Shulou(Shulou.com)11/24 Report--

Recently, the whole world has ignited a momentum of encircling and suppressing ChatGPT. In addition to the industry leaders, EU countries and the White House have also stepped in. The scary thing, however, is that GPT-4 has quietly acquired the ability to evolve itself.

The threat of GPT-5 has been overshadowed by dark clouds.

After the birth of GPT-4, the possible subversive impact of AI on human society has already made many industry leaders feel frightened.

Recently, it seems that the whole world is encircling and suppressing ChatGPT.

Not long ago by Musk, Turing Award winner Bengio and other bigwigs jointly initiated the suspension of advanced AI research and development of the open letter, now the signature has risen to more than 9000.

Immediately after that, the Center for artificial Intelligence and Digital Policy (CAIDP), a US non-profit organization, asked the Federal Trade Commission (FTC) to investigate whether OpenAI violated consumer protection rules and banned it from issuing GPT-4 further.

Italy has banned ChatGPT directly across the country, giving OpenAI only a 20-day adjustment period, and if it fails to come up with a qualified plan, it will have to pay a fine of up to 20 million euros.

What is interesting about the frenzy of banning ChatGPT across the world is that since Italy started this, other countries are also eager to operate on ChatGPT.

Privacy regulators in France and Ireland have contacted their Italian counterparts for more details about the investigation.

"in principle, it is possible for Germany to adopt a similar procedure," said Ulrich Kelber, Germany's federal data protection commissioner. "

In addition, the European Union has also proposed a draft to regulate ChatGPT as a high-risk AI system.

The United States also stepped in, and just today, US President Joe Biden also held a meeting with science and technology advisers at the White House to discuss the risks and opportunities that AI brings to human, social, and national security.

At a meeting of the President's Council of Advisers on Science and Technology (PCAST), Biden stressed, "in my view, it is the responsibility of technology companies to ensure that their products are safe before they are made public," Reuters reported. "

Asked if AI is dangerous, he said, "it remains to be seen. It is possible." "

Biden said that the harm that powerful technology can do without protection is already clear on social media.

"because of the lack of safeguards, we see that technology affects people's mental health, causing them to feel inferior and desperate about their self-image, especially among young people. "

He again called on Congress to pass bipartisan privacy legislation that would limit personal data collected by technology companies, ban advertising aimed at teenagers, and give priority to health and safety in product development.

Affected by the news, AI's shares fell sharply before Biden called the meeting.

Shares of C3.ai, the ChatGPT concept stock, continued to fall, closing down 26.34% at $24.95, with a market capitalization of $2.8 billion, the biggest intraday drop since the stock went public.

Earlier, C3.ai, like many companies in AI, has become one of the hottest companies on Wall Street since Microsoft announced a multibillion-dollar partnership with OpenAI, and C3.ai 's share price has risen more than 130% so far this year.

Gates: suspending AI research and development will not "solve the challenge." on April 3, local time, Gates publicly talked about the open letter signed by thousands of experts in an interview with Reuters.

Calling for a moratorium on artificial intelligence development will not "solve the challenge". It is best to focus on how best to use the development of artificial intelligence. It is difficult to understand how the suspension of AI research and development on a global scale will be achieved.

'i don't think asking a particular group to suspend work will solve the challenge,'he continued. Obviously, these things have huge benefits. What we need to do is to identify difficult areas.

Gates is a staunch supporter of artificial intelligence.

On March 21, the day before the launch of the joint open letter, Gates published a blog post entitled "the Age of artificial Intelligence has begun."

It is mentioned that artificial intelligence should be used to help reduce some of the world's worst inequalities.

In his blog, he praised artificial intelligence: "the era of AI has begun, and the revolution of artificial intelligence is as important as the birth of mobile phones and the Internet. This is the only revolutionary technological progress in his life." "

So far, Microsoft has invested billions of dollars in OpenAI. OpenAI is building AI technology for Microsoft Bing, Edge, Microsoft 365 and other products.

In an interview, Gates questioned the feasibility of suspending AI development and wanted to know how to implement it.

"the details of any pause will be complex and difficult to implement. I don't quite understand which people they are talking about may stop, whether every country in the world will agree to stop, and why, there will be many different opinions in this field. "

GPT-4 has learned to "self-reflect" and its performance has soared by 30%, but even if the world is facing an enemy of super AI, it may be too late to pull back from the brink of the precipice.

Because, today's AI, has learned to self-evolution!

There is evidence that GPT-4 is improving itself!

Some time ago, Eric Jang, a former Google brain research engineer, found something crazy: GPT-4 was able to criticize his work in a reasonable way.

Instead of looking for perfect tips for a large language model (LLM) (let's think about it step by step), let LLM evaluate its output and correct its mistakes immediately.

Even OpenAI research scientist Andrej Karpathy, former AI director of Tesla, praised it as a good example that we haven't seen the best of GPT-4 yet.

