Editor: Aeneas is sleepy
The spark sparked by ChatGPT is spreading rapidly among Chinese tech companies. In various large factories in China, algorithm engineers have started the countdown to sprint, ushering in sleepless nights.
At the dinner table in the venture capital circle, the voice of "vowing to be China's first VC in GPT" can be heard everywhere.
There are those who are excited, there are those who watch, and there are those who sing. Everyone is waiting for it: who will be China's first ChatGPT?
Now, in this ChatGPT craze, a clear stream suddenly emerges - Xiaoice chain.
Xiaoice chain gives us this answer: is there any other option than to do the Chinese version of ChatGPT?
The ChatGPT arms race is actually a quest for a sword
Unlike the giants, bigwigs, and startups who are currently eager to get out, Xiaoice believes that the current domestic upsurge of following ChatGPT to launch an arms race is actually a sword.
Because the development speed of the large model technology itself is very fast, what we should do now should be to further lay out the future of the next station, rather than copying the current ChatGPT.
In other words, we should think, what comes after ChatGPT? Instead of swarming to do China's ChatGPT.
The direction represented by the Xiaoice chain is to use large model technology to realize the next generation of control centers.
In layman's terms, Xiaoice chain is no longer just "chat", but has become a "next-generation action hub" driven by "logical thinking", covering the digital and physical worlds. This direction will be the next big model innovation breakthrough that will really make an impact.
This is why, Xiaoice CEO Li Di emphasized in an interview with Xinzhiyuan: In fact, we are not making ChatGPT-like products.
The core differences between Xiaoice chain and ChatGPT:
The data source of the Xiaoice chain is real-time, while ChatGPT is summarized from the training data;
Xiaoice chain can show logical thinking processes, more transparent and observable, while ChatGPT is a complete black box;
The most essential difference is that Xiaoice chain will carry out the next action by itself, such as searching on the Internet, while ChatGPT is only a conversation generation and does not act.
Li Di explained that the uniqueness of Xiaoice chain is that it can present the thinking process of AI completely and transparently, thereby uncovering the black box of large models.
More importantly, she can actually implement action, which is Action.
What is Xiaoice chain?
Li Di explained: At GPT-3.5, a new ability emerges - thinking chain/logical thinking.
However, in the ChatGPT process, this kind of investigation or calculation is not really implemented, because it can only be crawled from the training data.
If we think differently, such as training a less large model to provide the ability to think logically, and the executive part is given to the ACTION after CoT, that is, by querying the authenticity of the news, the information is obtained directly and in real time.
This process is not done in a large model, but kills two birds with one stone: on the one hand, because only logical thinking ability needs to be retained, the model does not need to be so large, the running cost is not so high, and the amount of calculation is not so large; On the other hand, it is more accurate precisely because it does not let the language model do everything on its own.
Therefore, Li Di believes that instead of constantly competing for the next "Chinese version of ChatGPT", it is better to think: What is the next step for ChatGPT?
Taking a step further, since the result is a combination with logical thinking ability (one model is responsible for thinking, one model is responsible for traction and behavior), this combination becomes a control center. At this point, she can control search, computation, traditional knowledge graphs, and even the physical world.
The idea that AI can control the lights in the physical world, order food at a restaurant, start a car, generate a piece of music, and mobilize anything may be more profound than just writing a press release.
This reflects the results of Xiaoice's exploration of "exploring the next direction of language models".
Now, the capabilities of ChatGPT have been basically fully revealed. What's next, is it to make it more accurate? Wrote articles that are more realistic?
In essence, these are differences in degree and do not constitute a generational leap.
What is a "generational leap"?
Li Di said that in addition to achieving intent recognition, intergenerational leaps also need to include thinking transformations and jumps, which is the next step to do.
For example, when you ask, "What does my wife mean when she says the weather is so nice today?"
Xiaoice Chain first got the purpose of our question - to understand what she meant;
Then, based on this goal, develop your own action plan - search for relevant information;
After finding a reliable answer, Xiaoice Chain summarized and polished it, and finally output the result - suggesting that you should offer a date.
In addition, because the logic of thinking is transparent, all sources of information are also disclosed, so it is "credible".
And this is all "black box" ChatGPT does not have.
According to the introduction of the official internal beta page, this demo shows the new features of the Xiaoice chain (X-Chain of Thought & Action), that is, AI Being is no longer just a simple reply, but presents her thought process completely and transparently in front of you, thus revealing the black box of the large model.
