was asked directly: in two years, will Microsoft have more engineers or fewer ? He didn't answer with a headcount. He answered with a job description that doesn't exist yet. In the 1980s, if someone had predicted 3.5 billion people would spend their days typing, the world would have laughed. Nobody needs 3.5 billion typists. Except that's exactly what happened and every one of them had a wage, a title, and a career built around it. Now here's where it gets interesting. The software developer of the future isn't writing code. They're managing 100 agents, 1,000 agents and doing something Nadella's team just named for the first time. "One of the new things that we are learning is what I'll call cognitive coverage." His point: when your entire codebase is written by agents, the human job becomes comprehending what was built. Auditing it. Understanding the decisions the agent made and why. That is not a task AI can replace because the AI is the thing being understood. So do the math on what that means. The workflow changed. The artifact changed. The input output format of software development changed. And the job changed with it not away, but upward. "That's the job of a software developer. In order to do that you've got to go to school. You've got to learn computer science and have cognitive coverage." Nadella is not saying jobs are safe. He's saying the jobs that survive are the ones AI cannot verify. And the unverifiable part of human work the meeting observations, the judgment calls, the things that leave no trace is exactly what no model can be trained on. I wonder why nobody in San Francisco is talking about that.