The Mystery Ahead of Us for Mankind

Hello, I’m Ylli Bajraktari, CEO of the Special Competitive Studies Project. In this edition of 2-2-2, SCSP Chair Eric Schmidt takes stock of the generative AI ecosystem, assesses where it is going and what that will mean, and offers ideas for what we can do to prepare for the Age of AI, including a new task force that SCSP will be convening on this topic.

We have entered a new age. The Age of AI  is here, and it will profoundly shape the future of humanity. We are now seeing intelligent systems perform tasks we once thought only humans could do, such as comprehending languages, making complex decisions, and generating elaborate images from simple inputs. Truly astounding breakthroughs are being made in science with AI applications able to see patterns and structures that we humans don't see. These innovations are revolutionizing industries, from healthcare and finance to education and entertainment, enabling entirely new applications and increasing the ability of humans to execute complicated tasks with greater speed and efficiency. 

In my 50 years of working with and involvement in computer applications I have never seen adoption rates as fast as what we are witnessing today with the latest large language models (LLMs). How fast? Well, it took two months for ChatGPT to reach 100 million users. It took Gmail five years to reach that number. And if you haven’t spent time with these applications, I recommend you try out GPT-4 or Bard to gain an understanding of their truly transformative nature. The speed of adoption and the pace of improvements we’re seeing in these models is why I believe AI will soon be integrated into every business and every industry. 

I’m also convinced that this new AI era will turn scientific research on its head. Take mathematics for example. Today, we have models generating math conjectures that challenge the human ability to prove them out. When both the problems and the proofs happen automatically, computer systems are entirely likely to suggest extensions to STEM fields in ways that are beyond human comprehension. 

Such scenarios will lead to new questions that we have yet to consider. How will humanity view systems that we use that are known to be true but that are not fully understandable by humans? What happens when this work self-extends in the sense that the system is learning and develops new capabilities while it's being used? We are on the brink of new systems that we don’t understand, depend upon, and that are changing in ways we can’t understand and can’t predict.

This image was generated with the assistance of AI.

---

What does the future look like?

One thing I am certain of is that the future will look very different than it does right now. While hard to predict precisely how these systems will evolve, we can extrapolate from observable trends and recent breakthroughs. I believe that in fairly short order we will see the following developments: 

1. The emergence of a small number of incredibly intelligent LLM systems that are very powerful and therefore – and rightfully so – heavily managed and regulated. Simultaneously, there will emerge a much larger number of more specialized systems deploying LLMs that are used everyday in all our digital systems. 

2. These more specialized systems will become ubiquitous in the devices we own from the humble toaster to powerful algorithms in our cell phones, cars, and homes. Chips with smaller versions of the same technology may be embedded in commonplace hardware that perform specific tasks. Just look around your house and think about all the digital devices and imagine each of them has one of these chips in it. Then imagine that it can deeply understand the meanings and nuances conveyed by human language. This is coming very rapidly. 

3. LLMs currently under development will soon have both rapid training abilities as well as memory, which they currently lack – they can generate an answer but then they forget that answer. Once true memory is invented, the system should be able to choose alternative futures for itself by imagining a different future from the past. At that point, these systems will begin to approximate artificial general intelligence. 

---

With promise comes potential peril

This future, however, is not without serious concerns. Even among those of us in the field, the diffusion of this new technology is happening so fast that it can be hard to stay up with the changes, let alone how it is being used. To highlight a few of the emerging risks: We understand that there are significant biases in humans and that these systems reflect the data that they're trained with. Also, with so few numbers of systems able to complete this kind of work, there could be a concentration of power and influence in the hands of a few AI developers and corporations. On the other hand, as open-source variations of these LLMs proliferate, we could see rogue actors using them to generate persuasive disinformation at machine speed and scale, or to create a new deadly virus. LLMs could be used to enhance cyber threats and exploit human vulnerabilities to steal data and identities.  

