Tuesday, I was thinking I could write a story The effect of canceling the executive order of Biden in the AI of the Trump administration. (The biggest impact: labs that are no longer asked to report dangerous powers to the government, though they can do so.) But then two big and more important AI stories were dropped: a technical and an economic between them.
This story was first displayed in Future Perfect Newsletter.
Sign up here to explore big, complex problems and most effective ways to solve the world. Are sent twice a week.
Stargate is a job program – but maybe not for humans
Economic story Stargate. Together with companies like Oracle and Softbank, OpenAI co-founder Sam Altman declared a $ 500 billion has planned to invest “New AI infrastructure for OpenAI” – that is, for data centers and power plants that will be required to get them.
People were immediately questions. First, Elon was mask Public announcement “They don’t really have money,” then Microsoft CEO Satya Nadella’s reply: “I’m good for my $ 80 billion“(Microsoft, Remember, OpenAI has a big partnership.)
Secondly, something Challenge OpenAI claims that the program is “hundreds of thousands of American jobs.”
Why? Okay, the only reasonable way to invest in investors in this project is that if the company is betting, OpenAI will soon create AI systems that can do most of the work that people can do on the computer. Economists Controversy Exactly what Economic impact That would be, if it happened, though hundreds of thousands of employment creation seemed like one, at least not in the long run.
Happened before mass automation, at the beginning of the industrial revolution, and some people Hoping sincerely That long term it would be a good thing for society. (My idea: it really depends on whether we have plans to maintain the benefits of democratic accountability and adequate supervision and sharing the benefits of worrying new science-science world. At the moment, we are not at all, so I am not cheering on the possibility of being automated. )
But even if you are more excited about automation than me, “We will replace all office work with AIS” – which is fairly understood as a business model of OpenAI – this is an irrational plan to rotate as a job program. But then, $ 500 billion investment to eliminate countless jobs probably will probably not get the impression of President Donald Trump like Stargate.
Deepseek may find reinforcement on AI feedback
Another huge story of this week was Deepseek R1, A New release From the Chinese AI Startup Dipsic, the company that advertises OpenAI’s O1’s competitor. And what makes 1 bigger is that the economic impact is low and technical issues are high.
To teach AI systems good answers, we rate the answers they give us, and we train their homes for the high rates we give high rates. This is the “Reinforce Learning from Human Respi” (RLHF), and this is the main method of modern LLM training since working on an OpenAI team. (The process of the process This is described in the 2019 paper.)
But RLHF is not the way we have received the superhuman AI Game program Alphazero. It was trained using a different technique based on self-esteem: AI was able to invent new puzzles for itself, to solve them, learn from solutions and improve from there.
This technique is particularly useful for teaching a model Quick Can do something Expensive and slowly. Alphazero can gradually and time-timely consider many different principles, which one can find the best and then learn from the best solution. It is a self-style that has made Alfaziro greatly improve on previous game engines.
So, of course, labs are trying to find something similar for big language models. The basic idea is simple: you let a model consider a question for a long time, using potentially very expensive calculations. Then you will be trained on the answer found in the end, trying to create a model that can get the same results more cheaply.
But so far, “main labs do not seem to be very successful with this kind of self-development RL,” machine learning engineer Peter Smid-Nilsen Wrote In the explanation of the technical significance of Deepseek R1. Engineers seem so fascinated (and so concerned) by R1 that the team seems to have made significant progress using that strategy.
This means that AI systems can be taught to do anything quickly and cheaper that they know how to do slowly and expensive – which will only make the world witness to the economy some rapid and startling improvement with the economy. Games are much more important than playing.
Here is another significant information: these advances are coming from a Chinese AI company. Given that US AI companies are not ashamed to use Threat of Chinese AI domination To push their interests – and there is really a geopolitical competition around this technology – which says a lot about how fast China can catch up.
Still January
Many people I know are sick of hearing about AI. They are sick AI slop on their newsfeed And AI products that are worse than humans but dirt is cheap, and they are not exactly rooting to be OpenAI (or someone else) The whole industry automated the world’s first trillionier.
But I think that in 2025, AI would be really important – not for whether these powerful systems were created, which is going well at the moment, but whether the society is ready to stand and is insisted on whether it is done with responsibility.
When AI systems began to work independently and began committing serious crime (all of the major labs are To work at the “agent” Those who can work independently right now), will we be held accountable to their creators? If the OpenAI offers a ridiculously low offer to convert its non -profit entity into a full profit, will the government take steps to implement non -profit laws?
Many of these decisions will be taken in 2025 and the bet is very high. If AI makes you uncomfortable, this is a much greater reason to demand a step rather than the reason for tune.
A version of this story originally appeared in the Future Perfect Newsletter. Sign up here!
