AI's Next Leg, and Watching the Clouds Rise
Somehow we have reached the last of the twelve columns I set out to write. Over half a year they have wandered from software, AI valuations, data centres and jobs to academia, mathematics and the kampung.
When I first pitched this series to Sin Chew, it came out of a very plain feeling. ChatGPT arrived at the end of 2022, but it was only late last year, when a new generation of models crossed some threshold in writing code and handling data, that it became obvious these chatbots could actually do people's work. The coverage at the time swung between rapture and doom. Having spent years working where AI meets science, I wondered whether I could be a little calmer and a little more quantitative, and simply set out what I was seeing.
The first column, written at the end of March, was about AI upending the software industry. I described myself then as a dove, someone who doubted AI would turn the world upside down any time soon. Little more than half a year later it has reached into the work I care about most, scientific research. A few columns ago I wrote about OpenAI's claim that its model had solved one version of the Navier-Stokes problem. In early October it released 722 mathematics papers written by an internal model, to applause from some mathematicians and doubts from others about whether anyone had checked them properly. Even a dove has had to recalibrate, more than once.
The upside of being an uncle
Somewhere along the way I became an uncle. The one advantage is that I can still remember the internet arriving. The grown-ups worried then too, that it would lead children astray, that people would stop reading. Looking back, we learned to live with it.
So for months I have been trying to gauge how big AI really is. To call it just another internet probably undersells it. The internet changed how we look things up. AI changes who does the work. But to compare it with the steam engine, or with going from no electricity to electricity, oversells it.
What the internet left behind were things like Google search and the smartphone. One put knowledge a search away. The other put a computer in every pocket, and along the way rewrote how we hail a ride and how we pay for it. AI so far has given digital work a fivefold or tenfold speed-up, but mostly by reworking industries that already existed. It has yet to grow an industry that is genuinely its own. And the impact is still largely confined to screens. As I wrote when discussing superintelligence, AI has some way to go once it has to act in the physical world. Teach a robot to plug in a socket and its engineers want to run laps of the field.
On speed alone, though, the internet cannot compete. In my own research, with AI behind us, we have started taking on problems we would not have dared to attempt before. New things are beginning to sprout, too. In August, a melanoma vaccine made to measure for each patient's tumour, with machine learning choosing what to target, reported a positive result in a Phase 3 trial of more than a thousand people. Whether science and medicine will change paradigm as a result, nobody can yet say.
What has warmed me to the technology most is something small. I set up ChatGPT for my mother, and when I am not around she can ask it anything, down to how to feed her plants. Older people a generation ago had nothing like it. She has, of course, also caught it talking nonsense.
Learning is the same. For introductory material there is almost nothing AI cannot teach you, provided you want to learn. What once meant paying for tuition and buying a shelf of revision guides is now a question away. A hundred ringgit a month will not cover heavy work, but for learning it is a resource that people of modest means could once only dream of.
Used well, it could genuinely make the world a little fairer. Just as nobody today can imagine life without electricity, a future that cannot do without AI need not be the apocalypse of the headlines. On that, I remain an optimist.
What still worries me
Astronomers are particularly alert to what technology costs. A few years ago I wrote in this column about light pollution, how the electric light brought brightness and quietly took the Milky Way from the skies above our cities. AI is no different. "Used well" is easy to say. It needs AI companies, governments and the rest of society all pulling in the same direction.
The first to lose out are usually those with the least to begin with. In the age of AI, the gaps between rich and poor, and between countries, can easily become winner-takes-all. The kampung in my last column is one example.
The young take the hit early as well. Entry-level jobs are squeezed first, and research in the United States has found that in the jobs AI can most readily replace, employment among people just starting out has fallen noticeably. Several young people I know with mathematical ambitions have begun to waver, reasoning that AI will solve everything anyway. But a mathematician's value was never only in proving theorems. It lies in the judgement that generations develop by walking the paths themselves. As I put it a few columns ago, someone dropped on a summit by helicopter cannot draw the map. My students are the first cohort to compete with AI from their first day in the field, and what I can teach them is less and less about technique and more and more about judging which questions are worth asking.
