Found something new. Feed AI a picture and it can guess a person's personality for you. Just by relying on the playlist, you can make a guess about a person's recent mental state. And these are all from KimiThe latest k1 visual thinking modelhand.
No, it has only been a month since the last mathematical model k0-math was launched and ranked o1, and k1 has made its debut quickly.
Of course, this K1 is not just as simple as the beginning, just looking at pictures and analyzing personality.
When we tested k0-math last time, we had already experienced the ability of "problem master", and the thinking process of solving the problem was astonishing to the reviewers. It's a pity that some of the math questions and geometry questions that revolve around logic are a bit less interesting.
But this time K1 has something to say,Have both reasoning and visual abilities, which means that you can directly take photos and upload them to solve problems, and it is also claimed to be able to equal or even surpass OpenAI's o1 in mathematics, physics and chemistry.
If we compare like this, we will be very excited. It just so happens that the new k1 model does not need to wait for internal testing now. It can be used in both the App and the web version. Without further ado, we will start the whole process directly.
When I came up, I threw a K1 geometry question from this year's college entrance examination.
First of all, k1 interprets the question carefully enough and knows what his goal is.
The cosine theorem that may be involved in the conditions given in the question is also taken into consideration. It is similar to our thinking when solving the problem. When we see a²+b²−c²=2ab, we will immediately think of the cosine formula c²=a²+b²-2ab·cosC.
Then continue to deduce according to the formula and conditions, and you can quickly find the angle B = 60°.
Swipe up to see more
Question (2) is a little difficult, but the reviewer carefully checked the problem-solving process of k1 and found that the ideas and solutions are all correct. The final answer of side length c=2√2 is also correct. (Because the thinking process of question k1 is too long, the screenshot is not shown.)
The same question was asked to o1. First of all, in terms of reasoning speed, o1 already lost in 58 seconds.
In terms of accuracy, o1 and k1 are tied, both got it right.
The difference is that o1 hides the idea of answering the question and does not give a complete thinking process like k1.
However, one thing to say is that the reviewer is not particularly surprised by the k1 model's way of imitating human thinking. Because the k0-math model shocked me last time. It seemed that it was aware of its mistakes and would conduct repeated verifications.It looks like me racking my brain when writing math problems.
In contrast, this time K1 is more outstanding in making up for the shortcomings. I tried the junior high school geometry question on k0-math last time with k1 again, and now I can get it right. Even the difficulty of the college entrance examination is not intimidating.
And I also found that k1 is not only good at math problems, but also physics problems.
Swipe up to see more
Then, I took out another logic trap question with a slightly convoluted logic and tried it:A watermelon was purchased for 50 yuan and sold for 70 yuan. The boss received 100 yuan in counterfeit currency. How much money did he lose in the end?
This question seems simple at first glance, but the answers given by netizens to this question are varied. Some say they lost 150, some say 180, and some say 100...
Let’s take a look at a question that even many humans can’t figure out, and see if k1 can spot the trap inside.
Moreover, I deliberately hand-wrote this question in a sloppy way, and by the way, I also tested whether K1’s visual ability is as good as advertised.
Don't tell me, don't tell me, the "eyes" of this model are indeed not bad.
In terms of the accuracy of the question, the analysis in the first half of k1 first came up with an answer of losing 100 yuan, but it quickly denied itself.
continue toCounterfeit currency, change and cost profitTaking these complex factors into consideration, we finally figured out that the boss lost 80 yuan. (The correct answer is 80 yuan)
Swipe up to see more
This logical ability is indeed a bit strong.
Including I gave k1 several analogical reasoning questions for the test. Although the logical analysis path was different from the reference answer, the final answers were all correct.
Swipe up to see more
Anyway, after this test, the reviewer found that k1 can think logically, has good eyes, and has a high IQ. Kimi is"Being a question writer"His name is finally confirmed.
But in addition to doing the questions, I also figured out more fancy ways to play this time.
There is no point in analyzing data and looking at reports. Doesn’t the k1 model make inferences based on pictures? Then it must be good at identifying ancient coins, right?
I specially found a picture of silver coins from the Republic of China period on the Internet. Two silver coins were fake on the top and real on the bottom. I sent it to k1 for a quick review."AI version of listening to the spring to identify treasures".
Picture source Xiaohongshu user @古古金来 (commented by public blog agent)
K1 not only knew that the coin was from the Republic of China period, but also outputted all the details of the coin. In the end, he actually saw that the coin above was a fake.
Let’s just send a picture of the room and let k1 take a look at the “feng shui”.
What about "air vents", symmetrical layout, energy balance... He talked about it clearly and clearly, and even gave us suggestions, such as changing the position of the bed, pruning the plants regularly, and changing to a simpler chandelier.
Swipe up to see more
Take a picture of k1 while eating, and it will be clear how many calories you consumed during the meal.
Swipe up to see more
But what amazes me the most is k1Guess the movie by looking at picturesability.
I gave it a screenshot from the movie "Seven Deadly Sins". There are no lines, only pictures. For many people who have never seen the movie, it is difficult to guess.
When I first read k1’s analysis, I thought this one was probably going to be bad. But the next second I said, “The shooting angle and tone reminded me of David Fincher’s movies.” I also deduced that the scene in the screenshot was a scene from “The Seven Deadly Sins.”
It's really too strong...
Even if you throw some obscure memes to k1, it can still explain the funny point in a serious manner.
Although it is a bit over-interpreted, the general meaning is basically understood.
Let's put it this way, based on k1's visual and reasoning abilities, answering questions is basic. As long as your brain is big enough, you can unlock more ways to play.
This ability of k1 is largely due to a person calledCOT (Chain of Thought) thinking chaintechnology.
The general meaning is that before the model outputs the answer,Imitate the way the human brain thinks, break down complex tasks and then solve them step by step.This technology can make the model's IQ higher.
On the other hand, with the help of reinforcement learning technology, the model also learns to evolve in the process of continuous trial and error to achieve optimal results.Just like training a dog.
As for why Kimi took the lead in choosing mathematics as the entry point for the reasoning model, I think it is the same reason as us humans to learn mathematics well and exercise our thinking.
On the basis of "learning mathematics well" in the model, we can then apply this logical reasoning ability to physics, chemistry, and even all aspects of our daily lives, until we finally truly understand the world.
And obviously,Generalization ability of Kimi inference modelIt's already starting to show.
Under the premise that the data has peaked, this path based on reinforcement learning technology may allow the model to achieve better results.
But in the final analysis, what technologies are used in the model and how high its paper score is?In fact, everyone is more concerned about whether the model is easy to use and practical.
Kimi, which has always been good at long texts, now focuses on both long texts and reinforcement learning, which is also a manifestation of adjusting its tool attributes to slowly move closer to user needs.
After all, when technology is no longer superior and can help people solve practical problems, it will truly complete its mission.