Early this morning, OpenAI announced the latest version of its large-scale language model, GPT-4. The company says GPT-4 outperforms the vast majority of humans on many professional tests.
Specifically, GPT-4 achieves leaps and bounds in the following areas: smarter and better problem solving; image input support and powerful graphical recognition, but currently limited to internal testing; longer context and a text input limit of 25,000 words; significantly better response accuracy; and more security and less harmful information.
For ordinary people, how to understand how strong and smart GPT-4 really is?
According to OpenAI, GPT-4 passed all the basic exams with high scores. For example, the GPT-4 ranks in the top 10% of test takers on the mock bar exam, in the top 7% on the SAT reading test, and in the top 11% on the SAT math test. In contrast, the GPT-3.5, which was once shocking, has a real score in the bottom 10%, and the GPT-4 is already predictably strong.
One user commented, “It would be terrible if it was really like the report says, I feel the ability has far surpassed me.” Others laughed and said, “I’m laid up! I was born in the right time for me!”
The most frequent voices were worried about the future of job insecurity, “What other industry can be irreplaceable?” “Eliminate any illusions that the AI era is sweeping the world and think about what you can do that GPT-4 can’t do.”
“Bigger” than previous versions, and perhaps more expensive
OpenAI says GPT-4 is “bigger” than previous versions, meaning it has been trained on more data and has more weights in the model file, which makes it more expensive to run.
At the same time, OpenAI did not announce how many parameters the model has.
OpenAI said it uses Microsoft (MSFT.US) Azure to train the model, and that Microsoft has invested billions of dollars in the startup. openAI did not release details on the exact model size or the hardware used to train the model, citing “competitive conditions.
Many researchers in the field now believe that many of the latest advances in artificial intelligence have come from running increasingly large models on thousands of supercomputers, which can cost tens of millions of dollars to train.
OpenAI is also building an “infrastructure” for its research and development in order to keep improving GPT-4 performance.
Over the past two years, OpenAI has rebuilt its entire deep learning stack and worked with Azure to design a supercomputer from scratch for its workloads. A year ago, OpenAI ran the supercomputer for the first time while training GPT-3.5, and they have since found and fixed bugs and improved its theoretical foundation. The result of these improvements was an unprecedentedly stable training run of GPT-4.
Greg Brockman, co-founder and president of OpenAI, said OpenAI expects that in the future, cutting-edge models will be developed by companies investing in billion-dollar supercomputers, and that some of the most advanced tools will come with risk. breathing room to really focus on security and do it right.”
With obvious limitations, GPT-4 still isn’t completely reliable
Despite its power, GPT-4 has similar limitations to earlier GPT models, not the least of which is that it’s still not completely reliable, meaning it’s still possible to spout nonsense.
OpenAI also warns that GPT-4 is not perfect, and in many cases it is less capable than humans. The company says, “GPT-4 still has many known limitations that we are working to address, such as social bias, hallucinations, and adversarial cues.”
Overall, GPT-4 has significantly mitigated the hallucination problem relative to previous models (which have undergone multiple iterations and improvements). In OpenAI’s internal adversarial realism evaluation, GPT-4 scored 40% higher than the latest GPT-3.5 model.
Also, GPT-4 training data is still available as of September 2021. This also means that GPT-4 does not know enough about information beyond this point in time and does not learn from its experience.
After the release of GPT-4, OpenAI founder Alterman tweeted, “It’s still flawed, still limited, and seems more impressive when you spend more time using it for the first time than it actually is.”
The most shocking thing is that, according to the OpenAI engineers in the demo video, the training for GPT-4 was completed last August, and the rest of the time was spent fine-tuning enhancements and, most importantly, removing dangerous content generation.
This also means that OpenAI’s internal technology is more years ahead of the outside world than everyone thinks. What’s even scarier is that OpenAI has also opened up the API interface and related papers in one go!
Wang Sheng, a partner of Enoch Angel Fund, told “State of the Art”, statically speaking, maybe in two or three years, I think it’s already very optimistic.”
But with the release of the more powerful GPT-4, it’s clear that OpenAI’s technical capabilities are still improving and progressing much faster than we can catch up. Wang Sheng judged in a previous interview, “Unless this thing suddenly hits a bottleneck and the whole direction of technology development comes to a head, and the other side has to stop, maybe we still have a chance to catch up, or we tap into a new technology path that has a chance to catch up, otherwise there may be no chance.”
Meanwhile, OpenAI also has an update on the landing and application of the model.
The new model will be available to paid subscribers of ChatGPT and will also be available as part of an API that allows programmers to integrate AI into their applications. openAI will charge 3 cents for about 750 words of command information and 6 cents for about 750 words of response information.
OpenAI also introduced Morgan Stanley, which is using GPT-4 to organize data, and electronic payments company Stripe, which is testing whether GPT-4 can help combat fraud. Other customers include language learning company Duolingo, Khan Academy and the Icelandic government, and OpenAI partner Microsoft said Tuesday that a new version of its Bing search engine will use GPT-4.
GPT-4’s achievements are exciting, but the Chinese ChatGPTs, who are still on the starting line and have a lot of work to explore and research, must be feeling a lot of pressure. The bigger the storm, the more expensive the fish, and there will be tremendous opportunities for these companies that are entering this new field quickly.
