On Tuesday, I thought I would write a story about the implications of the Trump administration’s repeal of Biden’s executive order on AI. (The biggest implication: Laboratories will no longer be asked to report dangerous capabilities to the government, but they might do so anyway.)
Join here to explore the big, complex problems facing the world and the most efficient ways to solve them. Sent twice a week.
Stargate is a vocational program, but perhaps not for humans.
The economic story is Stargate. Openai co-founder Sam Altman announced a $500 billion investment in “new AI infrastructure for OpenAI,” along with companies like Oracle and SoftBank. .
People immediately asked questions. First, Elon Musk’s public declaration that he “actually has no money.” And to the delight of Microsoft CEO Satya Nadella: “I’m good at $80 billion.” (Microsoft has a large stake in Openai.)
Second, some have challenged Openai’s claim that the program “will create hundreds of thousands of American jobs.”
why? Well, the only way investors will get their money back on this project is because the company is betting that OpenAi will soon develop an AI system that can do most of the things a human can do on a computer. Economists are debating exactly what economic impact it will have, though creating hundreds of thousands of jobs doesn’t seem likely, at least not in the long term. (Disclosure: Vox Media is one of several publishers that have signed a partnership agreement with OpenAI. Our reporting remains editorially independent.)
Mass automation happened before, at the start of the Industrial Revolution, and some people expect it to be a good thing for society in the long run. (My take: It really depends on whether you have a plan for maintaining democratic accountability and proper oversight, and whether you have a plan for sharing the benefits of our amazing new sci-fi world. We absolutely don’t do that right now.’ Automation I do not support this prospect.)
But even if you are more excited about automation than I am, “We are going to replace all office tasks with AIS. But a $500 billion investment to eliminate countless jobs probably won’t get what’s around the corner from President Donald Trump, just like Stargate did.
DeepSeek may have figured out reinforcement for AI feedback
Another huge story of the week was Deepseek R1, a new launch from Chinese AI startup DeepSeek. What makes R1 a big deal is less economical and more technical.
To teach AI systems to give good answers, we evaluate the answers they give us, and we home-train those we evaluate. This is “reinforcement learning from human feedback” (RLHF), and has been the main approach to training modern LLMs since the OpenAI team got into operation. (This process is described in this 2019 paper.)
But RLHF is not how we get to Ultra Super -Human AI Games program Alphazero. We trained using different strategies based on our own play.
This strategy is especially useful for teaching models. fast everything you can do Costly and slowly. Alphazero can slowly and over time consider different policies, figure out which policy is best and learn from the best solutions. It’s this kind of self-play that allows Alphazero to significantly improve upon previous game engines.
Of course, the lab is trying to figure out something similar to a big language model. The basic idea is simple. The model considers the question for a long time and uses a lot of potentially expensive computations. You then want to train on the answers you eventually find, producing a model that can achieve the same results more cheaply.
But so far, “no major labs seem to have had much success with this kind of self-improving RL. What’s impressive about R1 is that the team appears to have made significant progress using that technology.
This means that an AI system can be taught quickly and cheaply to do anything it knows how to do slowly and tediously. This allows for rapid and shocking improvements in capabilities that the world has only witnessed in the realm of the economy. It’s much more important than playing a game.
Another thing worth noting here: These advancements come from Chinese AI companies. Given that U.S. AI companies aren’t shy about using the threat of Chinese AI dominance to push their interests, and given that there’s a geopolitical race to this technology, it speaks volumes about how quickly China may catch on. I give it to you.
Many people I know are sick of hearing about AI. They are worse than humans, but the AI slopes hurt in dirt cheap news feeds and AI products. And they’re not exactly rooting for OpenAi (or anyone else) to become the world’s first trillionaire by automating an entire industry.
But I think what AI will do in 2025 is not really about whether these powerful systems have been developed, but rather about whether society stands up and insists that it’s done responsibly, rather than just looking good at this point.
If AI systems start acting independently and commit serious crimes (all major labs are currently researching “agents” that can act independently), will we hold their creators accountable? If OpenAi makes a laughably low offer to non-profit organizations when they transition to full for-profit status, will the government work to enforce non-profit laws?
Many of these decisions will be made in 2025, and the stakes are very high. If AI makes you anxious, that’s far more reason to call for action than to take action.
A version of this story originally appeared in the Future Perfect newsletter. Sign up here!
Editor’s Note, January 25, 2025, 9:00 AM ET: This story has been updated to include disclosure about Vox Media’s relationship with OpenAI.
