AI & Trends

    The Future of Large Language Models: Introducing OpenAI's "o1" Series

    Mikhail Shevtsov// September 13, 2024// 5 min read

    As a tech enthusiast and engineer, I'm excited to share a massive milestone in AI evolution. Yesterday, OpenAI released a new series of "Thinking" models called "o1". This isn't just another incremental update; it's a fundamental shift in how machines process logic.

    The Limitations of Traditional LLMs

    To understand why o1 is special, we first need to look at how typical Large Language Models (LLMs) like GPT-4 work. They operate on a straightforward principle: generating the most probable next token based on the input.

    Imagine the prompt: "Continue the phrase in one word: I was riding a..." The model predicts a bicycle. It's much more probable than a bomb or a snail. By looping through input and generated output, the model gives you a decent result.

    The Problem? The model doesn't actually "think" about the answer; it just predicts it. This is why models consistently fail at the infamous "strawberry" test. Ask most LLMs: "How many R letters in the word 'strawberry'?" and they'll confidently tell you 2. Wrong.

    Introducing OpenAI's "o1" Series

    The o1 series solves this by introducing "reasoning tokens". Instead of rushing to an answer, the model uses an internal "Chain of Thought" before it ever types a word. It iterates over the problem, checks its own logic, and corrects itself in the background.

    Let's look at a real-world logic test:

    There are 5 people in the room. A is watching TV, B is playing table tennis, D is sleeping, E is eating. What does C do?

    Standard models often get confused. But o1-mini understands the human context: table tennis requires two people. Therefore, C must be playing with B. This level of contextual reasoning is a game-changer for engineering and complex problem-solving.

    Pricing & Technical Specs

    Quality comes at a price. The o1 series is significantly more expensive than the gpt-4o models we are used to:

    ModelInput CostOutput Cost
    o1-preview$15.00 / 1M tokens$60.00 / 1M tokens
    o1-mini$3.00 / 1M tokens$12.00 / 1M tokens

    You can explore the full breakdown on the OpenAI Pricing page. For developers, I highly recommend checking out the Official Reasoning Guide to understand how to implement these "thinking" steps in your own API workflows.

    What's Next for Wiregate?

    At Wiregate, we believe o1 is not a replacement for gpt-4o, but a powerful addition. While gpt-4o remains great for creative writing and simple tasks, o1 is the heavy lifter for coding, complex architecture, and scientific reasoning.

    We firmly believe that models with integrated reasoning will eventually replace all "simple" models. We are already looking into how to integrate these capabilities into our clients' infrastructures to automate even more complex business logic.