Welcome to the Era of AI 2.0

This article was originally published on LinkedIn:  `

Progress can feel incremental, don’t you think? It’s subtle and, like the proverbial frog in slowly heating water, can go mostly unnoticed.

Did you notice?

Did you notice the moment when we entered the era of AI 2.0?

If you didn’t then it’s important you now take note, because taking full advantage and avoiding pitfalls of AI 2.0 will requires your full attention!

We must collectively recognise the incredible, pivotal moment, so we can

a) capitalise on the economic opportunity presented, and

b) exercise the caution and foresight needed to meet a new age of safety challenges.

• As the general public rise to appreciate the joy of self-managing social media profiles, auto-responding email accounts and books that read themselves (😉), the software engineers of the world have come to appreciate quite how valuable Large Language Models can be, too.

  AI 2.0 is marked distinctly by the orchestration of LLM-based agents. It is AI language models that are capable of managing, directing and modulating other AI. This not merely an incremental step. It’s a leap in artificial intelligence that redefines what is possible for both business and government.

  AI 2.0 is characterised by advancements in generative machine learning for NLP models. Recent LLMs have shown a level of reliability not seen before and have exhibited evidence of writing self-improving code [1]. Consider for a moment the impact of this statement. The implication is that we stand at the foot of a graph that is about to become exponential.

  (Or perhaps exponentially exponential? Who knows. This author humbly leaves the advanced mathematics to the research team. Let’s just say Al Gore would need a much bigger scissor lift.)

And it is the very same advanced model that does your homework and writes your emails, GPT-4, that exemplifies this change.

Human-level problem solving has thrust our species to prominence because we can consider a problem, project into various futures, and plan what is likely to be the most successful angle of attack. AI 2.0 agents are now showing these traits, too. We’re in relatively early stages, but these models are increasingly adept at planning and memory retention. They are becoming ever more capable of task decomposition and employing strategic thinking to meet broad objectives.

This means they are capable of autonomous decision-making.

The ‘loop’ function is a foundational tenet of software engineering. It enables programmes to continuously cycle through code until some condition is met. Condition met, the loop ends, the objective achieved.

A programme can have as many loops as it likes but those loops still need to be created. AI of the modern age will be considerably more useful than the previous generation of AI agents because it will always be able to plan the next phase of attack. The loop never ends.

We wrote a fun piece on the Q* rumours at Open AI, which explains enhanced reasoning and its implications in greater detail here: https://www.linkedin.com/pulse/when-computers-beat-us-our-own-game-advai-0d1xf%3F

This evolution understandably brings with it anxieties surrounding trust, reliability, and accountability. If an agent is capable of planning its next step, then who exactly is responsible for its decision? Deploying ‘black-box’ decision agents makes decision makers justifiably nervous. In our world, it doesn’t matter what the gains are if we can’t blame someone for something going wrong.

Society has collectively agreed that the buck needs to stop with someone.

How can this be? How can we both adopt the use of autonomous decision-making agents while retaining accountability – someone responsible for those decisions?

Robust assurance and strenuous testing are the methods for retaining control over their autonomous agents. The rigorous evaluation of AI 2.0 and the development and integration of safeguards are therefore paramount and must be a priority for any organisation seeking to deploy these agents.

The testing and evaluation of these advanced models is not a simple ‘click a button’ task. The variety of vulnerabilities demands a multifaceted approach and one that is specific to the model and context in question.

It will involve, paradoxically, the use of other AI agents!

(We explored ‘Superalignment’ in more detail here: https://www.linkedin.com/pulse/superintelligence-alignment-ai-safety-advai/ )

Assurance agents – agents that control other agents – are the only feasible way to effectively deploy safeguards. We’ve all seen the Matrix, so we all know what happens if an agent gets out of control. Neo dies. Trinity dies. Zion hosts its final sweaty dance party.

These agents will be dedicated to safeguarding deployed AI 2.0 and will be deployed in swarms. It’s important to deploy these agents in swarms to take advantage of the error correction properties of groups of agents.

In other words, by working as a group, these agents can cross-reference one another's decisions and outputs. Teamwork, it seems, is just as important in the world of artificial intelligence as it is for us meagre organic intelligences.

As with people, working effectively in teams reduces the likelihood of a group making a mistake. If one assurance agent fails to detect a problem or makes an incorrect assessment, the idea is that others in the swarm will catch the error and step in to correct the oversight.

AI 2.0 totally transforms how we conceptualise AI. No longer do we think of contained, bounded algorithms that are restricted to singular domains, single functions, single contexts. No longer will we create singular algorithms to solve a particular problem, which are hard to make and greedy for compute power.

The paradigm has shifted: AI 2.0 is now the amalgamation of intelligent language agents capable of collaboration, whose behaviour is guided by natural language, rather than code.

For Advai, the implications are clear and significant: the assurance ‘job to be done’ will increasingly sit in the domain of natural language. The AI Assurance sector in general is transitioning from testing and assuring models, to assuring LLM agents and harmonising their functionality with legacy AI. Facilitating effective communication and assuring the robustness of these communicating agents will be the protection for the AI 2.0 future.

Decision-makers will (/should) only approve the deployment of autonomous agents if the industry successfully delivers a new set of tools capable of testing and evaluating this latest generation of AI agents. And we must do so in a rapidly evolving landscape – because who knows when AI 3.0 will arrive!

[1] Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., & Anandkumar, A. (2023). Voyager: An Open-Ended Embodied Agent with Large Language Models. arXiv preprint arXiv:2305.16291v2. Retrieved from https://doi.org/10.48550/arXiv.2305.16291

[2] Weng, Lilian. (Jun 2023). LLM-powered Autonomous Agents". Lil’Log. https://lilianweng.github.io/posts/2023-06-23-agent/.

Here are a few links that you might find useful if you are interested in learning more about Robust AI:

1. Artificial Intelligence Risk Management Framework (AI RMF 1.0) (nist.gov)
2. https://pages.nist.gov/AIRMF/
3. US Government Releases Artificial Intelligence Governance Framework (friedfrank.com)
4. How to Prepare for (or Implement) the AI Bill of Rights - RTInsights