How collaborative AI agents can create disruptive AI systems?
I should start by defining disruptive AI systems.
AI is seeing a wave of breakthroughs from major tech companies, such as prediction machines and generative AI. Despite all these advancements, artificial intelligence is still in its early stages, with the ultimate objective being AI-based decision-making that is independent of human intervention (albeit its ethical implications remain debatable). Artificial General Intelligence (AGI) is another name for relying less (or no) on humans and more on AI systems.
Businesses are integrating AI into their business divisions piecemeal, but no company can claim to be fully AI-run—at least not yet. Therefore, technology-wise, AI is not producing a lot of disruption. However, why?
The HBR book “Power and Prediction” explains it. Any technology must pass through several stages before it may be disruptive, as demonstrated via an electrical demonstration. Even though Edison invented the electric light in 1879, hardly 3% of homes and hardly any industries were using electricity in the early 1900s. For forty years, electricity existed in “The Between Times.”
Even though there was plenty of enthusiasm of electricity, it was not disruptive. The reason behind all of this was implementation economics. Industries were looking at it as a “point solution” that could occasionally (though not significantly) lower their expenses. The primary energy source was steam-power, and businesses were experimenting electricity in certain areas. However, as technologies advanced and the emphasis shifted from discrete solutions to “system solutions,” everything changed. Any technology becomes disruptive when it revolutionizes the entire system in addition to offering an effective answer. At times during the 1930s, factories sought to create industries dependent on electricity rather than electrification.
Does this also apply to artificial intelligence? Most organizations are looking towards AI for single solutions rather than systems solutions because there aren’t enough AI systems.
However, things are about to change. Over the past three to four years, generative AI has advanced significantly, but it was still a point-and-click approach in producing text, images, or videos.
What is Agentic Workflow?
In a laymen language, its saves us from creating the complex prompts to get the desired output from any LLM.
Agentic workflows are AI collaboration that leverages specialized agents (large language models (LLMs)), advanced prompt engineering, and iterative processes to tackle complex problems and drive technological innovation,
As explained by Andrew Ng in his video, non-agentic LLMs work on prompts for outputs but if you need another work to be done, you need another prompt, but Agent workflow does not need several prompts.
Here is an example.
Source: What’s next for AI agentic workflows ft. Andrew Ng of AI Fund – YouTube
Agentic workflow shows great potential in creating the AI systems where number of point-solutions are bundled into a repetitive task aimed.
What is the current state of this technology?
Number of cloud platforms are providing the agent workflows like Vertex AI from google. This platform hosts 130+ models; both open source and closed source but nothing on the collaborative agents so far.
Multi-agent workflow shows greater accuracy as compared to other ways is shown in below picture. Even chat GPT 3.5 shows better results with multi agents as compared to GPT 4.
Source: What’s next for AI agentic workflows ft. Andrew Ng of AI Fund – YouTube
Although multi-agent systems are operating efficiently, we are seeking to disturb the state of multi-agent collaboration, which is still in its infancy and could require a year or two to fully develop.
Not only does this technology increase productivity and yield better outcomes even with subpar LLMs, but it is also a component of AGI (Artificial General Intelligence). Without artificial general intelligence, which is not too far off, system level AI deployment is not feasible.
