Xinzhiyuan Report
Editor: Editorial Department JYH
【Xinzhiyuan Guide】Sam Altman's essay "The Gentle Singularity" has caused a huge stir in the AI community. He predicts that humanity is moving towards an era of superintelligence, and the singularity will not arrive overnight, but rather permeate quietly.
After the o3-pro release, Sam Altman immediately published his latest blog post – The Gentle Singularity.
The article begins with a sense of urgency:
Humanity has crossed the "event horizon" of AI development, entering a phase of exponential acceleration.
Altman believes that the singularity will not arrive overnight but will permeate quietly: wonder becomes normal, and normal becomes the baseline.
In the text, he also provides his predictions for the technological timeline over the next five years:
2025: AI agents capable of real cognitive work (such as programming) will emerge, reshaping the way work is done;
2026: AI may gain the ability to discover new knowledge;
2027: Robots are expected to perform tasks in the physical world;
2030: Individual productivity will far exceed that of 2020, leading to significant changes in societal production methods.
It's worth noting that this is Altman's last personally written article, with no AI content.
He also pointed out another overlooked aspect: the consumption of each ChatGPT query.
An average ChatGPT query consumes about 0.34 watt-hours of electricity, roughly equivalent to a toaster running for just over a second, or an energy-efficient light bulb illuminated for several minutes.
Additionally, each query consumes about 0.000085 gallons of water, approximately 1/15th of a teaspoon.
Next, let's explore the future of AI through Altman's perspective.
Humanity Crosses the Tipping Point,
It's Now
We have crossed the event horizon; the "ascension" has begun.
In astrophysics, an event that crosses the event horizon cannot affect an external observer.
Humanity is one step away from building digital superintelligence, but so far, the process has been far smoother than imagined.
Robots aren't yet walking the streets, and most people aren't talking to AI all day. We still experience birth, aging, sickness, and death, and we still can't easily travel into space; there are still many unsolved mysteries in the universe.
However, we have indeed built intelligent systems that in many ways surpass humans, and they will greatly boost productivity. Scientific breakthroughs once thought impossible have been achieved; it is these hard-won insights that have given rise to systems like GPT-4 and o3, and they will lead us further into the future.
In many respects, AI will contribute to the world.
Among these, AI will accelerate scientific progress and increase productivity, leading to a huge leap in quality of life. The future will be much better than the present. Scientific progress is the greatest engine of overall development, and the thought of infinite future possibilities is incredibly exciting.
In an important sense, ChatGPT is already more powerful than anyone in history.
Hundreds of millions of people rely on it daily for increasingly important tasks: tiny new features can have a huge positive impact, while small deviations, when used by hundreds of millions, can cause enormous negative impacts.
Five-Year Timeline
2030 Becomes the Next Milestone
By 2025, AI agents capable of real cognitive work will have arrived, and the way code is written will be permanently changed. By 2026, humanity may welcome systems capable of discovering new knowledge. By 2027, robots capable of performing tasks in the physical world may emerge.
More people will be able to create software and art. But global demand for both will also soar. As long as new tools are embraced, experts are likely to remain far ahead of novices.
All in all, by 2030, individual productivity will far exceed that of 2020, which will be a striking change, and many people will find ways to benefit from it.
In some very important respects, 2030 may not be dramatically different. People will still care for their families, express creativity, enjoy games, and swim in lakes.
But in other equally important ways, 2030 may be fundamentally different from any past era. We don't yet know how much AI can surpass humans, but we are about to find out.
In the 2030s, intelligence and energy—that is, creativity, and the ability to turn ideas into reality—will become unprecedentedly abundant.
For a long time, these two have been fundamental limiting factors for human progress; with abundant intelligence and energy (and good governance), we could theoretically have everything.
AI Self-Reinforcing Flywheel Accelerates
Singularity Permeates Quietly
We are already living in an era of incredible digital intelligence, and after the initial shock, most people have become accustomed to it.
We quickly went from marveling at AI's ability to generate beautifully written paragraphs to expecting it to write beautifully written novels; from marveling at its ability to make life-saving medical diagnoses to expecting it to develop cures; from marveling at its ability to write a small program to expecting it to found a new company.
This is how the singularity develops: yesterday's miracles become commonplace, then become standard.
From scientists, OpenAI learned that their productivity increased two to three times compared to before using AI.
Advanced AI is significant, but perhaps nothing is more important than using it to accelerate AI research itself. We may discover new computational substrates, better algorithms, and other pleasant surprises. If a decade of research can be completed in a year or even a month, the pace of progress will obviously be completely different.
From now on, existing tools will help us gain deeper scientific insights and assist us in creating superior AI systems.
Of course, this is not the same as AI systems autonomously updating their own code, but it is undoubtedly the embryonic form of recursive self-improvement. Other self-reinforcing loops have also begun to operate.