Then there are our own minds. Outsource your thinking wholesale and it will rust. After the 722 papers came out, an advisory group on mathematics and AI at the Institute for Advanced Study in Princeton, where I once worked, put it well. The release, they wrote, is the beginning, not the completion, of human understanding. Essays, songs and pictures can now be produced at a click, and a screen full of work that is hard to fault only makes the ability to tell good from bad scarcer.
And there is attention. One of the directions people are most excited about is video and games generated in real time. Google has already let its top-tier subscribers step into virtual worlds that AI builds around them as they walk. The internet has left plenty of people curled up in their comfort zones. AI will only make those zones more comfortable, and children will find them hardest to resist.
Further out is society as a whole. Used well, AI is genuinely economical. Hand it the jobs that need no thought and it often uses fewer resources than a person would, and frees that person for something else. But if vast amounts of computing are burned on addictive entertainment, then the data centres that may take around 30% of Peninsular Malaysia's electricity by 2035 are burning what they should not. Communities also become easier to split. Deepfake videos are already doing the rounds on WhatsApp, and the next general election will be a test. When a video seems a little too convenient, ask where it came from and who stands to gain.
Finally, safety. Several frontier AI companies are heading in the same direction, getting AI to help with AI research. The industry calls it recursive self-improvement, AI improving itself. Some already say their AI can handle, under human direction, research tasks that would take a skilled researcher several days, and they want AI working as a researcher in its own right within a few years. Yet even the lab furthest ahead admits it does not yet know how to make that journey safely. The risk is real. Before believing the more lurid headlines, though, it is worth asking how much commercial interest sits behind them.
As I wrote in the column asking whether AI was a bubble, what people do not understand is where the premium tends to hide. The millennium bug was once billed as the end of the world, too, and passed without incident, because countless engineers had quietly fixed the problem in advance. Take it seriously, but there is no need to panic.
The world ahead will be more fragmented
You can probably tell by now that I am still an AI dove.
The world ahead will probably be more fragmented and more fluid. A great deal of work will be broken into pieces and assembled by people and AI together. A small team with AI behind it can take on a large company and even overtake it. That is an opportunity, especially for the young. The other side is that setbacks will come more often, and changing jobs, even changing fields, will become routine. Rather than panic, better to get used to the rhythm early.
The Tang poet Wang Wei has a couplet for this. "I walk to where the water ends, and sit to watch the clouds rise." When there seems to be no way forward, you might as well sit down for a while and watch the clouds come up.
Think of this afterword as a few questions left for myself a year or two from now. Is the curve from my first column, doubling every few months, still holding? Has AI grown its own Google? Has the first AI-discovered drug reached the market? Is it any easier for new graduates to find work? Can children in the kampung get hold of decent AI? And those windowless grey buildings in Johor, have they filled up?
If I have called it wrong, never mind. Anyone who tells you with total confidence what AI's future holds should not be wholly believed, me included.
The world will get better, perhaps even fairer, and there is a great deal still to do. The one constant is that it keeps changing.
A year or two from now, I will look up and see whether the clouds have risen.
(End of series)
About the author: Graduate of Chong Hwa Independent High School, Kuala Lumpur. Earned his PhD in Astrophysics from Harvard University in 2017, then held a NASA Hubble Fellowship at the Institute for Advanced Study (IAS) in Princeton. Former joint professor of Astronomy and Computer Science at the Australian National University, and currently Associate Professor of Astronomy at The Ohio State University, where his work focuses on using AI to accelerate discovery in astronomy and the fundamental sciences. Now on sabbatical in Malaysia, watching the local scene as an outsider, and welcomes letters from readers. Personal website: https://www.ysting.space