Steve Oberlin: “I’m most excited about the hybridisation of HPC and AI for scientific applications”

Luck, mostly. People underestimate the role chance plays in their lives. I was on one path, had an unlikely dinner with a friend who worked for NVIDIA, then I was on a different path.

I’ve been incredibly lucky in my career from the start. I got to meet and work with Seymour Cray, the father of supercomputing, help design awesome vector machines and architect Cray’s MPP supercomputers. Now I get to hang around with more smart people using GPU-accelerated computing to power an even more disruptive and impactful revolution happening in HPC [high-performance computing] and AI [artificial intelligence].

AI, of course. Call it machine learning, deep learning, or whatever, if you’re a modern human with a smartphone, AI is already touching your life. And we’re just at the beginning.

One thing we know is true: Scale – more data, bigger models, faster compute – is a primary factor in achieving state-of-the-art results in AI. Pressure to scale AI-capable system infrastructure is enormous, on the hardware and software sides.

I’m most excited about the hybridisation of HPC + AI for scientific applications, like weather and climate modeling, or simulating material physics and chemistry. Typical numeric scientific applications compute a mathematical model of the phenomena being simulated.

Those models are frequently so computationally expensive that scientists have to substitute heuristics, contrived approximations that trade accuracy for speed, for parts of the model just to achieve a simulation that can be run at practical scale and resolution.

Turns out artificial neural networks are universal approximators, so one idea is to use the accurate but slow numeric code to generate the training data for an AI approximator that might then be deployed at scale, be more efficient and possibly even more accurate. The gains can sound crazy, like the quantum chemistry neural net that’s 5-6 orders of magnitude faster than the numeric code.

Look for the economic engine. If there isn't one, or it’s implausible, it’s probably hype. We know the current AI boom largely isn’t hype because it’s already made billions for large-scale internet companies by using AI trained on our data to provide more valuable services back to us, and that value is continuing to expand in capability and across applications and markets at a rapid pace.

“Keep It Simple, Stupid?” “Ignore sunk cost?” I guess I’m bad at curating advice.

I read a lot and listen to a number of long-form technical, economic, and science-based podcasts.

Best general-purpose advice I distill is to not trust your own senses or memory. Anybody can be fooled and you’re the person you should distrust the most. That kind of humility and caution is the basis of the scientific method. It’s so easy to be wrong.

Reading a lot of science fiction as a kid, probably, and getting exposed to tools from a young age, taking things apart, occasionally putting them back together again.

I have a strong urge to make things. It might be genetic, since my father was a carpenter/metal worker/mechanic/electrician – whatever the job called for.

When I’m not traveling, I work from home. Lots of phone time and video conference meetings, lots of email. I read, I write. My effectiveness influencing technical direction is dependent on my ability to communicate and persuade.

I make infrequent use of carefully-curated Twitter, mostly following science and technical people who pass on links to new papers or interesting tech, and use LinkedIn as a Rolodex, but am otherwise pretty thin on social media.

Windsurfing, badminton, play with our cats. Play with 3D computer graphics programs. Make things. Read.

Actually, I’m a pretty relaxed guy most of the time, even without those activities...

Seymour Cray is still my role model. Seymour was a true Rennaissance Man, able to innovate in any discipline necessary to create supercomputers, including mechanical design, power and cooling, software, circuit and logic design, and of course system architecture. He was phenomenally productive, creatively.

These days, I’m also admiring tech billionaires who are using their fortunes to advance humanity, especially Elon Musk. Tesla has changed the course of automotive technology and will probably be first to produce fully-autonomous vehicles anyone would trust. SpaceX absolutely blows my mind.

I think it’s obvious that the biggest challenge humanity faces with a technological solution is limiting and mitigating climate change. The technology aspect isn’t actually that difficult, if you don’t count fusion as a necessary part of the solution, but it’s a big planet and there are a lot of people.

There’s a lot of momentum to the system, we’re very late taking action, and it requires global cooperation in the solution. I’m afraid we’re going to lose a lot of species and many of our most populous cities are going to be dramatically transformed.

If only climage change were merely a technological challenge I would not be concerned, but it’s really a political problem and a market problem and politicians and economists seem way less capable of solving problems than engineers and scientists.