Hello everyone, I am Ergou
With OpenAI’s official announcement the day before yesterday that Sam Altman will return and continue to serve as CEO, the OpenAI “gong fight incident” has come to an end.
However, for Ergou, who specializes in eating melons, there is still a core issue that he still hasn’t figured out:
Why was Sam Altman fired from the board?
There had been various speculations on the Internet before, but none of them seemed to have obtained the stone hammer.
Until yesterday, the latest news from Reuters came:
“Four days before Sam Altman was fired from OpenAI, several researchers sent a letter to the board warning that a powerful AI discovery (Project Q*) could threaten all of humanity. Sources revealed that the previously unreported letter The letter and AI algorithm breakthrough were among the reasons the board fired Sam Altman."
Some insiders believe that the Q* project may be OpenAI's key breakthrough in AGI. The new model behind the Q* project is capable of solving some mathematical problems (albeit only to the level of an elementary school student), but the researchers believe mathematics is at the forefront of generative AI development. Currently, generative AI is good at writing and language translation by statistically predicting the next word, but the answer to the same question can vary widely.
And the ability to conquer mathematics with only one correct answer meansAI may have stronger reasoning capabilities similar to human intelligence.
So why is this one of the reasons why the board fired Sam Altman?
It is reasonable to speculate that several board members of OpenAI, such as chief scientist Sutskever, Tasha McCauley, a former executive of multiple technology companies, and Helen Toner, director of the Strategic and Basic Research Fund of the Center for Security and Emerging Technologies (CSET) at Georgetown University, are " Believers of "effective rationalism" can be simply understood as "AI conservatives". Their priority is to ensure the creation of AI that is beneficial to all mankind and to ensure that AI cannot threaten mankind. This is also the mission of the OpenAI board of directors.
Helen Toner once said that even if something happens that leads to the dissolution of OpenAI, it doesn't matter. The mission of the board of directors is more important.
And Sam Altman is an AI accelerator. Altman believes that AI will not get out of control. His first priority is to get OpenAI more financing and make better commercialization to make money. After all, the GPT series of large models are too expensive, and this is the only way. Only by doing this can you ensure that AGI will be built slowly later.
It is possible that the breakthroughs behind the Q* project are considered by several board members to be potentially threatening to humanity, so the pace of development needs to be slowed down and AI safety and alignment issues prioritized, which is what Sutskever has been working on in recent months. to do.
Altman and several board members are not aligned directly on AI safety issues, and Altman's long-term commercialization route has relatively large differences with several other board members.
This may be why several board members want to remove Altman at all costs.
OK, the above are just reasonable guesses. The real reason for Altman's removal needs to be further officially revealed. Let's continue to take a look at what this Q* project is?
Q* project background and more information exposed
According to The Information and people familiar with the matter, Sutskever, OpenAI’s chief scientist, has been working on how to let language models like GPT-4 solve tasks involving reasoning such as mathematics or science for many years. In 2021, he launched a project called GPT-Zero. The name is a tribute to DeepMind’s chess master AlphaZero.
Earlier this year, the project led by Sutskever achieved a technological breakthrough and was able to "produce" its own data - in theory, it can use computers to generate infinite high-quality data like AlphaZero's self-playing game. This overcomes the problem of how to obtain high enough data. There are limitations in quality data to train new models, because it is understood that OpenAI has almost trained on data publicly available from the Internet and can no longer obtain more data for the next stage of training.
Abacusai CEO Bindu Reddy quoted the news on Twitter:
As suspected, OpenAI invented a way to overcome the limitations of the training data using synthetic data, and when trained with enough examples, the model started to summarize well!
Good news for open source and decentralized AI – we are no longer beholden to data-rich corporations.
Two researchers, Jakub Pachocki and Szymon Sidor, used Sutskever's research results to develop a model called Q* to build a system that can solve basic mathematical problems, which has always been a problem for existing AI models.
If you look at the name alone, Q* may be related to the Q-learning algorithm in reinforcement learning. This is a method of evaluating how well an AI takes a specific action in a specific situation, and is used to guide the AI to make decisions in different situations. Optimal decision making.
But it is more likely that Q is just a code name. Reddit users broke the news and speculated that Q has more abilities:
-
The model behind Q* may already be capable of autonomous learning and self-improvement.
-
The model behind Q* is capable of making complex decisions in a wide range of scenarios and may already be slightly self-aware by assessing the long-term consequences of its actions.
AI already has slight self-awareness?
This sounds too "nonsense"! I don’t even believe Ergou who only went to elementary school.
After all, countless scientists have not made any breakthroughs in the difficult problem of consciousness, and they are still in the stage of philosophical discussion and neuroscience exploration.
But just a month ago, OpenAI chief scientist Sutskever said in an exclusive interview with MIT Technology Review: "ChatGPT may be conscious." The following is quoted from the MIT Technology Review report:
Ilya said he does not intend to build the next GPT or DALL-E, but rather to figure out how to prevent super artificial intelligence from becoming uncontrolled. As a believer in futurism, he believes that this still hypothetical future technology will eventually appear.
