AI in 2024: A Year in Review

In this reflective podcast, the team at Winder.AI — Dr. Phil Winder, Charles Humble, and Jonathan Hunter — take a deep dive into their year, discussing trends, lessons learned, and their vision for 2025.

The Allure of Code and Data

Charles and Phil discusses how alluring writing code is, and how it provides an intrinsic reward to work. Phil talks about how he loves working with data for the same reason. Phil goes on to talk about public data and how stories can be told with it. Phil demonstrates a visualisation of the north-south divide in the UK, plotting the average income for small geographic regions. He then compares that with a visualisation after compensating for housing costs. This tells a story of how wealth is spreading up towards Birmingham and towards Manchester. Wales and the East Coast are clearly poorer.

Most Popular Blog Topics and AI Trends

Charles shared insights into the year’s most popular blogs on their website. Predictably, large language models (LLMs) dominated, with articles on token counts, fine-tuning, and prompt best practices resonating widely. He noted that generative AI has rapidly commoditized, drawing attention from businesses of all sizes. Key insights include:

• Generative AI’s Rise: The accessibility of AI tools like ChatGPT has broadened adoption, with organizations exploring diverse applications.
• Tokenization and Context Windows: Practical guides on optimizing LLM usage received significant interest.
• Reinforcement Learning: Older posts on reinforcement learning continue to attract steady traffic, reflecting ongoing interest in this technical domain.

Here are links to articles that Charles mentioned:

• Big data in small context windows
• Calculating LLM token counts
• RAG vs. fine-tuning
• Reinforcement learning articles

Challenges and Lessons from AI Projects

The team discussed the rapid commoditization of AI technologies and the challenges associated with it:

• Market Maturity: Jonathan noted a shift in the types of clients that Winder.AI works with. Many organizations now recognize the potential of AI but lack clarity on how to implement it effectively.
• Phil highlighted the change in terminology over the past decade from signal processing to machine learning to artificial intelligence.
• Rapid commoditisation of AI: Charles highlighted the rapid commoditisation of AI, and how this has changed the market.
• Proprietary vs. Public Data Models: Phil highlighted the potential for businesses to use proprietary data to build domain-specific models, addressing niche challenges while maintaining security and performance.

LLM Fine-Tuning to Ease the Pain of Large Prompts

Phil highlighted a recent example where a prompt was getting too long and complex, leading to the inability to improve performance. The solution is to fine-tune the model, which allows the prompt to be simplified and performance to continue to be improved. In this vain, if you’re reading this, you should also consider splitting the problem up into smaller agents, like we did for a recent legal project that leverages LLMs.

Emerging Trends: Small Language Models

Looking forward, the team predicted a shift toward smaller, efficient language models in 2025. Phil explained:

• Efficiency Gains: Distilling large models into smaller ones can reduce costs and improve speed, making AI more viable for smaller organizations.
• Customization: Smaller models allow businesses to focus on specific needs while retaining high performance.

Regulations and Ethical Considerations

Phil discussed the implications of the EU AI Act, a groundbreaking regulation targeting the safety and ethical use of AI. Key takeaways include:

• Risk Categorization: AI applications like use cases in HR are labeled as prohibited, or high-risk, requiring compliance measures.
• Innovation vs. Regulation: Striking a balance between regulation and fostering innovation remains a challenge.

Also see:

• The EU AI Act: What it means for your business
• Interview with Kai Zenner on the EU AI Act

AI Failures and Testing

Jonathan brought up humorous and serious examples of AI failures, from bizarre fast-food voice-order mistakes to chatbot inaccuracies. These incidents underscored the critical need for robust testing and validation. The team emphasized:

• Non-Deterministic Testing: AI’s inherent unpredictability demands new approaches to testing and validation.
• Fine-Tuning for Accuracy: Tailoring models to specific tasks can improve outcomes and reduce errors.

Also see:

• Testing LLMs
• Helix.ML - your testable private LLM stack

AI in 2025 and Beyond

The team speculated on future AI trends:

• Organizational Transformation: Jonathan highlighted the need for companies to build internal AI capabilities, blending technical expertise with traditional roles.
• AI touches all areas of the business: Jonathan and Phil highlighted how necessary it is to upskill the workforce and our children.
• Data Sovereignty: Phil predicted a growing focus on private, secure AI models, particularly for sensitive industries.
• Small-language models: Phil predicted a growing focus on small-language models, which are more efficient and faster.

Also see:

• Small Language Models

Non-Tech Highlights

On a lighter note, the team shared personal favorites:

• Books: Charles recommended Not the End of the World by Hannah Ritchie and Orbital by Samantha Harvey, both of which delve into optimistic perspectives on the future.
• Books: Phil recommended Competing Against Luck by Clayton Christensen and A Skeptic’s Guide to the Future by Dr. Steven Novella.
• Motorsport Advances: Jonathan shared excitement over innovations in Formula 1 engineering.

Closing Thoughts

The conversation concluded with reflections on the evolving AI landscape and a shared optimism for the future. As Phil remarked, “The next few years will bring foundational changes, not just in AI, but in how businesses operate and innovate.”

The team expressed gratitude to their clients and followers, inviting continued collaboration as they shape the future of AI.

What was your highlight of the year? Let the Winder.AI team know your thoughts and explore their resources for staying ahead in the AI revolution.