OpenAI launched its new AI model called “o1” which surprised users. And unlike ChatGPT, which delivers short, quick responses, this new technology has the ability to think and reason, a double-edged sword for some experts.
A few weeks ago, OpenAI —the company behind ChatGPT — launched its new model of Artificial intelligence “o1” that, contrary to what we already knew, it is a chatbot which has the ability to “think” or “reason” before responding.
This new AI capability would make the most complex reasoning tasks more efficient. It would be a perfect tool for the field of research, science and programming.
This is why A group of researchers helped test this new language model, better known as OpenAI o1, and the result shocked them.
This, not only because of the incredible capabilities it has to facilitate work and provide more complex and precise answers, but also poses a risk, especially for “newbies” who cannot distinguish between an AI hallucination and a real response.
What is OpenAI o1
OpenAI o1 is the new Artificial Intelligence model from the OpenAI company which has much more advanced capabilities than other versions like GPT-3.5 or GPT-4.
The difference with other chatbots lies in the system: the older ones try to respond as quickly as possible with the information they have. OpenAI o1 takes some time to “think” about the answer and provide something more complex and robust.
As the company explains, This AI operates according to “chain-of-thought logic”. First, she explains a series of reasoning steps to herself as she attempts to solve the problem and corrects herself in the same process.
Why scientists were shocked by OpenAI o1
“In my field of quantum physics, OpenAI o1 gives significantly more detailed and coherent answers” said Mario Krenn, head of Artificial Scientists Laboratory at the Max Planck Institute in Germany, at the magazine Nature .
Krenn was one of the few scientists who first tested the trial version of o1.
Like him, other researchers have been able to test this AI which promises to spend more time thinking about its responses. Clearly, The time required to generate them is longer than a simple question asked to ChatGPT, however, this research could help different people who need more precise solutions.
According to Andrew White, a chemist at Future house an organization that applies AI to molecular biology, He and his team had been a little disappointed with previous versions of ChatGPT, as they did not make their work in science easier due to a “lack of capacity”.
However, Now that they have tried the o1 series, “everything has changed”.
According to a test (Graduate-Level Google-Proof Q&A Benchmark) that evaluates doctoral-level academics, OpenAI o1 managed to outperform the best humans even on the most difficult questions.
OpenAI claimed that its own academics scored just under 70% on this exam. the o1 AI managed to achieve an overall score of 78%, with a particularly high score of 93% in physics.
“It seems plausible to me that this represents a significant and fundamental improvement in the basic reasoning abilities of the model.” said David Rein, who was part of the team that developed this test.
“This is significantly better than the second best performance reported by any chatbot.”
o1 was also tested with an exam to qualify for the International Mathematics Olympiad. ChatGPT-4 correctly solved 13% of the problems, while o1 managed to answer 83% of the test correctly.
On the other hand, Krenn incorporated o1 into a tool he developed to suggest “interesting” avenues for future research and ensured that The new technology “generates much more interesting ideas than GPT-4 or GTP-4o.”
Kyle Kabasares, data scientist at Bay Area Environmental Research Institute at Moffett Field California too He implemented o1 to reproduce the code from his PhD project that calculated the mass of black holes and got a big surprise.
And that’s it It took him an hour to accomplish what he had done in several months.
In addition, o1 would also be implemented in other tasks, such as in hospitals: Catherine Brownstein, geneticist at Boston Children’s Hospital in Massachusetts, assured that they used this tool to connect patients carrying certain genes in order to discover rare diseases in time.
The doctor said that “o1 is more precise and offers options that I didn’t think were possible in a chatbot.”
Why OpenAI o1’s capabilities spark fear among some scientists
The same company reported that o1 models began to hallucinate (invent incorrect responses) more frequently than their predecessors, such as GPT-4.
In addition, the scientists who tested this new AI highlighted “missing safety information related to harmful steps, such as failing to highlight explosion hazards or suggesting inappropriate chemical containment methods. »
This, they write, would indicate that “The model is not suitable for high-risk physical security tasks.”
The chemist White added that o1 “it’s not yet perfect enough or reliable enough that you wouldn’t want to check it closely.”
He also explained that its use might be safer for experts than novices.
“For a rookie, this is beyond their ability to immediately inspect.”
In other words, a person who does not have sufficient knowledge about a potentially risky action could be exposed to dangers that, generally, specialists can immediately detect and avoid.
Until nowThe o1 trial is available to paying customers and select developers. There is also o1-mini, which is a smaller, more economical version.
Source: Latercera
I am David Jack and I have been working in the news industry for over 10 years. As an experienced journalist, I specialize in covering sports news with a focus on golf. My articles have been published by some of the most respected publications in the world including The New York Times and Sports Illustrated.
