© The Indian Express Pvt Ltd
Tags:
Journalism of Courage
OpenAI has dropped its latest AI model, which is part of its secretive ‘Project Strawberry’. The new OpenAI o1 is the first in a planned series of “reasoning” models that have been designed to solve more complex tasks and problems in science, coding, and maths.
The company, in a statement, said, “We are releasing the first of this series in ChatGPT and our API. This is a preview and we expect regular updates and improvements. Alongside this release, we’re also including evaluations for the next update, currently in development”.
How does the new model work?
According to OpenAI, the new model is designed to “think” more carefully about queries before answering, much similar to how humans do while solving a complex problem. The company said through its training, the model learns how to approach problems from different perspectives, check its output and learn from its mistakes.
Based on their tests, OpenAI said the next model update in the series can perform as well as PhD students in subjects like physics, chemistry, and biology, and it is particularly strong in maths and coding. For instance, in a tough maths contest, the model solved 83 per cent of the problems when compared to earlier versions that were only able to get about 13 per cent of the problems. Similarly, in coding, the model ranked higher than 89 per cent of coding participants.
However, OpenAI has said that the o1 model is an early version and is yet to be equipped with other handy features such as browsing the web or managing files and images. It excels in solving more complex tasks, which is a big leap forward in AI technology. The first of the series is being introduced by the company as “OpenAI o1”.
OpenAI has said that the o1 series excels at accurately generating and debugging complex code. The company has also released OpenAI o1-Mini for developers, especially to endow them with a faster, cheaper, and reasoning model that is effective at coding. The o1-mini is 80 per cent cheaper than the o1-preview version, making it a cost-effective application with reasoning albeit broad knowledge.
What about the safety of the models?
OpenAI has said that to make these models safe to use, they have come up with a new way of training. This new method helps the model to understand and follow safety rules better by using its problem-solving skills to stick to the guidelines more effectively. OpenAI tested how well the model follows safety rules, even when users try to trick the AI into breaking them (something known as AI jailbreaking).
The company said that in one of its toughest tests the earlier version of the AI scored 22 out of 100, and the new version scored 84 — a significant improvement. OpenAI has also improved its safety efforts by working closely with the UK and the US governments and their AI safety teams. This includes running through tests, having experts try to spot weaknesses in what is called “red teaming”, and getting inputs from the company’s board, including a special safety committee.
The AI firm has also been given early access to a version of the model to safety groups for research purposes. OpenAI believes that this partnership would help it to test and improve future models before they are publicly available.
How is it going to impact jobs and research?
Since the new model can solve complex problems, it may impact jobs that could include tasks such as software development, data analysis, coding, mathematical modelling, etc. Reasoning is an intellectual task, and automating it would likely reduce the need for human involvement in some routine coding and troubleshooting in industries like IT, finance, and engineering.
Since OpenAI o1 can perform tasks at a level that is on par with PhD students, the existing workers may need to acquire higher-order thinking skills like creativity, critical analysis, and problem-solving, which AI models may not fully imitate. A model like this could lead to reduced demand in some traditional roles. However, it may also lead to the emergence of new opportunities in AI safety, ethical use of AI, and AI maintenance.
When it comes to researchers, OpenAI o1 can be an assistant in fields like physics, chemistry, biology, and healthcare as it can speed up problem-solving. The model’s ability to generate formulas and analyse large datasets makes it a great tool for researchers looking for major breakthroughs.
In this sense, OpenAI o1 is a big step forward in AI’s handling of tasks that are traditionally managed by highly skilled professionals.
How to use OpenAI-o1?
The OpenAI o1 model can be accessed by ChatGPT Plus and Team users starting today. The o1-preview and o1-mini can be selected using the model picker. At launch, OpenAI said that the weekly rate limits will be 30 messages for o1-preview and 50 for o1-mini.
The company is currently working towards increasing these rates and allowing ChatGPT to automatically opt for the right model for a given prompt or query. ChatGPT Enterprise and Edu users will get access to the models from next week.
