Original author: Sam Altman
Original article: AIGC Open Community
At 5 a.m. today, OpenAI co-founder and CEO Sam Altman published an in-depth article Three Observations on his personal blog.
Three main observations were made about the world of AI: The intelligence level of an AI model is roughly equal to the logarithm of the resources used to train and run it;
The cost of using a given level of AI falls by a factor of 10 approximately every 12 months, and lower prices lead to greater usage; the socioeconomic value created by linearly growing intelligence is super-exponential.
The article also specifically emphasizes that OpenAI is making every effort to deploy AI agents, and these tens of thousands of intelligent agents will eventually be widely used and become virtual colleagues in various business scenarios. The performance of primary agents may be worse, but they will still penetrate into various fields and reshape the world economy.
The following is the original content:
Our mission is to ensure that general artificial intelligence (AGI) can benefit all of humanity.
Systems that are beginning to point toward AGI are already emerging, so we feel it’s important to understand the moment we’re in. AGI is a nebulous term, but generally we define it as systems that can solve increasingly complex problems at a human level across a wide range of domains.
Humans are toolmakers, with an inherent drive to explore and create that makes the world a better place. Each generation builds on the discoveries of the past, creating more powerful tools—electricity, transistors, computers, the Internet, and, soon, AGI.
In one sense, AGI is just another tool on the ladder of progress that humanity has built together. But in another sense, the advent of AGI marks a fresh start, and it’s hard not to feel like “this time it’s really different.” The economic growth ahead of us looks amazing, and we can even imagine a world where we can cure all diseases, spend more time with our families, and fully unleash our creativity.
Perhaps within ten years, every person on Earth will be more capable than today’s most influential people.
We continue to witness rapid progress in AI development. Here are three observations about the economics of AI:
1. The intelligence of an AI model is roughly equal to the logarithm of the resources used to train and run it. These resources mainly include training computing power, data, and inference computing power. It seems that as long as you are willing to invest any amount of money, you can get a consistent and predictable return. The scale laws predicting such returns are accurate over multiple orders of magnitude.
2. The cost of using a given level of AI drops by about 10x every 12 months, and lower prices lead to more usage. You can see this in the change in the token cost from GPT-4 in early 2023 to GPT-4 o in mid-2024, where the price per token dropped by about 150x in a year and a half.
Moore’s Law changed the world by doubling performance every 18 months; AI’s costs are falling even more incredibly fast.
3. The socioeconomic value created by linearly growing intelligence is super exponential. This result shows that we do not see any reason why exponentially growing investments will stop in the near future.
If these three observations continue to hold true, the impact on society will be enormous.
We are now starting to roll out AI agents that will eventually be used as widely as virtual colleagues.
Lets take the example of a software engineering agent, which is something we think is particularly important. Imagine that this agent ends up being able to do tasks that most top corporate software engineers with a few years of experience can do (a few days at most). It probably wont generate the greatest new ideas, will require a lot of human supervision and guidance, and will probably excel at some things but not others.
But you can still imagine it as a real, but relatively junior, virtual colleague. Now imagine there are 1,000 of these agents. Or a million. And imagine there are such agents in every area of knowledge work.
In some ways, AI may be economically similar to the transistor—a major scientific discovery that scales massively and permeates every corner of the economy. We don’t pay much attention to transistors or transistor companies, but we expect our computers, TVs, cars, toys, and so on to work miraculously.
The world won’t change overnight. It never has. In the short term, life will remain largely the same, and people in 2025 will still spend much of their time the same way they did in 2024. We’ll still fall in love, start families, argue online, go hiking, and so on.
But the future will come at us in a way that cannot be ignored, and the long-term changes to society and the economy will be huge. We will find new things to do, new ways to help each other, and new ways to compete, but these may be very different from the work we do today.
Autonomy, will, and determination are likely to become extremely important. Deciding correctly what to do and how to navigate a changing world will be of great value; resilience and adaptability will be skills worth cultivating. AGI will be the strongest lever of human will, and it will enable individuals to have a greater impact than ever before, not less.
We expect the impact of AGI to be uneven. While some industries may not change much, scientific progress will likely be faster than it is today, and this impact of AGI will likely outweigh everything else.
The prices of many goods might eventually fall dramatically (right now, the cost of intelligence and energy limit many things), while the prices of luxury goods and some inherently limited resources (like land) might rise more dramatically.
From a technical perspective, the path forward seems fairly clear. But public policy and collective opinion will be critical in how AGI is integrated into society. One reason we’re launching early is to give society and technology time to evolve together.
AI will permeate every area of the economy and society, and we expect everything to be smart. Many people expect that we will need to give people more control over technology than we have historically, including more open source and accepting that there will be trade-offs between security and personal empowerment.
While we would never want to act recklessly, and there may be some major decisions and restrictions related to AGI safety that will be unpopular, directionally we believe it is important to lean toward more empowerment of individuals as we get closer to achieving AGI. Another possible path we can see is AI being used by authoritarian governments to control their populations through mass surveillance and removal of autonomy.
Ensuring that the benefits of AGI are widely distributed will be critical. The historical impact of technological progress suggests that most of the metrics we care about (health outcomes, economic prosperity, etc.) get better on average and in the long run, but increasing equality does not seem to be determined by technology, and getting it right may require new ideas.
In particular, the balance of power between capital and labor could be disrupted, which may require early intervention. We are open to some ideas that sound exotic, such as allocating a certain computation budget to everyone on the planet so that they can use a lot of AI, but we can also see many ways to achieve the desired effects by continuously reducing the cost of intelligence.
By 2035, anyone should be able to call upon the intellectual capacity of all human beings in 2025; everyone should have access to unlimited wisdom to harness in any way they can imagine. Today, there are many talented people in the world who do not have the resources to fully express themselves. If we could change this, the world would create tremendous value, which would bring great benefits to all of us.
