Advancing Toward the Third Generation of AI

Industry pioneer Tom Kehler shares insights on the evolution of AI, the promise of collective intelligence, and his optimistic vision for the future

---

Can you tell us who you are and what’s your take on both Generative AI like Midjourney or Stable Diffusion and current LLMs such as GPT-4 in a nutshell ?

Tom Kehler: My name is Tom Kehler. I have been involved in natural language processing and AI since the late 60s and early 70s. I published my first computational linguistics paper in 1976. I left academia and joined the commercial AI world in 1980 as a leader in Texas Instruments AI group—enough of history.

The paper ‘Attention is all you need‘ and transformer models changed the world of natural language understanding. My academic journey is rooted in applied physics, the magnetic behavior of materials. Transformer models linked the power of statistical physics to NLU by linking language modeling to the immense power of the mathematics of hyperdimensional vector spaces. 

In 2017, I was working on a topic-learning problem, and the advent of transformer models immediately changed everything. It was a major breakthrough, and we started using the early encoder models early on. 

GPT-4 is amazing! However, with all my enthusiasm, I am more aligned with guys like Hinton, who suggests we may be getting ahead of ourselves in commercialization.

In this series, we meet interesting founders and creators from the AI scene and discuss with them not only their take on this new era of technology, but maybe learn a few secret tricks from them. If you’d like to share your story and tips, you can get in touch with us here.

Part of my former life was about creating and facilitating workshops, so when I read that your current venture, “Crowdsmart” is about helping collective human intelligence with AI, it rings a bell. Many of my facilitators and innovators Linkedin friends are exploring this topic, what kind of results really impressed you when it comes to AI + human collaboration to tackle problems?

Humans are the best at imagination and creativity. Judea Pearl clarifies this in his piece “The Book of Why”: “You are smarter than your data”. Human collective intelligence has produced all the scientific results we see to date – including generative AI. 

For that reason, I immediately thought we should apply these generative models to magnify human creativity and human collective intelligence. The source of data for CrowdSmart generative modeling technology is humans working together actively, making predictions, and solving problems.

We are early in the process, but an area that is one of the biggest opportunities is ‘seeing ahead’ – market intelligence. Collective intelligence science shows that a cognitively diverse group of humans can outperform any individual expert in ‘seeing ahead’ – predicting outcomes.

Combine that observation with applying generative AI techniques, and you can see a path to a new world that pushes aside all that we have done before with focus groups, design thinking sessions, surveys, etc. Collective intelligence and generative AI have the greatest power for us to see around the corner and co-create like never before.

We often hear that once the AI will be more reliable (as in won’t do as many mistakes as now) we will have nothing left to do, and we might even become a ” stupid specie ” like in “Wall-E” for example. We also hear that Wisdom might become the main purpose of humans, to be able to leverage AI for the greater good. How do you picture the impact of AI on societies in a 20 or 30 years time frame?

I am extremely optimistic about the future. Those who have hitched their wagons to the current technology, as good as it is, are forced into a mindset of an automaton – and that future is dismal. Those who hitch their wagon to a future based on human collective intelligence and, in general, an AI that links to natural laws will see a brighter future. Let me unpack that a bit. Those who have followed AI, know that the Defense Advanced Research Projects Agency (DARPA) has played a key role in its development.

DARPA’s vision of a third generation of AI is an AI that incorporates, context, can transparently explain, and can seamlessly integrate with human thinking. That is the AI that leads to a brighter future.

My first company in AI, IntelliCorp, had funding from DARPA in the area of knowledge representation and reasoning. DARPA’s bird’s eye view of AI states that we are in the second wave, second generation of AI. The first was hand-coded AI, where we used symbolic AI to mimic the reasoning and knowledge representation activities of humans.

The second wave of AI is ‘statistical classifiers’ using the mathematics of statistical modeling (largely from the statistical mechanics in physics) to learn intelligence embedded in data. This is largely where we are right now with ‘statistical’ generative AI. It generates plausible concepts from existing data. 

That is very cool but it stops short. DARPA’s vision of a third generation of AI is an AI that incorporates, context, can transparently explain, and can seamlessly integrate with human thinking. That is the AI that leads to a brighter future. There is not time to go into it in this interview but this AI links to neuroscience and its pioneer is Karl Friston.

This is the form of AI we use at CrowdSmart because it links to a model of the human brain and is grounded in the intelligence of living systems (the physics of living systems).

You’ve got 50 years of experience in the AI field and you even hold several patents related to AI, We would like to know, what do you think about the new model MEta is working on, based on Yann LeCun vision: “the Image Joint Embedding Predictive Architecture (I-JEPA), who learns by creating an internal model of the outside world, which compares abstract representations of images (rather than comparing the pixels themselves)”., is it what will allow AI to be much more innovative and creative?

LeCun’s vision is along the lines of what I suggested in my prior answer. I think we will make the best traction on moving from pixels to concept formation by working with neuroscience-based models where we can link to brain imaging and the immense amount of work going on in understanding how the human brain works. 

Maybe we should respect nature and do a bit more studying before we think we have nailed a model for general intelligence.

I heard recently that training GPT-4 costs about $65 million dollars and takes an immense amount of power. The human brain operates on the power of 35 watts. Maybe we should respect nature and do a bit more studying before we think we have nailed a model for general intelligence. 

What do you think is the next big thing that is going to happen in the next couple of weeks/months, AI-tech related?

It is hard to contain the level of excitement I feel about the future of co-creativity with humans and AI working together. Linking the power of human collective intelligence with AI will change everything. We are just at the beginning of a whole new world of co-creativity – this will have its biggest impact I believe in how we do scientific research.

With your experience using the tools, you probably discovered a couple of tips and tricks. Which ones would you be ok to share with our audience?

This is a hard one. Any and all tools that help us find common ground and learn together are of interest to me. One of the biggest opportunities of generative AI in its current form is coding. Coding is a means of creative expression that is unequaled. Generative AI is opening that door for everyone through things like prompt engineering.  Experiment and play. 

Is there anything else you want to share with our audience?

Keep an eye out for an initiative called “Common Good AI”. We are forming a new organization that aims to create an AI future that integrates with natural laws – the laws of living physics – and is aimed at using AI to tackle our greatest challenges.

Where can people find out more about you?

You can find me on LinkedIn and follow me on Towards Data Science and Medium. This paper picks up on what was discussed here.