Beyond the Hype: What AI Actually Can (and Can't) Do • Jodie Burchell & Michelle Frost • GOTO 2026

In a wide-ranging discussion, Jodie Burchell and Michelle Frost offer a pragmatic and research-grounded perspective on the current state of AI, moving beyond the hype to discuss its real-world applications, limitations, and historical context. Drawing on her unique background transitioning from clinical psychology and biostatistics to Natural Language Processing (NLP), Jodie emphasizes the importance of measurement, scientific rigor, and ethics.

Defining “AI” and the Importance of Fundamentals

The conversation begins by addressing the ambiguity of the term “AI.” Today, when people say “AI,” they are almost always referring to generative AI—models based on the transformer architecture from the fields of NLP and computer vision. This creates confusion, as it’s often conflated with the broader application of classical machine learning, which has proven, specific use cases. Generative AI, in contrast, is still a relatively new and unproven branch of models whose best use cases are still being discovered.

Despite the accessibility of pre-trained models, foundational knowledge of machine learning remains vital. Jodie illustrates this with the example of building a Retrieval-Augmented Generation (RAG) system for vector search:

• Data Quality is Paramount: You cannot “magic up meaning” from poor-quality text. The process of dividing documents into meaningful “chunks” is an art that requires domain knowledge.
• Understanding the Data: Issues like missing data are complex. The reason data is missing can point to underlying biases in collection, which affects model performance and generalizability.
• Bias is a Performance Issue: Ethics in AI is not just about fairness; it’s a practical concern. Biased data leads to poor generalization and business risk. Even in a seemingly simple RAG system, the model can’t create meaningful embeddings from something that isn’t there. The fundamentals are crucial for measurement, performance, and generalization.

Historical Cycles: Are We in an AI Bubble?

Jodie points out that the patterns of the current AI boom are identical to previous “AI summers” in the 1950s and 1980s. The cycle is predictable:

1. A new AI technology emerges that performs well on human-like tasks.
2. Researchers and the public become dazzled, leading to speculation that Artificial General Intelligence (AGI) is imminent.
3. The technology inevitably hits a plateau, and the initial excitement gives way to disappointment and reduced funding (an “AI winter”).

The current excitement, particularly around the capabilities of models like GPT-3 and beyond, mirrors these past cycles.

Deconstructing the AGI Hype

The speakers stress that today’s Large Language Models (LLMs) are not AGI. The term “AGI” itself is not well-defined, but it generally refers to a system that can perform a wide range of human-like tasks at or above the level of a skilled human.

A key challenge is measurement. Referencing François Chollet’s 2019 paper, “On the Measure of Intelligence,” they highlight the difficulty in defining the scope of tasks and what it means for a system to be truly “generalizable.” We have achieved Artificial Narrow Intelligence (ANI)—systems that excel at specific tasks—but the leap to general intelligence is not a matter of simply adding more parameters.

LLMs are generalists at language tasks, but language is not equivalent to intelligence, reasoning, or sentience. The discourse around AGI is heavily influenced by science fiction narratives and the deeply held beliefs of some industry leaders, which can diverge from the practical, concrete problems that most data scientists and engineers are trying to solve.

The Nuanced Reality of AI in Developer Productivity

While skeptical of the hype, both speakers are pragmatic about the real-world utility of LLMs, particularly in software development. They point to the Stanford 100K developer study as a source of meaty, empirical evidence that moves beyond anecdotes. The findings reveal a complex picture:

• Productivity Gains are Real but Context-Dependent: Overall, developers saw productivity gains. However, these benefits were most pronounced in specific scenarios: greenfield projects, smaller codebases, less complex tasks, and popular programming languages like Java and Python, which are well-represented in the training data.
• A Trade-off with Code Quality: The study also found that LLM-assisted coding introduced more bugs and led to more refactoring. While the net time savings were still positive, it underscores that these tools are not a silver bullet.
• Core Use Case: Code Understanding: A significant benefit is in code understanding and translation. LLMs can be incredibly helpful for new team members trying to understand a large, existing codebase, which brings the technology full circle to its original use case of machine translation.

Ultimately, the key is to use these tools for what they are—brilliant systems for working with language—and to remain realistic about their limitations rather than pursuing moonshots.

• Tags:
• LLM
• Generative AI
• Developer Productivity
• AGI
• Machine Learning
• AI Ethics
• AI Hype Cycle