The creation of immense economic value has initiated a flywheel of accelerated infrastructure expansion to support these increasingly powerful AI systems.
And robots building other robots is no longer distant; in some sense, this also includes data centers building other data centers.
If we have to build the first million humanoid robots using traditional methods, but then they can operate the entire supply chain—mining and refining minerals, driving trucks, operating factories, etc.—to build more robots, and these robots can then build more chip manufacturing facilities, data centers, etc., then the pace of progress will obviously be vastly different.
As data center production becomes automated, the cost of intelligence should eventually approach the cost of electricity.
(People often wonder about the energy consumption of each ChatGPT query: on average, each query consumes about 0.34 watt-hours, equivalent to a toaster running for just over a second, or an energy-efficient light bulb operating for a few minutes. It also consumes about 0.000085 gallons of water, roughly one-fifteenth of a teaspoon.)
Approaching AGI Endpoint, AI Miracles Become Commonplace
Technological progress will continue to accelerate, and humanity will adapt to all changes as always.
There will certainly be growing pains, such as the overall disappearance of jobs in certain industries, but on the other hand, the growth of global wealth will be so rapid that we can seriously discuss new policy ideas that were previously unimaginable.
Perhaps we won't instantly establish a new social contract, but looking back decades from now, quantitative changes will eventually lead to qualitative changes.
Drawing on history, we will find new goals and new needs, and quickly absorb new tools. Expectations will rise, but capabilities will improve just as quickly, and everyone's quality of life will get better.
The changes in occupations after the Industrial Revolution are a good example.
Humans will create increasingly wonderful things for each other.
Compared to AI, humans possess a long-term, important, and unique advantage: we are inherently concerned with others and their thoughts and actions, whereas we care less about machines.
A farmer a thousand years ago, seeing the work of many people today, would say it's simply "fake work," believing modern people are just well-fed, indulgent, and entertaining themselves with games.
A thousand years from now, when people look at the work of our era, I truly hope they will also find it very "fake," but I have no doubt that those engaged in it will find it extremely important and satisfying.
New miracles are emerging at an astonishing speed.
Today, we can't even imagine what new discoveries 2035 will bring—
Perhaps this year we solve the mystery of high-energy physics, and next year we begin space colonization; or perhaps this year we achieve a major breakthrough in materials science, and next year we have truly high-bandwidth brain-computer interfaces.
Many will choose to live as they always have, but at least some may decide to "interface."
Looking to the future, all of this seems incredible. But when we are in the midst of it, we may be shocked, yet still cope gracefully.
From a relative perspective, the singularity unfolds gradually, and integration proceeds slowly.
We are climbing a long curve of exponential technological growth: looking forward, it seems too steep to climb; looking back, it appears deceptively flat. But it is always a smooth curve.
(Imagine how incredible it would sound if, in 2020, someone told you that AGI-like products would appear by 2025, and then compare that to the actual journey of the past five years.)
Superintelligent Brain, a Long Way to Go
Great opportunities also come with severe challenges.
We certainly need to address safety issues on both technological and societal levels, but given its enormous economic impact, widespread access to superintelligence is also crucial.
The optimal path might be this:
First, solve the alignment problem. This means ensuring with strong guarantees that AI systems will learn and act towards the true goals that humanity collectively pursues in the long run.
Social media feeds are an example of misaligned AI: the underlying algorithms are incredibly good at making you keep scrolling, and they clearly understand your short-term preferences, but they do so by exploiting a mechanism in your brain that overrides your long-term preferences.
Then, focus on making superintelligence cheap, accessible, and not overly concentrated in any single individual, company, or nation.
Society is resilient, creative, and highly adaptable. If we can harness humanity's collective will and wisdom, then despite making many mistakes and some situations becoming very bad, we will ultimately learn and adapt quickly, leveraging this technology to maximize benefits and minimize risks.
Within the broad boundaries that society must define, giving users ample freedom seems crucial. The sooner the world begins discussing these boundaries and defining collective alignment, the better.
The entire industry, not just OpenAI, is building a brain for the world.
It will be highly personalized and easily accessible to everyone; our only bottleneck will be good ideas.
For a long time, tech people in the startup world always mocked "idea people"—those who had ideas but needed to find a team to implement them. But in Altman's own view, these people are about to have their moment.
Today, OpenAI plays multiple roles, but its primary identity is a superintelligence research company.
The road ahead is long and challenging, but most of the path has been illuminated, and unknown areas are rapidly disappearing. We feel privileged to be engaged in work we love.
"Intelligence so cheap it's immeasurable" is within reach. This might sound crazy, but if we had told you in 2020 what the world would be like today, it would probably have sounded even crazier than our current predictions for 2030.
May we smoothly, exponentially, and uneventfully move towards the era of superintelligence.
References:
https://blog.samaltman.com/the-gentle-singularity
https://x.com/sama/status/1932547247243505924