OpenAI has presented o3, its new family of reasoning models that promises to bring humanity closer to artificial general intelligence (AGI). The proposal is in the experimental phase and has been made available to external researchers to evaluate its capabilities and ensure its safety before the official launch.
The set includes two versions of the algorithm. The standard o3 is designed to process large amounts of data and perform general tasks. For its part, the 03-mini variant operates with a smaller number of parameters and is ideal for responding to more specific tasks.
The company led by Sam Altman claims that its new products outperform OpenAI o1 in all performance benchmarks. In SWE-Bench Verified coding tests, o3 is 22.8% more accurate than its predecessor. Reaches almost 90% correctness in the benchmark GPQA Diamond that evaluates expert-level scientific problem-solving skills. The system was able to solve 25% of the most advanced mathematical and reasoning challenges. Until now, no other AI had managed to complete more than 2% of these tasks.
Technological advancement suggests that OpenAI is getting closer to AGI. Their new set of models scored 87.5% on the ARC-AGI indicator, a test that measures the algorithm’s abilities to efficiently acquire knowledge on its own beyond the data it was trained on. The o3 generation tripled the performance of o1 in this field.
OpenAI wants to guarantee the security of its new AI
The startup shares that o3 has been trained to “reflect through a private chain of thought” before responding. When it receives an instruction, it pauses, fragments the request and relates it to other previous instructions to deliver a more precise result. The procedure is a widespread feature in artificial intelligence reasoning resources and is associated with a decrease in response speed.
One of the new features of o3 is that it allows you to define the calculation time. Users can choose between low, medium and high data processing speeds. The greater this range, the more reliable the answers proposed by the algorithm will be, according to the firm.
The creator of ChatGPT shares that her new AI has been trained with a technique known as “deliberative alignment” to align its operation with internal security principles. This method forces the system to analyze and “actively reason” whether a user’s request fits the criteria for appropriate use of OpenAI technologies.
Sam Altman says the goal is to launch the o3-mini at the end of January 2025. The standard version of the model will arrive “in later months.” The executive has stressed that the company is working to guarantee the security of its innovation before opening its access to the public.
OpenAI has been questioned for alleged security omissions in the production and marketing processes of its AI solutions. Some executives at the organization left their positions citing concerns. The firm is preparing to adopt a fully commercial business model, according to various reports. Analysts suggest this approach could raise new concerns about the security of its AI. The firm has taken various measures in response to these allegations, although experts warn that they are not enough to address the risks that the misuse of sophisticated AI tools may represent in the immediate future.
#OpenAI #launches #family #reasoning #models