AI's Agentic Ambition: A Promising Future or a Never-Ending Delay?

The promise of 2025 as the year of agentic AI, where generative robots would take over our tasks, has been delayed, with some experts questioning if it will ever come to fruition. A mathematically-backed paper published earlier this year has fueled this debate.

The Limitations of Language Models

The paper, titled "Hallucination Stations: On Some Basic Limitations of Transformer-Based Language Models," argues that current language models (LLMs) are incapable of performing complex computational and agentic tasks. The authors, including a former SAP CTO and his teenage prodigy son, claim that even advanced reasoning models cannot overcome these limitations.

Challenges in AI Coding

However, recent developments in AI coding, such as the success of coding that took off last year, and the reported breakthroughs by Google and Harmonic, offer a counterpoint. Harmonic, co-founded by Robinhood CEO Vlad Tenev and a Stanford-trained mathematician, has made strides in guaranteeing the trustworthiness of AI systems using formal methods of mathematical reasoning.

The North East Connection

The implications of this debate extend beyond the global AI industry. In North East India, a region rich in technological talent and potential, the development of reliable agentic AI could revolutionize various sectors, from agriculture to healthcare. However, the challenges in ensuring the reliability of these systems underscore the need for rigorous testing and regulation.

The Persistence of Hallucinations

Despite progress, hallucinations remain a significant issue in AI models, as demonstrated by OpenAI's findings last September. These inaccuracies can disrupt workflows and hinder the widespread adoption of agents in the corporate world.

Guardrails and the Future of Agentic AI

The industry's response to these challenges is the development of guardrails to filter out hallucinations. These systems, if effective, could pave the way for the widespread adoption of agentic AI, offering benefits such as increased efficiency and cost savings.

Reflections and the Road Ahead

As we move forward, the debate over the feasibility and desirability of agentic AI will continue. While mathematical verifiability may be beyond reach, the ultimate assessment of the impact of agentic AI on our work and lives may not be. As computer pioneer Alan Kay suggests, the focus should be on understanding the implications of this automation, rather than debating its merits or demerits.