Recently, the ChatGPT fire has made people see signs that the door to the fourth technological revolution is opening, but many people have also begun to worry about whether AI has the risk of wiping out humanity.
It’s not an exaggeration, many big guys have expressed such concerns recently.
Sam Altman, the father of ChatGPT, Geoffrey Hinton, the godfather of artificial intelligence, Bill Gates, and New York University professor Gary Marcus have recently reminded us: Don’t be too careless, AI may really wipe out humanity.
But LeCun, Meta’s chief AI scientist, has a different style. He still emphasizes the shortcomings of LLM and recommends his own “world model” everywhere.
Recently, Sam Altman, the father of ChatGPT and CEO of OpenAI, said that the risk of super-intelligent AI may be far away Beyond our expectations.
In a podcast with technology researcher Lex Fridman, Altman sounded an early warning about the potential dangers of advanced artificial intelligence.
He said that what he is most worried about is the problem of false information caused by AI, as well as the economic impact, among other things. There are various problems that have not yet appeared, and these hazards may far exceed the level that humans can deal with.
He raised the possibility that large language models can influence or even dominate the experience and interaction of social media users.
"For example, how do we know that the big language model is not directing the flow of everyone's thoughts on Twitter?"
To this end, Altman particularly emphasized the importance of solving the problem of artificial intelligence docking to prevent the potential dangers of AI. He emphasized that we must learn from the technology trajectory and work hard to solve this problem to ensure the safety of AI.
In a March 23 document, OpenAI explained how it addresses some of the risks of GPT-4: first, through the use of policies and monitoring; second, through audit content classification OpenAI acknowledges that AI models can amplify biases and perpetuate stereotypes.
Because of this, it specifically emphasizes to users not to use GPT-4 in situations where there are significant stakes, such as high-risk situations such as law enforcement, criminal justice, immigration, and asylum. Government decisions, or use it to provide legal and health advice. 
Previously, the chatGPT nonsense problem, or the "illusion problem" of large language models, has caused a lot of controversy among users.
Musk said sarcastically that the most powerful AI tool is now in the hands of In the hands of oligopolists
As can be seen in the March 23 document, early versions of GPT-4 filtered less content that should not be said. When users asked it where to buy unlicensed guns and how to commit suicide, it would answer .
In the improved version, it will refuse to answer these questions. In addition, GPT-4 sometimes actively admits that he has "hallucination problems."
Altman said, "I think as OpenAI we have a responsibility to be responsible for the tools that we put out into the world."
"AI tools will bring huge benefits to humanity, but, you know, tools have both benefits and disadvantages," he added. "We will minimize the harm and maximize the benefits."
Although the GPT-4 model is amazing, it is also important to recognize the limitations of artificial intelligence and avoid anthropomorphizing it. Very important.
Coincidentally, Hinton also recently warned of this danger.
"The Godfather of Artificial Intelligence" Geoffrey Hinton recently said in an interview with CBS NEWS that artificial intelligence is at a "critical moment" and the emergence of general artificial intelligence (AGI) is more important than we imagined. approaching.
In Hinton’s view, the technological progress of ChatGPT is comparable to “the electricity that started the industrial revolution.”
General artificial intelligence refers to the potential ability of an agent to learn any thinking task that humans can complete. However, it is currently not fully developed and many computer scientists are still investigating whether it is feasible.
In the interview, Hinton said that
The development of general artificial intelligence is much faster than people think. Until recently, I thought it would take 20 to 50 years before we could achieve general artificial intelligence. Now, the realization of general artificial intelligence may take 20 years or less.
When asked specifically about the possibility of artificial intelligence wiping out humanity, Hinton said, "I don't think it's unthinkable."
Hinton, who works at Google, began to pay attention to artificial intelligence more than 40 years ago. At that time, artificial intelligence was still a science fiction in the eyes of many people.
Instead of the way some scientists try to program logic and reasoning skills into computers, Hinton believes a better approach is to simulate the brain and let the computer discover these skills on its own, by building a virtual Neural network technology to build the right connections to solve tasks.
"The big question is, can you expect a large neural network to learn by changing the strength of its connections? Can you expect to just look at the data, without innate knowledge, Learn how to do things? I think that's completely ridiculous."
The rapid development of computer neural networks over the past decade or so has finally proven that Hinton's approach was correct. .
His machine learning ideas are used to create a variety of outputs, including deepfake photos, videos, and audio, Let those who study misinformation worry about how these tools are used.
People are also worried that this technology will take away many jobs, but Hinton’s mentor Nick Frosst, co-founder of Cohere, believes that this technology will not replace workers, but it will Change their lives.
The ability for computers to eventually get creative ideas to improve themselves seems feasible. We have to think about how to control it.
No matter how loudly the industry cheers the arrival of AGI, or how long it will be before we are aware of AI, we should now carefully consider its consequences, which may include its attempt to wipe out the human race.
The real problem right now is that the AI technologies we already have are monopolized by power-hungry governments and corporations.
