AI Search Tools are Evolving – But Not Without Issues

AI-powered search tools like Perplexity and Arc are rapidly gaining popularity among users seeking quick, conversational answers. Although these platforms strive to emulate human-like assistants with cited sources, there's a growing concern: they often produce hallucinations.

What Are Hallucinations in AI?

In the world of AI, hallucinations occur when systems confidently generate false information, misquote sources, or rely on outdated data. A striking example involved Air Canada's chatbot dispatching a fictitious refund policy to a distressed customer, which ultimately resulted in legal repercussions for the airline.

Transformers: The Backbone of AI – But Flawed

AI models like GPT-4 rely on transformers, a system designed to predict the next word in a sentence by assessing the relationship across all words simultaneously. While this mechanism produces fluid and coherent text, it has fundamental shortcomings that contribute to hallucinations.

Why Transformers Get It Wrong

1. **Token Prediction Over Truth-Seeking**: Transformers generate statistically likely text rather than factually accurate content. Gaps in training data lead them to make educated guesses that can sound correct but are contextually and factually flawed. 2. **Computational Inefficiency**: Transformers analyze relationships for every word, which can become computationally expensive and lead to shortcuts that overlook critical context. 3. **Source Blindness**: Without the ability to discern reliable information, transformers sometimes cite fabricated or outdated sources, like an AI-generated LinkedIn post that misrepresented content.

Enter State Space Models: The Future of AI?

State Space Models (SSMs) are emerging as a potentially superior alternative to transformers, particularly for sequence-based tasks. Unlike transformers, SSMs process information incrementally, akin to how humans read and understand.

How Do SSMs Work?

SSMs build understanding piece by piece, mitigating context overload. They require significantly less computational power over lengthy texts because their memory usage grows linearly, allowing for greater efficiency.

Real-world Applications: From AI Search to Robotics

1. **Perplexity's Hallucination Pitfalls**: Despite incorporating real-time data retrieval, Perplexity has mistakenly cited non-existent markets and AI-generated travel guides. Its reliance on unreliable sources emphasizes how transformers often equate all retrieved information regardless of its authenticity. 2. **RoboMamba's Precision**: This SSM-focused robotics model demonstrates real-time error correction by adapting to changing conditions, prioritizing safety. This capability could minimize the risks lodged within performance-based errors.

How SSMs Stack Up Against Other Models

While other methods, like reinforcement learning with human feedback, help somewhat in mitigating hallucinations, they still don't address the root of transformers' guessing tendencies. Knowledge-Augmented LLMs involve external databases, but still fundamentally rely on transformer frameworks.

What Does This Mean for You?

For everyday users, the adoption of SSMs signals a move toward fewer inaccuracies, enhanced complex question handling, and even improved privacy with potential local processing. Imagine receiving precise medical advice based on continual sourcing rather than fabricated data.

The Path Ahead: SSMs and the Future of AI Search

The transition to SSMs is in progress across various industries, with applications in banking, healthcare, and law where accuracy is paramount. The prospects of integrating SSMs with existing transformer-like architectures are also being explored, potentially merging the strengths of both models.

Conclusion: Trust in AI Through Elegant Architecture

The race for the most effective AI search engine highlights the importance of trust over mere speed and aesthetics. By addressing fundamental flaws associated with transformers, SSMs pave the way for AI that truly comprehends and verifies information. The future of AI is not just about retrieving answers, but about constructing them one verified fact at a time.