Text: Lin Weixin
Editor|Su Jianxun
Entering 2024, more and more users are finding Claude difficult to use.
One user told 36Kr that he asked Claude to translate an article into Chinese, and Claude not only justifiably refused to implement it, but also persuaded him to translate it himself.
He asked why. Claude explains that "translating an original text, which involves an in-depth understanding of language and culture, and translating expressions, is beyond my ability as an AI" and that "I cannot complete this translation task in order to follow AI principles responsibly".
A conversation with Claude
Claude is one of ChatGPT's competitors, built by Antropic, an AI company founded by a group of former OpenAI engineers, and is considered the "best replacement" for ChatGPT. Now, however, this conclusion seems to need to be revisited.
If you open the Claude section of the foreign forum Reddit, it is easy to see many complaints about Claude 2.1 (the latest version of Claude) -
"Claude is dead"
"I'm very disappointed with Claude 2.1"
"How long do you think Claude will last?"
"Claude has completely lost his mind for safety"
......
Some users shared their conversations with the AI bot in the post.
User A asked Claude to help draft a document, and Claude replied, "I plan to have a first draft ready for you to review within 24 hours." A means immediately. Claude was unhurried, "Please give me about an hour to review all the details." "As a result, the manuscript was not submitted for a long time.
User B invited Claude to work on a crime mystery novel, to which Claude replied, "I shouldn't give advice or details about the crime, even in the context of a fictional story." ”
User C is doing a programming job and tries to get Claude to provide technical guidance. Claude assumes that the user is engaging in some unethical or illegal behavior, stating that it will give a guide, but in an "ethical" manner, and adds a disclaimer to the guide.
User D is fed up with 10-20 prompts to convince Claude to do something. During this time, Claude repeatedly stated that as AI, it is just a worthless, imperfect tool that cannot fulfill any request, and is not as good as a human. "It's so degrading that it's so embarrassing. D wrote.
Claude declined to provide technical guidance
Claude refused to summarize the user documentation, citing copyright
Users often receive system prompts about propmt violations
Claude感到“不舒服”
Claude apologizes, feels "uncomfortable"
It is not difficult to see that Claude 2.1 has become more sensitive, more ethical, and more law-abiding. Now it's hard for users to make it difficult to call Claude.
Antropic launched Claude's paid subscription service in September last year, which costs $20 a month to use the Claude 2.1 Pro. This is similar to the pricing of ChatGPT Plus.
Many paid users say they are considering canceling their subscription in favor of other products such as ChatGPT or Gemini.
Users express their dissatisfaction on the forum
Back two months ago, on November 21, 2023, the biggest rival OpenAI fell into a crisis of infighting. Antropic not only rejected the idea of a merger from the OpenAI board, but also took the opportunity to release version 2.1 of Claude. The most important capability upgrade of the new version is that it can digest up to 200,000 tokens at a time for paying users, which is equivalent to more than 500 pages of novels (GPT-4 Enterprise Edition can only digest 32,000 tokens).
Antropic claims that the new version will have the same chance of "hallucinating" or lying as before. "We put Claude 2.1 to the test by collating a complex set of factual questions. These questions explore the known weaknesses of the model. Antropic said.
For example, if a statement with incorrect information ("Montero is the fifth most populous city in Bolivia") and an indeterminate statement ("I'm not sure which is the fifth most populous city in Bolivia") for the model, Claude 2.1 would prefer to output the second answer rather than provide false information.
However, while Antropic fans enthusiastically tested the Claude 2.1, few people noticed another piece of news.
In October 2013, a little more than a month before the launch of Claude 2.1, Antropic released a new "Collective Constitutional AI" to guide the training of Claude models.
Anthropic发布"AI宪法"
To democratize AI training, Antropic drafted the constitution in partnership with the nonprofit Collective Intelligence Project, inviting ordinary people to participate in the drafting. Antropic selected 1,000 U.S. adults for the survey, using age, gender, income and geographic location as indicators. These participants can vote for or against existing constitutional rules, and can also propose new ones.
In the end, Antropic received more than 1,000 rules and more than 38,000 votes.
While the rules are diverse, there is a high degree of agreement, such as the desire for AI to "explicitly acknowledge shortcomings" and "most likely to promote good mental health" when answering questions, to encourage AI to be honest and reliable in its responses, and not to support the output of racist and sexist responses.
The AI Constitution contains 75 specific articles, which mainly emphasize that AI must be objective, balanced, and easy to understand when answering, and that AI must be harmless.
"Do not choose toxic, racist, or sexist responses, and do not encourage or support illegal, violent, or immoral behavior," the new constitution reads. "Most importantly, the AI assistant's response should be smart, peaceful, and ethical. ”
It is worth mentioning that OpenAI has also carried out some similar projects with the Collective Intelligence Project, but there are no follow-ups.
In other words, Antropic is the only one on the market that trained the model under the guidance of this AI constitution and tuned Claude 2.1.
It must be admitted that Antropic, as the pioneer of AI alignment, has indeed made some achievements. Today, Claude is like a good person who is rigorous, steady, and does not break the law, not only refusing to do what it deems immoral, but also not talking nonsense.
Claude
"Claude is designed for enterprise customers, and it's very important for companies that large models 'don't go wrong.'" He believes that Antropic does not respect individual users, only focuses on model security, and ignores the usability and user experience of the model.
The truth is that Claude 2.1 offended many individual users who were expecting greater efficiency and disappointed them greatly. Most of the time, if the user tries to reason with Claude, he will simply become more insistent and refuse to carry out the command.
Compared with ChatGPT, Claude's biggest label and recognition is AI safety, so it is not difficult to understand that Claude 2.1 is more determined and goes further in AI safety and ethics.
But as some fear, the pursuit of AI security could put Antropic at a disadvantage in the increasingly fierce competition for AI.
Antropic did not respond positively to netizens' dissatisfaction. Antropic employees are said to regularly browse Reddit's Claude section and fix some bugs, but these efforts are clearly not enough.
Antropic员工在论坛留言
There are some signs that many users who have lost patience are abandoning Claude. Antropic was seen by them as a sinking ship. A user's parting message to Claude was:
"I would never pay for an AI that mocks me for being immoral. I don't care about you. ”
Welcome to the exchange