He thinks ChatGPT may be conscious. He also believes people need to realize the true power of the technology that OpenAI and other companies are racing to create. He believes that some people will choose to merge with machines in the future.
ChatGPT, he said, has changed many people’s expectations of what’s to come, from “never going to happen” to “moving faster than imagined.”
Before predicting the development of general artificial intelligence, which refers to machines as smart as humans, he said: "It's important to talk about where it's going. At some point, we're really going to see general artificial intelligence. Maybe. OpenAI will build it, and maybe other companies.”
Is the big data paradigm just a stopgap measure?
Various celebrities on Twitter discussed the above incident.
Jim Fan, senior artificial intelligence scientist at Nvidia, said:
It is clear that synthetic data will provide the next trillion high-quality training tokens. I bet the vast majority of big model teams know this. The key issue is how to maintain data quality and avoid plateauing conditions.
Richard Sutton's painful lesson continues to guide the development of AI: there are only two paradigms that can scale infinitely with computing, and that is learning and search. This was true in 2019, it is true today, and I bet it will be true until the day we solve the AGI problem.
Musk said: Yeah, it's kind of sad that you can store the text (the amount of information contained) of every book ever written by humans on a hard drive. But there will be infinite amounts of synthetic data.
Perplexity AI CEO pointed out: Tesla has used synthetic data for training, which is the so-called automatic labeling project.
However, Turing Award winner Yann LeCun believes that the big data paradigm is just a stopgap measure:
Animals and humans can quickly become very smart with very little training data. My money is on new architectures that can learn as efficiently as animals and humans. Due to the limitations of our current methods, using more data (synthetic or non-synthetic) is a temporary expedient.
For example, parrots, dogs, and octopuses have about 2 billion neurons. How do we make a machine with only 2 billion neurons/10T parameters become like octopuses, dogs, parrots, and crows through several months of real-time training data? Just as smart?
Some netizens posted: Aren’t millions of years of human evolutionary adaptation similar to pre-training, and our lifetime experience similar to continuous fine-tuning?
LeCun said this data is insufficient:
An AI researcher responded to LeCun:
We humans also use large amounts of data to train. You forget that we receive massive amounts of video, audio, and sensor data all the time, not to mention DNA-encoded "instructions." We are not trained from scratch, and our output is much more general than that of large language models;
Also I agree with you about the new architecture.
Lecun made a rigorous calculation:
Eduardo Slonsk was convinced by Lecun:
Lecun concluded: “Current large language models are trained on text data that would take a human 20,000 years to read.But they still don’t know if A is the same as B, then B is the same as A (reverse curse).With relatively little training data, humans will become smarter. Even crows, parrots, dogs and octopuses can be very, very fast The earth is getting smarter than this, with only 2 billion neurons and trillions of "parameters."
Are big language models the path to AGI?
Not long ago, Sam Altman said in an interview with the Financial Times:
-
Although OpenAI has been successful with ChatGPT and user usage, neither ChatGPT nor the GPT store is the real product that OpenAI wants to build. The ultimate goal is to build general artificial intelligence; the large language model (LLM) behind ChatGPT is just to build general artificial intelligence." One of the core parts;
-
In the race to develop general artificial intelligence, the “biggest missing piece” is the fundamental “leap of understanding” that such AI systems would need to make.
-
For a long time, Newton's regular approach was to read more mathematics textbooks, talk to professors, and practice problems (which represents the big data training paradigm behind it); but Newton would never invent micros just by reading geometry or algebra. Points (but need to find a new paradigm), the same is true for OpenAI to implement AGI;
This incident has also been discussed in China. Xie Lingxi, a famous Zhihu V and a doctor of Tsinghua University, wrote a very sharp article with wonderful views:
-
To realize AGI, it is very unlikely to achieve a breakthrough in just one algorithm.
-
Currently, there is no interactive environment in the industry that can train computer vision algorithms like ChatGPT. 1.** In order to realize a truly large visual model, a visual interaction environment such as dialogue must first be established. **
-
Unless we see OpenAI's robots running around the streets one day, interacting with humans to collect data, or OpenAI making a rich enough virtual environment that can simulate various specific tasks; otherwise I don't believe that the ChatGPT paradigm can be transferred well. Come into the visual world.
Xie Lingxi then added some background knowledge to explain:Any technological leap is often not a breakthrough at a single point, but the result of the accumulation of multiple technologies.
The current significance of studying network architecture design or self-supervised learning algorithms is far less significant than designing a real world model (or providing an implementation method for a sufficiently complex interactive environment). Only by realizing the latter can we see substantial progress in AGI.
Professor Ma Yi also updated an update on Weibo, expressing that we have only just begun to understand the nature of intelligence.