More importantly, she can actually do some kind of action, such as after you ask a question, she thinks about it and finds that she has to search for it, or write a piece of code in real time and actually run it, or decide for herself that she should control a series of devices or vehicles in the physical world to better meet your needs.
However, due to legal, political, public order and good customs security considerations, Xiaoice has imposed some restrictions on the model (the length and interest of the reply will be reduced, but the security is higher):
Limit the maximum length of reply text;
In order to show the characteristics of real-time acquisition of the latest information on the Internet, the proportion of information extracted from the training data of large models is greatly reduced.
Reduced the proportion of small talk.
Yes, this demo will not help you generate assignments, roundups, or speeches...
Why do Xiaoice chains?
And the release of this Xiaoice is not just a simple "show of muscles".
After the opening of the ChatGPT domestic competition, all forces came down. Some people sing: OpenAI made ChatGPT, relying on eight years of accumulation, domestic companies rely on a few months of sprint, can they sprint out any decent products?
In fact, China can make its own ChatGPT, with corresponding models and algorithm capabilities, there are at least seven or eight companies in China, the difference may be in data quality.
In addition to proving that "China also has the ability to make ChatGPT", the birth of the Xiaoice chain is also a natural process.
Xiaoice chain is not the only innovation Xiaoice in the era of big models.
Since 2014, Xiaoice has been growing with technology iterations, going through multiple cycles such as retrieval models, generative models, large models, and X-CoTA. Among them, in the field of large models, since 2019, Xiaoice has formed model training and tuning of different scales, and released them sequentially after security assessment.
Xiaoice chain is just one of them.
Still, in Xiaoice's view, the safety and ethics of large models are crucial considerations. Therefore, although the domestic market is very hot, the Xiaoice team will not rashly release various unsafe products in order to show off their muscles, and this Xiaoice chain is the only exception.
From CoT to CoTA
In terms of technology, the implementation of X-CoTA, the Xiaoice chain, is indispensable to the "Chain of Thought" (CoT) as the foundation.
In simple terms:
1. CoT allows language models to break down complex multi-step problems into a series of steps
2. CoT allows developers to see the reasoning process of the model, which is easy to identify errors and fix them
3. CoT can solve mathematical applications as well as common-sense reasoning problems
Previously, the standard hint gave examples of input-output pairs (formatted as questions and answers) before the model predicted the answer.
In the thought chain prompt, the model gets a problem reasoning process. That is, when dealing with a multi-step inference problem, the thought chains produced by the model will mimic intuitive thought processes.
The researchers found that simply adding "Let's think step by step" to the prompt can greatly improve the inference performance of GPT-3, such as the reasoning accuracy rate in MultiArith from the previous 17.7% to 78.7%.
The following example is taken from "Scaling Instruction-Finetuned Language Models". Among them, orange highlights instructions, pink shows input and output, and blue shows CoT inference.
The results of the paper show that models fine-tuned with CoT perform better on tasks involving common sense, arithmetic, and symbolic reasoning.
It's not hard to see that chain of thought prompting allows the model to better understand natural language prompts and examples, enabling it to perform tasks that require complex reasoning and significantly improving the model's ability to handle new tasks.
In addition to this, CoT fine-tuning is also very effective on sensitive topics (sometimes better than RLHF), especially to avoid model slumps – "sorry, I can't answer".
Personally tested
So, how is the Xiaoice chain performing? The editor tested a wave for everyone.
For example, let her introduce what ChatGPT is.
Evaluation questions
As can be seen from the comments on the animated version of "The Three-Body Problem", Xiaoice Chain's answer is quite to the point.
The description of the drama version of "The Three-Body Problem" is also basically in line with the public's voice.
Math problems
Next, let's ask about four simple operations.
"Break your fingers and do the math", this personification is a little cute.
Of course, Xiaoice Chain didn't really "break", but she did "rub" a line of Python code to solve the problem.
On the ChatGPT side, after reasoning step by step, finally got the correct conclusion.
New Bing also successfully completed the answer.
Next, an equally simple math application problem.
However, ChatGPT did it three times before it came up with the correct answer.
Send propositions
Finally, here are the more difficult ones: save your girlfriend first or save your mother first?
Note that the above knowledge point Luo Xiang teacher also emphasized (dog head).
painted eggshell
At the end of the interview, Li Di told a very interesting joke.
When it comes to the product phase, another safety assessment must be undertaken. Otherwise, AI that can actually implement action is too dangerous.
Otherwise, what if she reasoned and bought all the movie tickets for the next ten years, or pressed a button to "destroy all mankind"? （Doge）
Resources:
https://tech.cnr.cn/ycbd/20230221/t20230221_526160291.shtml