OpenAI o1 Explained: Why ChatGPT Decided to Slow Down to Speed Up

Speed isn’t everything. While most AI models have been optimized to crank out responses as fast as possible, the new OpenAI o1 model has been designed to slow down and take its time responding to user inputs.

“We’re releasing a preview of OpenAI o1 — a new series of AI models designed to spend more time thinking before they respond,” OpenAI said in a post on X.

“These models can reason through complex tasks and solve harder problems than previous models in science, coding, and math.”

However, while OpenAI o1 has some interesting implications for the development of large language model (LLM) reasoning capabilities, it’s unlikely to excite those who were anticipating a GPT 5 release. After all, a $150 billion valuation comes with unforgiving expectations, particularly in the fast-moving world of AI.

We explore OpenAI’s claims about better reasoning and why slowing down is the new speeding up — or, in other words, thinking before you speak.

Table of Contents Table of Contents

1. Why OpenAI o1 Has Better Reasoning
2. OpenAI o1 Keeps Up Momentum… Just
3. Why We Shouldn't Get Too Excited About OpenAI o1
4. Experts Initial Reactions to OpenAI o1
5. The Bottom Line
6. FAQs

Show Full Guide

Table of Contents

1. Why OpenAI o1 Has Better Reasoning
2. OpenAI o1 Keeps Up Momentum… Just
3. Why We Shouldn't Get Too Excited About OpenAI o1
Show Full Guide
4. Experts Initial Reactions to OpenAI o1
5. The Bottom Line
6. FAQs

Why OpenAI o1 Has Better Reasoning

As part of its reasoning capabilities, o1 uses reinforcement learning (RL) to optimize its chain of thought, gradually refining the strategies it uses to generate responses. This approach means that it can recognize and correct mistakes without human oversight.

Computer science professor Tom Yeh released a post explaining that under “traditional” reinforcement learning from human feedback (RLHF), an initial prompt and the model’s output are fed into a reward model, which then assesses the quality of the response and helps better align the LLM.

You can add RLHF+COT—Chain of Thought, where the model processes a user’s prompt and then “thinks” in steps before generating an answer and?finally sending the data to the reward model.

Using COT in this context means that the model will develop a step-by-step explanation of its reasoning process for its response.

Bring Inference+CoT into the mix, as OpenAI has done, and the model reasons with itself without needing the guiding hand of a human to get to the answer

This enables the model to reason and question itself before providing an answer to the user and reduces the need for human oversight. CoT tokens in this version remain invisible to users.

OpenAI o1 Keeps Up Momentum… Just

Say what you will about OpenAI, but the world’s biggest AI startup knows how to keep up momentum even when it doesn’t have a big flagship release coming. Since the launch of GPT-4 in March 2023, people have been anticipating the release of GPT-5, but over a year later, there’s still no confirmed release date.

Of course, a lack of a GPT-5 release doesn’t appear to have slowed OpenAI down. Intermittent releases like GPT-4 Turbo, GPT-4V, GPT-4o, custom GPTs, and now o1 have kept the AI vendor’s name in the news cycle.

So what’s the selling point of Open o1? Well, its reasoning capabilities enable it to solve complex problems in key areas like science, coding, and mathematics.

As Sandi Besen, applied artificial intelligence researcher at IBM, told Techopedia:

“I think OpenAI is focusing on improving one of the biggest language model limitations — reasoning. The evaluations of how the o1 preview has performed in a question / answering setting seem promising.

“Particularly its performance on law benchmarks where the answer usually requires multi-step reasoning and lots of information to be used as context.”

In fact, according to OpenAI, the model performs similarly to PhD students on “challenging benchmark tasks” in physics, chemistry, and biology while also excelling in math and coding.

It offers improved performance over GPT-4o in some areas. For example, in a qualifying exam for the International Mathematics Olympiad (IMO), GPT-4o correctly solved only 13% of problems, while o1 scored 83%.

Why We Shouldn’t Get Too Excited About OpenAI o1

Although this innovative approach to reasoning has exciting implications for the future, the release of OpenAI o1 as a solution in itself is underwhelming.

COT tokens are unavailable to users, offering little transparency over the model’s reasoning process, and the model itself is pretty slow, with responses to complex questions taking minutes.

Even Sam Altman seems underwhelmed with the release. “Here is o1, a series of our most capable and aligned models yet,” Sam Altman said in a post on X.

“O1 is still flawed, still limited, and it still seems more impressive on first use than it does after you spend more time with it.”

here is o1, a series of our most capable and aligned models yet:https://t.co/yzZGNN8HvD

o1 is still flawed, still limited, and it still seems more impressive on first use than it does after you spend more time with it. pic.twitter.com/Qs1HoSDOz1

— Sam Altman (@sama) September 12, 2024

OpenAI’s announcement blog post also noted that OpenAI o1 doesn’t yet have many of the features that make ChatGPT useful, including browsing the web for information and uploading files and images, though it does pose a significant advancement on complex reasoning tasks.

While there has been lots of positive sentiment surrounding the release, not everyone in the field has been too excited about it.

Sean Ran, CEO and co-founder at Sahara AI, told Techopedia:

“I don’t feel very excited, to be honest. Putting on my AI researcher/professor hat, I see this as a sign that the first principle of model development hasn’t changed — we’re still looking to apply reward-based learning over large amount of data to improve the model, just with more targeted algorithms.”

Experts Initial Reactions to OpenAI o1

To get a feel for the release, Techopedia reached out to a number of experts to find out how the tech community felt about OpenAI’s new model. By and large, the responses were quite positive.

Alon Yamin, co-founder and CEO of Copyleaks, said:

“The release of OpenAI’s o1 model marks a true milestone in generative AI development. Its capacity to evaluate steps before proceeding is genuinely groundbreaking.

“Up to now, AI models have struggled with nuance, so I am intrigued to see how well the O1 model actually performs in that area. I’m thrilled about the potential advancements in mathematics, science, and coding that this model could unlock, especially within higher education.”

Likewise, Besen was also quite enthusiastic about the potential of the release.

“Overall, I’m excited about the potential for models that have better reasoning skills out of the box without having to add on additional methods on top of it using prompting (ReAct, Chain of Thought, etc) or another additional language model call to plan for more complex tasks.

“I’m hopeful that the increased reasoning ability will improve accuracy and therefore gain the trust of business leaders to scale more GenAI use cases.”

Ultimately, OpenAI o1 demonstrates how an alternative approach to AI training can help enhance the reasoning capabilities of LLMs as a whole.

That being said, OpenAI o1 might receive a warm response at best from users who were expecting a big LLM release from Project Strawberry.

The Bottom Line

OpenAI o1 looks like an interesting niche release, but it’s unlikely to live up to the hype it had as Project Strawberry.

But that might be unfair to OpenAI; they are approaching the different approaches to AI in different ways and then offering the public the chance to try them out experimentally.

In any case, OpenAI o1 raises some interesting questions about how AI reasoning will evolve over the next few years.

FAQs