Welcome to this new (special) edition!

Hello readers, Douglas here!

This week, OpenAI released its new model, o1. I couldn’t resist doing a special edition of The AI Newsroom.

• What is o1?

• How does it work?

• Why does it matter?

• Can o1 create new knowledge?

All these questions deserve a special edition. I’ve broken down everything to answer the questions you might have.

I hope this helps clarify things for you. Enjoy!

What is o1 and how does it work?

OpenAI has just launched a new model called o1, previously codenamed "Strawberry." This model is a major leap forward in its ability to handle complex reasoning tasks. Scoring in the 89th percentile in competitive programming and surpassing Ph.D.-level knowledge in subjects like physics, biology, and chemistry, o1 brings a new level of intelligence to the table.

Unlike earlier models, which often provided quick, surface-level responses, o1 is designed to think methodically through each problem. It uses a technique known as chain of thought reasoning, which helps it solve tasks step by step rather than rushing to an answer. Much like how humans work through long division by writing out the steps, this approach allows the model to stay focused and deliver more accurate responses.

While the chain of thought was once just a prompting strategy to improve results in earlier versions of GPT, o1 has been trained to use it automatically through reinforcement learning. Now, when you ask a question, the model’s thought process becomes visible through an expandable indicator, showing you how it reasons through the problem before delivering its response.

Why we should care about it?

Previously, AI models couldn't reason—they just predicted answers based on patterns in data. o1 is different. It breaks down complex tasks step by step, mimicking human-like reasoning.

This is a big leap forward. With o1, we move beyond quick responses to thoughtful problem-solving. It opens up the potential for smarter AI agents, quicker scientific discoveries, and real breakthroughs in fields like medicine and technology.

Can o1 create knowledge?

If you're curious about how far o1 could go in creating entirely new knowledge, consider this thought experiment:

What if you trained a version of o1 only on writings from before 1500? Or before 1800? Or 1900? Would it discover geocentrism, calculus, the steam engine, or the assembly line?

My guess is that no matter how long it ran, o1 would still seem stuck in the past. A version trained on pre-1500 knowledge might predict the motion of the stars using the Ptolemaic system, but it likely wouldn’t propose geocentrism, which only gained popularity in the 1700s. It would feel like talking to a smart—but outdated—ancestor.

For now, I think o1 will excel at reasoning through existing knowledge, but it's not likely to generate entirely new ideas. It's more of an advanced tool for applying what we already know, refining the abilities of earlier models, rather than a source of groundbreaking discoveries.

Is inference-time computing the next big shift in AI?

Chain of Thought reasoning may be the next big leap in AI. Until now, AI improvements came from more data and computing during training.

With OpenAI's o1 (formerly Strawberry), there's a new focus: compute during inference. When o1 takes longer to process a prompt, its accuracy improves. Previous models like GPT-4 often lost focus when given more time, but o1 stays on track.

This opens a new way forward—improving performance by giving o1 more time to think, rather than relying on massive compute power for future models like GPT-7.

One model to rule them all?

A major question in AI is whether we'll end up with a single dominant model or a variety of specialized ones. Will everyone rely on GPT-7, or will different models emerge for different tasks?

In my opinion, we’re likely to see both—a couple of big leaders and a range of specialized models. ChatGPT is already versatile enough for most tasks, and OpenAI keeps extending its lead as a general-purpose tool.

However, there’s still demand for specialized models. o1, for example, excels in math, while Claude has proven to be a much better writer in my early tests. This suggests that specialized models will continue to hold value for specific use cases.

Thanks for your time!

See you on Tuesday at 9:12 am!

I'd like to sincerely thank you for taking the time to read this special edition of The AI Newsroom.

I've tried to condense as much value as possible to help you understand these new OpenAI models.

Please let me know how you found this special edition by replying to this email or by answering the questionnaire below. 👇

♻️ Please also feel free to share as much of the newsletter as you can with your friends, colleagues, or AI girlfriend. It helps me enormously!

Keep Learning!