In the example given by Eric Jang, I saw Microsoft researchers tweet that GPT-4 could not write "non-rhyming" poetry at all.

However, when you ask GPT-4 if he has completed the task. At this time, after it apologized, it generated a poem that did not rhyme! It can be said to be a full mark.

This just confirms that GPT-4 has the ability of "reflection".

Recently, researchers from Northeastern University, MIT and other institutions proposed Reflexion. This method gives agents the ability of dynamic memory and self-reflection.

Paper address: https://arxiv.org/ abs / 2303.11366 in order to verify the effectiveness of the method, the researchers evaluated the ability of agents to complete decision-making tasks in AlfWorld environment and the ability to complete knowledge-intensive, search-based question answering tasks in HotPotQA environment.

The success rates in these two tasks were 97% and 51%, respectively.

The overall architecture of Reflexion agent

Some netizens said that you can reflect on "Why are you wrong by asking GPT-4?" "and generate a new prompt for yourself, taking into account the cause of the error until the result is correct, improving the performance of GPT-4 by an amazing 30%.

In addition, the paper points out that Reflexion+GPT-4 is also significantly better than conventional GPT-4 when testing the model's ability to write code.

In this way, as Eric Jang said,

We can now consider a "suffix prompt project" instead of an ingenious "prefix prompt project". This will encourage LLM to find corrections and inconsistencies in the previously generated solution.

ChatGPT has a "hand". If GPT-4 can reflect, with a brain, it would be scary enough, but now ChatGPT has a "hand", not only with tools, but also with models.

Before ChatGPT launched the plug-in function, Meta researchers threw out a new paper, Toolformer, in February, aiming directly at the "weakness" of ChatGPT, allowing large language models to learn to use tools on their own.

Address: https://arxiv.org/ abs / 2302.04761 according to Meta, a large language model like ChatGPT can complement what it lacks.

For example, if you can't count, take out a calculator; if you need the latest information, connect to the search engine yourself; and you can even throw it what you want to translate.

As soon as the paper became popular, many AI researchers said that this paper opened up ideas for people so that the big language model could be the starting point for everything in the future.

Similarly, the recent paper "Jarvis" HuggingGPT published by Zhejiang University and Microsoft Asian Research Institute has attracted a lot of attention in the industry.

In the https://arxiv.org/ abs / 2303.17580 paper, the HuggingGPT proposed by the researchers uses ChatGPT as a controller to connect various AI models in the HuggingFace community to complete multimodal and other complex tasks.

Some netizens said that I strongly doubt that the first general artificial intelligence (AGI) will appear sooner than expected. It will rely on "glue" AI to intelligently glue together a range of narrow artificial intelligence and utilities.

Write code automatically recently, a developer worked on a project that allows GPT-4 to generate code automatically-Auto-GPT.

You can even execute code. In subsequent improvements, Auto-GPT can write code in GPT-4 and execute Python scripts. This makes it possible to recursively debug, develop and improve itself.

Auto-GPT gives people a better understanding of the power of GPT-4, and some netizens even made an outline:

In my opinion, the next frontier area of the hint project is "AutoGPT", says Andrej Karpathy. A GPT call is like an instruction on a computer. They can be concatenated into programs. Use prompts to define I / O device and tool specifications, define cognitive loops, page data input and output in context, and then run.

Netizens: when the singularity comes, some netizens say that this is basically the definition of the singularity.

The only thing an AI that can improve itself now is to enable it to modify the code of the model and train on its own (by browsing the Internet and tagging everything itself).

Some netizens even think that with the rapid progress of technology, we have reached the point where there is no turning back. By 2025, the world will change dramatically, for better or worse.

Learning from mistakes is the foundation of intelligence.

In this way, GPT-4 beat himself and wanted to become a super AI.

According to the latest news from a boss, OpenAI is well aware that the recent rapid iteration of their AI product has caused extremely strong response, and they also welcome public scrutiny.

The person in charge said that risk per token has not yet reached a tipping point, so the abuse of GPT has not done too much harm. A roadmap for AI alignment will be released in a few months.

Reference:

Https://www.youtube.com/watch?v=5SgJKZLBrmg

Https://www.reuters.com/technology/bill-gates-says-calls-pause-ai-wont-solve-challenges-2023-04-04/

Https://www.reuters.com/technology/biden-discuss-risks-ai-tuesday-meeting-with-science-advisers-2023-04-04/

Https://www.reuters.com/technology/biden-discuss-risks-ai-tuesday-meeting-with-science-advisers-2023-04-04/

Https://www.cnbc.com/2023/04/04/italy-has-banned-chatgpt-heres-what-other-countries-are-doing.html

Https://biz.crast.net/bill-gates-weighs-in-on-proposed-ai-pause-wont-solve-challenges/

This article comes from the official account of Wechat: Xin Zhiyuan (ID:AI_era)

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.