The Dilemma of non-linear Disruption

Non-linear disruptions have been predicted a lot – successfully, mostly in hindsight and I would like to draw parallels here with great transitions to steam, to electricity, to computers and information technology. It is a common tendency to perceive past disruptions as having been more predictable than they were. The question we are now facing as a civilization is – are we going to face such a non-linear disruption with the advancements in AI and if so, where is it going to take us – to a cleaner / greener environment by building greater futures or to self-inflicted domination by some super intelligent AI.

So, in short this is all about AI or Machine Learning or about applied statistics and us as a child-like civilization looking at the unfolding story of Machine Intelligence in awe and wonder.?

The speed of evolution is exponential and any written statement left for a day can already be outdated. The diagrams should therefore be taken as indicative and representative, the message is the key and not the representation. I also need to acknowledge all the sources explicitly or implicitly mentioned – especially the rubber band analogy – the potential disruption which is so fundamental and disruptive that the human mind feels more comfortable adopting a status quo perspective.

Looking at the applicability and current application of ML / Generative AI (GAI) there is an interesting phenomenon – people who have a thorough understanding of how ML / GAI works are becoming increasingly worried – they take the learnings from the first wave of AI and the impact on society and extrapolate that even the current “simple” second wave LLMs can drive that race for supremacy. Greed and FOMO seems to be driving the valuations of AI/ML companies (even if only remotely or by name associated to the topic). And the applications go beyond alluring teens to stay longer on their favorite social media site – we are now looking at military use cases or large-scale political persuasion, three seconds of voice being enough for mimicking your speech patterns and video personalization using deep fake technologies.

Is AI ready to make unsupervised decisions? If we look at early ML challenges - it took quite a bit of time for machines to approach, meet and beat human performance, from decades initially to two / three years and some of those “tests” are still elusive – by design probably - if we look at autonomous vehicles. While fully autonomous passenger driving will pass the test only with a loss function of 0.00001%, some existing systems have clearly passed the skills of everyday drivers … and for autonomous search and to destroy a 50% success rate is seen as sufficient already. But looking at the unfolding larger picture, we see tests being passed faster than the design is peer reviewed and only weeks or days later human benchmarks are being broken. ChatGPT has clearly established language superiority over most of us without breaking a sweat – only last week an article was published on ChatGPT scoring 155 on the WAIS-III standardized verbal IQ subtest – That’s higher than 99.9% of all humans ever tested in the standard sample (and had it not tried to show off it might have scored even better!).

As experts admit that we do not actually know what is “inside” those LLMs and how Theory of Mind is evolving with each new LLM generation, we are also observing there is neither an indication how this cognitive ability was deliberately engineered into these models, nor research demonstrating that scientists know how to achieve that. AI has now evolved, and we are noticing 4 months ago that, now a neural network has the intuitive skills of a 9-year-old. But the complexity of the topic and the potential implications for research teams (“Better than us”), seems to invite less interest than stories about ChatGPT giving marriage advice or MyAI failing to protect children.

When humans made contact with the first generation of AI we clearly strayed – the race for attention was a simple enough task and AIs perfected the same on a 1:1 level – from Facebook to Netflix to Instagram - we accepted addiction and a polarization of society. We are now seeing the same and even more dramatic challenges including the loss of online verification, missing transparency, mass persuasion, but we have also achieved a level of sophistication that helps solving scientific challenges, democratizes health care and unlock efficiency & productivity gains.?

People comparing the advent of AI with the development of nuclear technology are facing two challenges in any argument – nukes / nuclear technology was initially rushed due to perceived war time necessities but then developed more in a linear way. AI/ML has been around for some time but now suddenly is developing exponentially and secondly nukes did not improve nukes. If you look at just the brief period since ChatGPT has gotten released, we have seen the integration of external feature like WolframAlpha, search functions, mathematical models, self-supervised learning, auto verification, Code Optimization, Voice / Video Training sets, auto recursive features, H100 and other features in the wings. While on the other side not much has come in terms of safe deployment, guard rails and break the glass features.

Given the current state of affairs, AI Is Outpacing Moore’s Law with exponential growth. The cost of building an AI powered drone that could kill one enemy soldier is approaching 100USD and with 50% accuracy one soldier equals 200 USD – a step change over previous wars. AI powered tutors are expected to be 2-3 times more efficient than human tutors and are never in a bad mood. Language capabilities have not only established translation between 1000+ languages but also helped to identify subtle features of (mother) bat language, helped to develop a beebot to tell bees where to go for pesticide free nourishment, told us about dolphin dialects. AI has helped to improve the distance EVs can go and how to auto adopt changes during production. We are also seeing 90% reductions in clerical jobs, in reduction of false positives, in accident avoidance ... and a potential elimination of 300 million jobs (IMF estimate). I would also believe that the next US elections will already be won by whoever has more AI/ML firepower and deploys the better alphapersuade bot. Current guardrails do not stand up more than 3 hours, but we have also seen the significance of AI/ML in medical devices in healthcare e.g., in identifying blocked arteries.

Future is a destiny, not a predestination. Our use of AI/ML will determine how this is unfolding because – unchecked evolution has the potential to let the Gorilla prediction become true. With sensible regulation and especially with safe deployment - we can build a better future either by driving disruptive waves of elimination or focusing on blue ocean choices- building something new, sustainable, and circular.