Beyond Limits: Rethinking the next generation of AI

A new AI company called Beyond Limits took an idea from space exploration and turned it into a fascinating new self-training AI. This new AI – which is already being used in finance, healthcare and petrochemical industries – represents an impressive breakthrough for certain AI implementations that require high levels of autonomy.

A human profile containing digital wireframe of technology connections.
Sadsadang / Getty Images

I had an interesting talk with AJ Abdallat, CEO of a small firm called Beyond Limits doing interesting things with AI. Their differentiator is that their AI’s decisions can be audited, and the AI itself can be edited at a granular level so corrections generally don’t require retraining. As I was listening it struck me that if we could do this with people, particularly young teenagers, top executives, criminals and politicians we could almost instantly make the world a better safer place.

Granted this approach – particularly if it was being used for commercial aircraft or self-driving cars – should have a high requirement for substantial simulation before deployment. This could not only cut years off what would typically be needed for a complex AI development project, however, but would also allow for a level of customization at scale we don’t currently seem to have in this space.

Fixing a bad brain

For some reason I’m thinking of the movie “Young Frankenstein,” when Igor picked up Abby Normal’s (Abnormal) brain. Actually fixing people’s brains has always been problematic, but since we build these AIs ourselves, we can both diagnose problems and come up with workable solutions. Those solutions often entail wiping out the data set that forms the AI’s education and reloading it from scratch – reminding me more of the movie “Total Recall.”

But the difficulty in the wipe-and-replace method is that you can introduce more problems with the new data load, so you’re constantly playing a game of Whack a Mole, worrying that the new problem you may have introduced could be worse than the one you tried to get rid of.

The process should be: identify the problem, research the cause, craft a solution, implement the solution, test the solution and repeat as necessary until the test is clean.

This is basically what Abdallat walked me through at Beyond Limits. During development or post deployment they identify a problem and forensically audit the AI to determine the cause. Using the forensic data, they craft a fix, then apply the patch and test it to assure the result.

There’s another potential paradigm here: to see if you could contain this process in the solution so the AI could reliably fix itself.

That’s part of what makes this platform interesting, and it comes from the company’s roots.

Built for space

Beyond Limits evolved out of work with NASA's Jet Propulsion Laboratory (JPL) for remote rovers used to explore places like the moon and Mars. Due to the communications lag in space, real-time control is virtually impossible. Any AI solution must be not only fully autonomous, it must be able to train and, ideally, correct itself. When there is a problem it can’t correct, the bandwidth limitations for communication make full reprograming problematic…but point patches are certainly possible.

This resulted in an AI platform uniquely able to be updated, modified and, to a certain and initially limited extent, able to both teach itself and make corrections while disconnected. This unusual requirement likely has made the resulting AI nearly ideal for areas where the AI must often act independent of oversight – and/or in areas where problems can escalate very rapidly – and the AI must be able to both deal with a diversity of known and unknown issues.

Initial tests and deployments of Beyond Limits’ AI have been in:

  • Deep water oil field exploration – to avoid issues like sanding, where there are few qualified experts, but the resulting issues can cause a catastrophic well failure
  • Refineries – mostly for control but this would likely be ideal for disaster mitigation as well
  • Financial institutions – automating traders and assuring the audit trail
  • Healthcare – data portability while better assuring privacy (this is going very slowly due to the changing privacy regulations but could eventually be ideal because of those changes)
  • Distributed IoT – implementation is similar to the space rovers and used for pipe crawlers

A new class of AI

Although still its infancy, Beyond Limits represents a new class of AI. It’s better enabled to operate fully autonomously, it can both learn on the fly and increasingly make corrections to its own programing, and it may eventually include emulation as a feature so that it can more safely self-train. Using another, and far older science fiction movie as a reference (“Forbidden Planet”), this takes us to a Robbie the Robot-level AI and far closer to the AIs we all thought we’d eventually have.

Beyond Limits is a small, young company but firms like this have historically been incredibly disruptive once they get to scale. An AI that could self-train, provide a full audit trail, allow for point patching of its training and operate independently indefinitely is the future.

It seems that with Beyond Limits, that future is closer than I thought.

Copyright © 2019 IDG Communications, Inc.

7 inconvenient truths about the hybrid work trend
Shop Tech Products at Amazon