New AI from Open AI puts everything else in the shade

 "Open AI is launching a new model called o1. This "reasoning" model has been trained to answer more complex questions faster than a human and to reach the level of a doctoral student.

 

The rumors about the "Strawberry model" have come to an end: Open AI has presented its new model o1, which represents a further step on the way to a general artificial intelligence that is supposed to be as smart as a human. The new model is slower than the previous GPT-4o, but is better at forming "chains of thought" and thus solving more complex multi-stage problems, for example in mathematics.

 

While new models have always been faster and cheaper than their predecessors, Open AI is going in the other direction: o1 is more expensive and slower than GPT-4o, but "thinks" about a task for longer and should therefore lead to better results. 

 

Open AI has used its model in science competitions such as mathematics and physics, but also in programming competitions against humans and - according to its own statements - regularly ended up in the top group at the level of a doctoral student.

 

In online programming competitions such as the Codeforces competitions, the new model achieved a score of 89, meaning it was among the 11 percent best participants. 

 

Open AI also tested o1 with a qualifying exam for the International Mathematical Olympiad. While GPT-4o only solved 13 percent of the tasks correctly, o1 achieved a success rate of 83 percent. 

 

Open AI has also designed the user interface so that the individual "thinking steps" of the model are displayed while it "thinks".

 

o1's training method is fundamentally different from its predecessors. At the heart of it is "chain-of-thought technology", which uses chains of thought in a similar way to a human. With each step, errors are detected and avoided in the next round. This means that the models improve on their own. According to Open AI, the new training methodology should make the model more accurate and largely avoid hallucinations, although this problem could not be completely solved. Safety rules were built into the thought chains to prevent illegal behavior, for example.

 

At the same time, o1 is not as powerful as GPT-4o in many areas. It does not perform as well in factual knowledge about the world and cannot browse the Internet or process files and images. Nevertheless, the company believes that this represents a completely new class of skills. It was named o1 to indicate that "the counter is reset to 1."

 

On the way to autonomous agents

 

Open AI is bringing these reasoning skills to language models because it sees a future with autonomous systems or agents that are able to make decisions and act on behalf of users. For AI researchers, mastering reasoning is an important next step towards human-like intelligence. Reasoning is expected to enable breakthroughs in areas such as medicine and engineering. However, o1's reasoning capabilities are still relatively slow, non-agent-like, and expensive for developers.

 

"We've been working on Reasoning for many months because we believe this is actually the decisive breakthrough," Rob McGrew, Open AI's chief research officer, told industry service TheVerge. "Basically, this is a new modality for models to solve really difficult problems that are necessary to make progress toward human-like intelligence." [1]

1. Neue KI von Open AI stellt alles andere in den Schatten. Frankfurter Allgemeine Zeitung (online) Frankfurter Allgemeine Zeitung GmbH. Sep 13, 2024. Von Holger Schmidt