diff options
Diffstat (limited to 'misc/raw_english.txt')
| -rw-r--r-- | misc/raw_english.txt | 105 |
1 files changed, 105 insertions, 0 deletions
diff --git a/misc/raw_english.txt b/misc/raw_english.txt new file mode 100644 index 0000000..15068c4 --- /dev/null +++ b/misc/raw_english.txt @@ -0,0 +1,105 @@ +You can start by clicking the "start" button. +You will not know what to fill in +if you don't have the correct way of thinking. +Sometimes I don’t have any ideas. +To get a high score, +testers need to reason abstractly +from the information they extracted +from small samples. +A high score cannot be obtained by just doing a lot of exercises. +Towards general artificial intelligence. +Exploring cognitive intelligence with human-level intelligence quotient. +"Artificial intelligence" has always been a window for humans +to explore the boundaries of their capabilities. +In recent years, significant progress in artificial intelligence +represented by deep learning has been made at the perception level, +but there is still a long way for existing models +to achieve intelligence with general human-level cognitive capabilities. +Research has shown that in the case of determining +whether two figures are alike, +primates like capuchin monkeys can do it successfully. +This indicates that animals have an innate cognitive architecture +that allows them to find generic paradigms +for solving problems from small data. +These advantages of cognitive framing are particularly evident in humans. +For slightly more complex geometric problems, for instance, +the Amazonian indigene group in the rainforest +can still solve them easily. +However, the deep learning foundation model represented by Transformer +is dwarfed in similar tests +not only does the model require a large amount of labeled data for training, +but its ultimate performance cannot be comparable to that of humans. +Intelligence levels are generally measured +based on intelligence quotients, +or "IQ" as it is often called. +Psychologists have created a series of tests +to numerically quantify IQ +and have found that IQ level has a high correlation +with human achievement. +Among these tests, +a representative one is Raven's Progressive Matrices. +The following question is an example. +This example is complicated at first glance, +which has only 8 pictures and the shapes of objects are different. +However, a closer analysis shows that +the objects in each row are all dark gray, light gray and black, +and the size of the objects in each picture is basically the same. +Thus, it is not difficult to find the correct answer. +In the case of Odd-One-Out, on the other hand, +subjects are required to pick an outlier data point from several examples. +For example, in the next question, +only the third picture has a dark black hexagon. +For traditional perceptual intelligences, +we need to provide thousands of examples +for the machine to learn the concept of a cat or a dog. +For a cognitive intelligent agent, +however, the machine can abstract the corresponding events from a huge space +from just a few pictures and understand their spatial-temporal-causal relationships. +Exploring models with human cognitive intelligence is +a fundamental research project of the Beijing Institute of General Artificial Intelligence (BIGAI), +in which scholars from BIGAI and UCLA cooperate to address this challenging problem: +how to use small data to understand +spatial-temporal-causal relationships in IQ tests. +After several years of study, +we proposed the Tong-Hui model. +In this summer, we invited students from +top universities in China to +have a competition with our Tong-Hui model. +In the preliminary tests, +we had a rough estimate of the capabilities of the model. +But when faced with truly highly intelligent human opponents, +we were not sure how our model would perform. +All right, I will click the "start" button +to begin the competition. +Students often have a variety of wondrous ideas, +but our program may not have similar thoughts. +Thus, we are not sure +about the result of the competition. +It was quite easy at the beginning, +then it was a little bit tough, +and later I had no idea. +I made 6 to 7 mistakes. +One would make mistakes if he cannot find the correct way of thinking and the hidden pattern, +and he would not know what to fill in. +We needed to spend +a lot of time thinking, +but the machine could quickly +try various solutions in a short period of time. +OK, thank you all. +The Tong Hui model outperformed all the students +and the foundation model represented by Transformer. +The first item in the upper left corner is a pentagon, +the others do not have pentagons. +We have beaten the best students in the country in this task. +Our next step is to provide more robust criteria +for the grading of AI +and to evaluate our general AI systems in a more comprehensive setting. +We were always thinking that +if we ever really created intelligence that +could outperform the world's smartest brains, +we must have discovered some kind of universal algorithm +or even a whole new cognitive architecture. +Perhaps we are already on +the doorstep of general AI right now, +and the success of this competition +will be a further step to it.
\ No newline at end of file |