And then there are the very real concerns about how nations like China might employ LLMs. We can be certain that China understands this space and has programs that are similar to those that have been developed and released in the U.S. They won't be able to release them the way ChatGPT was released because China can't allow its population to have open access to unfettered information. But the national security implications and surveillance issues that arise from China’s development of powerful LLMs are profound. 

---

So what can we do next to be helpful?

There is a real possibility that this technology will move faster and have a bigger impact than our political, regulatory, and moral systems can sufficiently understand and effectively manage. While these systems may exhibit great self-confidence, they are not always right and are unlikely to be so for some time, if ever. Some have observed that commercial competition and rapid scaling risks producing a “race to the bottom” where the competition for new features will cause safety mechanisms and guardrails to be short changed. Already, LLMs are attracting considerable scrutiny and as people imagine the potential impact of future iterations they will move to regulate them. We’ve already seen the Italians ban ChatGPT for violating GDPR rules. That is likely just the beginning. 

That means that those of us who work in and are otherwise involved in this industry need to be: a) a source of truth of what is known, likely, and unlikely about the future of LLMs as well as associated risks; b) a reasonably good predictor of the time frames for future developments, particularly potential threats; and c) a communicator of constructive recommendations to society on how to accentuate the benefits and mitigate the risks of LLMs.

So I propose the following formulation: we comprise an agreed upon “short term” list of recommendations that we think our governments should take action on immediately and also a “longer term” list of recommendations to explore and debate. 

---

What can we do now:

1. Build an AI tutor and an AI doctor, so that learning and health can be more broadly available, worldwide. AI’s potential impact in these areas is enormous and will demonstrate the value of this incredible new technology.

This image was generated with the assistance of AI.

2. Require LLMs’ design details and methodologies be published, and that they include mechanisms to upgrade them and remediate any potential harmful impact. 

3. Propose easy to understand requirements to make social media less susceptible to aspects of AI like deep fakes and misinformation. These would include:  

   • Requiring that the the system knows who is on their platform (is it a person, or a bot, etc); 

   • Establishing provenance where the system knows the origin of the content being served, including if has been altered by using cryptography and watermarks;

   • Publishing the methodologies used by the platform and be held to those platforms for amplification; and 

   • Having an AI system that helps the user understand if the information in front of them is likely true or false.

---

What should we consider in the long term:

1. Emergent behavior

   There is evidence that at certain larger scales, new capabilities can emerge. How will we know that we have sufficiently explored the capability overhang in these models and how long is testing required before release? How do we have general, adaptable systems for safety and for behaviors we have not yet seen? 

2. Proliferation

   This will be a core problem with open source LLMs. It's clear that the very large models will not be publicly released. How do universities, researchers, and others do research without them? How do we constrain an open source model, for example by embedding a constitution within the training data? What does the “safety board” or “AI ethics committee” look like for open source models? Assuming diffusion of this knowledge, how do we prevent a rogue group developing a fully open source model with fully released weights with no constraints?

3. Low probability but very harmful events

   Develop a list of what are the potential society-level impacts that could prove harmful, and a theory of how to mitigate them (for example, synthetic bioweapons).

---

SCSP’s next steps

As the Age of AI continues to reshape our world, it's crucial that we understand the profound implications of LLMs and other generative AI systems on society, business, and individuals. To that end, we have created a task force under SCSP, co-chaired by Andrew Moore and Tom Mitchell, to assess the potential benefits and drawbacks of these powerful tools. Over the next few months, the task force will convene a diverse group of experts — from AI researchers and technologists to ethicists and policymakers — and develop a framework that fosters responsible innovation and harnesses the transformative power of generative AI while addressing the ethical concerns and potential risks. Together, we will strive to shape a future in which AI serves as a force for good, empowering us, and driving meaningful progress. Stay tuned for more on this in the months to come, and please reach out to the SCSP team if you would like to help.