The leaks from OpenAI turned out to be true. The world’s largest AI startup has released another model, GPT-o1 (currently in preview). There are many changes to previous models, but o1 is aimed primarily at scientists and researchers. Why does the GPT-o1 model have an even 10 times slower response generation time, and is the latest model from OpenAI “smarter” than other LLMs available on the market?
o1 preview, or the Strawberry model from leaks becomes a fact
OpenAI has unveiled the latest model in the GPT series, called GPT o1 preview. This model stands out from previous generations with significant improvements, especially in areas related to science, technology, engineering, and mathematics (STEM). But that’s not the only thing that makes it unique. GPT o1 is also less susceptible to jailbreaking and has a lower tendency to generate hallucinations, which makes it a milestone on the way to creating artificial general intelligence (AGI). It is worth adding that the latest model from OpenAI is 3 times more expensive than GPT-4o. Open AI has also released o1-mini, a reduced but still very powerful model, which in turn is 20 times more expensive than GPT-4o-mini!
Strengths of GPT-o1
The new model stands out for its extremely precise ability to solve STEM problems. Compared to its predecessors, the new model better analyzes complex mathematical, physical, and programming problems, which could contribute to the wide application of AI in the exact sciences. For specialists working with difficult data and analysis, this is a huge advance that allows for even more accurate and advanced solutions.
The new model is also supposed to be much more resistant to jailbreaking, a process that allows users to force the model to generate unwanted or dangerous responses. OpenAI focused on securing the model against this type of manipulation in the o1 preview model. Thanks to advanced security mechanisms, the model defends itself much better against attempts to bypass its limitations. This is a step forward in ensuring that the generated content will be ethical and safe for users.
One of the main problems in large language models is their tendency to create so-called “hallucinations” – generating information that looks plausible but is not true. GPT-o1 shows significant improvement in this area. Improvements in the model training algorithms make the generated answers more precise and based on solid data, which increases their credibility and value.
o1 smarter than the average person?
The results of the o1 model are very impressive. In the International Mathematical Olympiad qualifiers, the GPT-o1 model got 83% correct, compared to only 13% for its predecessor (ChatGPT-4). What’s more, when it comes to coding, the o1 is in the 89th percentile in Codeforces programming competitions. While I don’t really like comparing human intelligence to generative AI, such tests have appeared online. According to such tests, the o1 model has an “intelligence” of 120 points on IQ tests.
GPT-o1 Model – The Future of AI and AGI
The current achievements of the o1 model provide a strong foundation for further development of AI. As the model becomes more precise, safe and versatile, we begin to see how it could evolve into artificial general intelligence (AGI). AGI is the concept of a machine capable of understanding, learning and solving problems at the level of a human being, without being limited to specific domains of knowledge.
GPT-o1 preview is a step in this direction, especially with its advanced STEM capabilities and improved ability to avoid bias and misinformation. With each AI model, we are getting closer to the point where machine intelligence can function at the level of the human mind, collaborating with us in even more advanced and complex domains. GPT-o1 is one of OpenAI’s most important achievements, showing how close we are to revolutionizing AI. The model offers significant improvements in many areas, suggesting that AGI is becoming a more and more realistic goal.
