After chatbots imitating human dialogue, the development of artificial intelligence models can take a new direction. Four months after ChatGPT-4o, which responds lightning fast, has a sense of humor and imitates human-like gestures in speech, OpenAI announced its brand new model, the o1, and its slimmer version, the o1-mini, which are a completely different teaching method than before thanks to which they have human-like problem-solving skills,
i.e. they are essentially capable of logical thinking.
According to the description on the company’s website, the new model, previously only known by the code name Strawberry, was originally taught to spend more time solving a problem or task, to reach the result through different ways and approaches, and to be able to recognize its mistakes. According to the developers, with this, the o1 can prove to be much more effective in solving more complex tasks compared to all previous models.
HP Wolf Pro Security: An invisible layer of protection in the age of hybrid work (x)
94 percent of cyber attacks start with an email, so endpoint protection is of utmost importance, especially if several employees work remotely.
According to OpenAI, o1, which is still in an admittedly rudimentary state, can already solve certain physical, chemical, biological and mathematical tests, sets of tasks, and functions with efficiency similar to that of PhD students. At the same time, due to the different task-solving chain, the new model gives results for each complex question significantly slower than the previous ones, but it is still much faster than any human reaction.
The company also cites a specific example of the effectiveness of the model, so while the ChatGPT-4o presented in May, demoed with equation-solving capabilities among other things, only solved 13 percent of the tasks correctly from this year’s International Mathematical Olympiad, while the o1 test was 83% Latvian.
At the same time, the new model is not at all suitable for certain tasks, so its subject knowledge, for example, is far behind “traditional” models, it cannot interpret and process images or collect information on the web. o1 does not solve one of the most fundamental problems of artificial intelligences based on large language models, the hallucination, but at the same time, according to the company, the phenomenon occurs less often with this model compared to before.
In accordance with OpenAI’s practice so far, o1 will initially only be available to paying customers, but later o1-mini, which promises to be particularly effective in coding tasks, will also be available free of charge.
Source: www.hwsw.hu
