Artificial intelligence is constantly evolving and developers are improving it. Therefore, in 2024, Open AI's AI reached human level in general intelligence test.
We are talking about the new o3 system, which scored 85% in the ARC-AGI test. What is this test and what do its results mean – Channel 24 reports with reference to The Economic Time.
What do the test results mean
The previous record for AI in this test was 55%. Now, the AI has shown a result that is comparable to the average human. This event brings OpenAI one step closer to creating artificial general intelligence. Despite the fact that there are many skeptics of creating an AI, many researchers say that the prospect of its creation is closer than previously thought.
The ARC-AGI test that AI passed is a test of its ability to adapt to new situations. The test was used to determine how many examples are needed to understand the work of AI. For example, the well-known ChatGPT is not effective in this, although it is trained on millions of test examples, but for more difficult tasks it lacks data.
Until AI can learn effectively on a small sample, it will be used for repetitive tasks where errors are permissible. The ability to generalize new problems from a minimum of data is a key element of intelligence. The ARC-AGI test tests the ability to adapt to a sample by offering problems with grids. The AI needs to find a pattern that turns the left grid into the right. Each question contains 3 examples, after which the AI must generalize the rule for the fourth. This is somewhat reminiscent of IQ tests that are given in schools.
The o3 model has demonstrated a high ability to adapt. Although very little is known about this system, because the developers do not tell much. The prospects of this system will become clear after its release. In particular, it is now questionable whether AI can really generalize as well as the average person. If this is indeed the case – then this will cause a technological revolution and usher in the era of self-improving AI. This will require new guidelines for AI and new approaches to management.