Over the last couple of months I've invested a healthy amount of time in engaging with US representatives working at the intersection of academia,industry and policy. Earlier in May the US Dept of commerce organised a “Select USA campaign” that was meant to position the USA as the leading destination for investment and I had an opportunity to connect with several Foreign service officers in Delhi promoting USA to Indian entrepreneurs. Before that, last year in December, I met online with researchers at Princeton Plasma Physics Laboratory to understand how they are using AI to accelerate fusion research. My engagement continued with several non profit organisations like American society of Metals ,Society for automotive engineers and American institute of physics —plasma division and it enabled me to get the big picture view of the US research landscape. Which is why I was really excited to attend the inaugural genesis mission summit held last week to see what announcements would be made.
As I expected, AI took the center stage. This was not surprising because when president Trump signed the executive order launching the initiative it was done with the express purpose of creating a centralized system to accelerate scientific output of the USA.
Clearly the US administration, along with the extremely wealthy information technology lobby in Silicon Valley, believes that AI has created a new computing paradigm that has the potential to change how science is done. On the face of it this is a good decision because it reroutes a portion of the capital back to universities and research institutions.
This is obviously good from the perspective of science. It will make it possible for several bright young minds in the USA to make a career out of research. Politically this is an excellent decision because President Trump has been facing a lot of criticism regarding “cutting the funding from institutions" but this nearly $ 5 BILLION announcement proves that funding is actually being streamlined rather than frozen.
What I find really exciting about this initiative is that it is being led by the Department of Energy. Secretary Wright has recognised that it is important for the US to maintain its position as an energy leader. A good chunk of resources have been allocated to grid modernisation, nuclear energy capability building and building tools to advance US manufacturing capability that can allow it to harness energy from renewable sources on a massive scale.
While Secretary Wright recognised the importance of AI he also understood that necessity to strengthen the infrastructure necessary to make AI work. Energy was one part of that equation,materials were another. He emphasized the importance of securing critical minerals, and continuing the material discovery process that can enable new technologies. Equal emphasis was placed on improving construction processes,water management and on life sciences.
So while AI was the highlight of the event it's clear that the genesis mission is setting a more ambitious goal. And the fact that the goal is coming from the department of energy makes me certain that the US administration has got its innovation loop right. Invest in tech that allows you to build even more technology.
I do believe that the focus on AI is a bit exaggerated. AI does help continuing education, building skills and in facilitating cross pollination of ideas. People who don't have a background in STEM can get good at science with dedicated work and start contributing with their core competencies in policy,economics ,business and help in diffusion of technology or even breakthrough discoveries.
The focus on simulation of experiments is also not out of place. Science had already become computational long back. Platforms like MATLAB allowed researchers to perform virtual experiments in certain highly calculation intensive fields. AI is a natural evolution of that.
But it has its limits. No matter how much computing power you throw at it, AI is not going to get rid of geographic mineral imbalances in the world. New Li deposits are not going to pop out to make battery problems go away. These problems can't be solved by brute force but need to be sidestepped by building new systems. I have my doubts if these can be outsourced to AI because creating a new body of knowledge is not an optimisation problem. The means to achieve those objectives are not well defined and often require backtracking and branching. If AI does manage to achieve that it would be at best as efficient as a normal human. That discovery process will be just as long-winded as if it were done by humans.
I also don't really believe that novel research can be institutionalized. Often the example of ARPA is given as an agency that bought us the internet among other wondrous new inventions. But nearly all of them including the internet were evolutions of design that were built by painstaking hard trial and error processes. The internet is an evolved telephone network and it exists because theorists like Maxwell,experimentalists like Bell ,entrepreneurs like West and Bright together with private corporations like Atlantic Telegraph company were willing to take massive risks that would have sounded utterly foolish to everyone else.
Every invention can be traced back to a handful of people who were willing to believe in it and see it through multiple failures.
I have said this before and the time that has elapsed since then has only made me more confident that intelligence is largely overrated. Inspiring moments and intuition do open up new sources of knowledge to us but every time we acquire new information it fits naturally with truths that were established previously. Through action we refine our understanding and build new tools,give structure to new ideas that continue to enrich our lives. If AI ever reaches the point where it starts to make discoveries it would look very human-like. What might change is that you could have machines using electrical energy instead of food to give rise to that discovery but it would cost just about the same in terms of energy.
On the materials side AI promises to help discover new materials and new processes. Which could be useful but I'd be really cautious in estimating its overall impact. Abundance of materials have an overwhelming impact on the technologies that develop around it. Higher strength and toughness could be reached with exotic alloys systems but they would have to compete with moderate strength but vastly more abundant systems.
Advanced manufacturing too is very limited. As an example compare the promises of additive manufacture the hyperbole around it that promised to revolutionize manufacturing to what it has actually delivered.
This is the crux of the debate. Concentrated high power vs diffused low power. The industry as it currently stands is built upon centralised systems. Power generated at one spot and then distributed. Servers at one location sending information to billions of people. One steel plant producing material for the entire state. However it is also possible to distribute both energy and materials production, this would make the overall system more robust though it would require new techniques.
It is also what the genesis mission is doing with its super-computer strategy. It's creating massive compute power to simulate complex physical problems. This means smaller transistors on the device level and more power hungry servers on the system level. But if we look at the computing history every new invention promised scientific breakthroughs. Vacuum tubes did that,transistors did that,big data and cloud computing did that and now AI is doing that. Did it lead to a breakthrough? I don't know because we are still struggling to solve equations like Navier stokes built over a century ago. Maybe a server farm could do that but how would engineers benefit from it? Would they be expected to run off to a supercomputer to determine the outcomes of every small change? Paradoxically, experiments are cheaper and faster!
In science, materials changes are actual real computation. The materials under experiment are physically computing , or perhaps we should say changing their state, according to the laws that govern their behaviour. When you heat a slab of metal heat transfer occurs according to the heat equation. This is determinable via sensors. What goes on inside the metal, the response of the electrons and crystal lattices is real computation. It's the same in all scientific fields,chemistry,biology doesn't matter what field you choose. Material response is a common underlying theme that manifests change that is observed and controlled. Simulation can predict this because real computation is happening at the material level.
The discovery of figuring out the governing law is a one time process. Newton's laws are discovered. They will predict outcomes according to their model of the universe forever.
Supercomputing alternatives and more powerful AI systems would definitely make those predictions more accurate. And it pushes computer science to its limits for sure. Those who have the means and the ability to do it should go for it. But it's not a necessity for either scientific advancement or material production. Mathematical tools exist that allow probabilistic predictions of complex equations and don't even require a computer. It also allows us to build different kinds of computers that are not as powerful but still good enough probabilistic simulators relaxing the feature size limits.
This does not mean that I'm against simulations. I find them very useful and necessary during the discovery and formulation stage. Simulation should be reasonably accurate that it gives enough confidence to actually do a real experiment. If even models can't work then no way a real experiment would. But again it does not have to be precise, only accurate enough that it allows us to proceed with experimentation and flexible enough to allow feedback of those results to the original formulation and increase accuracy.
The fact that AI has been useful in learning is really more of a critique on how learning systems are currently designed than any inherent advantage of AI itself. Because the material is still the same only how it's approached and acquired changes.
Geopolitics is definitely an angle here as is nearly every new technology these days. AI is now a ‘national security’ issue as well. Geopolitically speaking there is no way that any nation can catch up or level with those who have built this tech. The only way to achieve parity is to make it irrelevant. But that's easier said than done.
Comments
Post a Comment