A note to all subscribers:

Welcome to The Sequence 2025! I’ve been eagerly waiting for the end of the year to propose some changes that I believe will tremendously improve the experience for you, the readers of this newsletter.

We started The Sequence a few years ago as a hobby project at a time when AI had yet to reach the mainstream levels of popularity it enjoys today. Little did I suspect that we would approach 200,000 subscribers, including members from some of the world’s top AI organizations.

What has always set The Sequence apart from other newsletters is its focus on deep technical content and original ideas rather than chasing news or hype. Over the past few weeks, I analyzed current readership patterns and came away with some important insights:

• The Sequence should feature content targeted at both AI scientists and engineers, covering both research and implementation topics.

• The educational series are highly popular, but readers sometimes skip the engineering sections because they’re too long or unrelated to the main topic.

• Our original and controversial topics, explored in longer pieces, consistently attract the highest readership, which I love.

• While we’re bombarded with sponsorship requests, I’ve decided to slow things down and work out a non-invasive strategy that aligns better with our audience.

• Interviews have received very positive feedback.

• Some of the branding, like The Edge and The Scope, feels outdated.

With this in mind, in 2025, we’re nearly doubling our coverage with the following weekly editions:

1. The Sequence Knowledge: Continuing with educational topics and related research. We’re kicking off an exciting series on RAG and have others lined up on evaluations, decentralized AI, code generation, and more.

2. The Sequence Engineering: A standalone edition dedicated to engineering topics such as frameworks, platforms, and case studies. I’ve started three AI companies in the last 18 months so have a lot of opinions about engineering topics.

3. The Sequence Chat: Our interview series featuring researchers and practitioners in the AI space.

4. The Sequence Research: Covering current research papers.

5. The Sequence Insights: Weekly essays on deep technical or philosophical topics related to AI.

6. The Sequence Radar: Our Sunday edition covering news, startups, and other relevant topics.

That’s six editions! Each one will be relatively short and easy to follow. We’re starting with this new structure next week!

We’re also discussing potential price changes for new subscribers (current subscribers won’t be affected), so I encourage you to subscribe in the next few days if you haven’t already.

To the companies reaching out with sponsorship opportunities: thank you for your patience. We’ll have something to discuss soon.

I hope you love these changes. If nothing else, they should make The Sequence even more enjoyable while doubling down on what we do best.

Thanks for your continued support.

Jesus Rodriguez.

Now, onto today’s edition! As mentioned before, we are not doing our typical Sunday edition given the limited market activity during the holidays and, instead, we are discussing another controversial AI topic.

Is Reasoning Exclusive to Massive Models or do Small Models Have a Chance?

The rapid advancement of artificial intelligence has led to the emergence of large language models (LLMs) like GPT-01 and o3, which exhibit remarkable reasoning capabilities. However, the prevailing notion suggests that such reasoning is primarily confined to models with extensive parameter counts, often exceeding hundreds of billions. This essay explores whether small language models (SLMs) can develop reasoning abilities comparable to their larger counterparts. We will discuss the nature of reasoning in LLMs, the techniques that enhance reasoning in SLMs, and challenge the assumption that reasoning is exclusive to larger models.

Understanding Reasoning in Large Language Models...