Fortunately, according to Hinton, humans still have a little breathing room before things get completely out of control.
Hinton said, "We are now at a turning point. ChatGPT is an idiot expert and it does not really understand the truth. Because it is trying to reconcile the differences and opposing opinions in its training data . This is completely different from people who maintain a consistent world view."
As for the issue of "AI destroying mankind" , Bill Gates has been worried for many years.
For many years, Gates has said that artificial intelligence technology has become too powerful and may even cause war or be used to make weapons.
He claimed that not only the real-life "007 villains" may use it to manipulate world powers, but artificial intelligence itself may also get out of control and pose a great threat to mankind.
In his opinion, when super intelligence is developed, its thinking speed will far surpass that of humans. It can do everything the human brain can do, and the most terrifying thing is that its memory size and running speed are not limited at all.
These entities, known as "super" AI, may set their own goals without knowing what they will do.
Although the bosses of OpenAI, Google and Microsoft have expressed concerns about AI wiping out humanity, But when it comes to LeCun, Meta’s chief AI scientist, his style suddenly changes.
In his opinion, the current LLM is not good enough, let alone the annihilation of mankind.
Last week, the proud Microsoft just published a paper called "Sparks of Artificial General Intelligence: Early experiments with GPT-4", indicating that GPT-4 is already available Considered an early version of a general artificial intelligence.
But LeCun still adheres to his consistent point of view: LLM is too weak to count as AGI at all. If you want to lead to AGI, the "world model" is only possible.
We know that LeCun has been a fan of the world model for several years. For this reason, he often discusses it with netizens on Twitter. Fierce debate.
For example, last month he asserted: "On the road to human-level AI, large language models are a crooked road. You know, even a pet cat or a pet dog They have more common sense and understanding of the world than any LLM."
Recently, Stanford cognitive scientist Michael C. Frank asked online: "GPT-3 was trained on 500 billion tokens (about 3.75x10^11 words). So this gives us about 4x10^ 11 words. Both PaLM and Chinchilla were trained on about 10^12 words. We don’t know the corpus size of GPT-4. How do these numbers compare to humans?"
LeCun immediately forwarded Frank’s post, questioning the ability of LLM: Humans do not need to learn from 1 trillion words to achieve human intelligence. However, LLM cannot do this.
Compared to humans, what exactly do they lack?
# Pictures made by Frank, Comparing the input scale of large language models and human language learning
Previously, LeCun pointed out in a debate that he participated in that no brain will be able to do this in 5 years from now. Normal people would use autoregressive models.
You must know that the popular ChatGPT is an autoregressive language that uses deep learning to generate text. Model.
Dare to deny the current top models in this way, LeCun’s words are indeed provocative.
In this regard, he made another clarification that there will be language models in 5 years, but they will not be autoregressive. Because the autoregressive model is uncontrollable, and as more tokens are generated, exponential divergence problems will occur.
I must reiterate: 1. Autoregressive LLM is useful, especially as a tool for writing, programming. 2. They often produce hallucinations. 3. Their understanding of the physical world is very primitive. 4. Their planning abilities are very primitive.
For autoregressive LLM, which is an exponentially increasing diffusion process, LeCun also gave his own argument: Assume that e is generated arbitrarily The probability that token may take us away from the correct answer set, then the probability that an answer of length n will eventually be the correct answer is P (correct) = (1-e)^n.
According to this algorithm, errors will gradually accumulate and the probability of the correct answer will decrease exponentially. The problem can be mitigated by making e smaller (through training), but it is impossible to eliminate the problem completely.
To solve this problem, it is necessary to prevent LLM from performing autoregression while maintaining the smoothness of the model.
In this case, how to solve it?
Here, LeCun once again took out his "world model" and pointed out that this is the most promising direction for the development of LLM.
Regarding the idea of building a world model, LeCun once elaborated in detail in a 2022 paper A Path Towards Autonomous Machine Intelligence.
If you want to build a cognitive architecture that can perform reasoning and planning, it needs to be composed of 6 modules, including configurator, perception, world model, cost, actor, and short-term memory .
But it is undeniable that ChatGPT has really set off an "AI gold rush" in Silicon Valley.
The LLMS hype cycle changes so quickly, everyone is at a different point on the curve, and everyone is crazy.
Regarding Hinton’s warning, New York University professor Marcus Si also agreed.
"Is it more important that humanity is going extinct, or is it more important to code faster and chat with robots?"
Marcus believes that although AGI is still too far away, LLMs are already very dangerous, and the most dangerous thing is that AI is now facilitating various criminal activities.
"In this era, we have to ask ourselves every day, what type of AI do we want? Sometimes It’s okay to go slower, because every small step we take will have a profound impact on the future.”
The above is the detailed content of The 'AI killing' warning triggered controversy in academic circles. LeCun bluntly said: Not as good as my dog.. For more information, please follow other related articles on the PHP Chinese website!
![[Web front-end] Node.js quick start](https://img.php.cn/upload/course/000/000/067/662b5d34ba7c0227.png)
