When will a company catch up with OpenAI? This question must be the confusion that has lingered in the hearts of many readers for more than a year.
If there is only one company in the world that can catch up with OpenAI, Google should be the most promising player.
Google, which is also a North American AI giant, has the same AGI goals, world-class technical talents, and global financial resources as OpenAI, and even the core Transformer architecture of OpenAI's large model is originally invented by Google.
However, from 2023 to the present, the AI field has been turbulent, and OpenAI will always be one step ahead of Google. Every time Google comes up with a "revenge killer" and wants to be ashamed, OpenAI will always steal the limelight.
For example, the newly launched nuclear bomb-level multi-modal large model GEMINI 15, only "exploded" on the list of science and technology topics for a few hours, and no one cared about it because the SORA launched immediately after it was too explosive and more eye-catching.
There is no doubt that the AI field is going on with the world's most exciting "Fast and Furious", with the leading OpenAI winning beautifully, followed by Google also losing decently. I found that their situation was accurately grasped by the popular movies of the Spring Festival in the Year of the Dragon.
If OpenAI is the amazing and exciting "Hot and Hot", then Google is like a middle-aged racer in "Flying Life 2", who struggles to chase his dreams, plucking up the courage to go full throttle, and the result is not a win, but a big overturn.
The outcome of who will take the holy grail of the finish of the course AGI is still unknown, but it has been incredibly exciting for more than a year just to enjoy the initial stages of this long-distance race.
Google and OpenAI's AI matchup can be described as repeated defeats and repeated defeats. Let's take a look at the overall industrial rivalry situation of North American AI giants from this wonderful "two-hero race".
Google has lost three times in a row, and the North American AI giants are passionately aligned
At present, there are three North American AI giants in the competition for the holy grail of AGI general artificial intelligence: OpenAI, Google, and Meta.
Among them, Meta is taking the open source route, and its large model series LLAMA is currently the most active AI open source community in the world. OpenAI and Google are on the same track, mainly building "closed-source" large models.
Although OpenAI was ridiculed as "no longer open", Google staff also boldly admonished that "we and OpenAI have no moat". However, from another perspective, in order to convince users to pay, closed-source business strategies must provide high-quality models and irreplaceable capabilities, which will also drive model manufacturers to continue to innovate and maintain competitive advantages, which is an indispensable business force in the AI industry.
Therefore, the confrontation pattern of the three giants of AI in North America is the Meta volume ecology, OpenAI and Google volume model.
So, what about the race when you focus on the model track?
Throughout 2023, Google, which is on the same track as OpenAI, has deeply tasted peer pressure.
This race can be divided into three stages:
round1.chatgpt vs bard.
Needless to say, this is a competition that Google started by OpenAI's "peach picking", and since then it can only follow OpenAI's exhaust all the way.
In November 2022, OpenAI released ChatGPT and became a blockbuster, starting a global boom in large language models.
Among them, ChatGPT's basic technology, Transformer, was launched by Google, and the emergence of large language models was discovered by Google researcher Jason Wei (who later switched to OpenAI). Using Google's technology to grab Google's people and impact Google's AI leadership, OpenAI can be described as a "face-riding output".
Google's response was "angry and angry".
In March 2023, Google released Bard on an emergency basis. But the performance of this model itself is relatively weak, the function is limited when it is launched, it only supports English, only for a small number of users, and it can't be played at all with ChatGPT.
round2.gpt-4 vs palm2.
Some people say that Google uses a "Tian Ji Horse Racing" strategy, and deliberately released a relatively weak machine learning model bard in the first game. There is some truth to this, but every horse that can't stand OpenAI is a good horse.
OpenAI soon launched an upgraded version of GPT-4 and opened up the GPT-4 API, leaving Google even further.
The Google IO 2023 conference in May, which was sent out to play GPT-4, is also a "transitional product". Zoubin Ghahramani, vice president of research at Google, said that Palm 2 is an improvement on the earlier model, which only narrows the gap between Google and OpenAI in AI, but does not surpass GPT-4 as a whole.
This round, Google is still lagging behind. Google is obviously aware of this, and at the same time announced at this conference that it is training a successor to Palm, named Gemini, who is betting on hundreds of millions of net worth, and is ready to stage a "prince's revenge" at the end of the year.
round3.Gemini family vs SORA + GPT-5.
At the end of December 2023, Google Gemini "arrived late", Google's most powerful and versatile AI model, which is called a "revenge killer". During this period, OpenAI staged a palace fight drama of "Zhen Huan returns to the palace", and there were no particularly explosive products. Will Google be able to take back everything that belongs to it this time?
Unfortunately, Google has not been able to stage the "return of the dragon king" in the field of AI.
Three sizes of Gemini: Nano, Pro and Ultra, of which Gemini Pro lags behind OpenAI's GPT model in the common sense reasoning task, Gemini Ultra has only a few percentage points advantage over GPT-4, which is OpenAI's product a year ago. Moreover, Gemini was also exposed, claiming to defeat the multimodal ** of GPT-4, with post-production and editing components, training with Chinese corpora generated by Chinese models, claiming to be Wenxin Yiyan.
Google launched the multi-modal large model Gemini 1 a few days after the release of Gemini Ultra5. It can stably process up to 1 million tokens, setting a record for the longest context window.
This is an exciting result, without Sora.
A few hours later, OpenAI launched the text ** generation model SORA, with unprecedented ** generation performance, as well as the productization of the world model, once again amazed the world and snatched the Gemini 15 topics, but also strengthened their AI leadership position. At present, people tend to believe that OpenAI is still ahead of Google.
Previously, everyone speculated that GPT-5 had been almost trained, facing Google's current strongest model, Gemini 15, someone has already shouted Ultraman in the air, asking him how long he is going to cover the baby, and if he doesn't hurry up and release GPT-5.
At this point, the North American AI "Tian Ji Horse Race", which lasted for about a year, came to an end with Google's three consecutive defeats.
The different paths of agi, the Google that is difficult to fly
AGI is a long game. Stretching the long axis, the one-year confrontation between Google and OpenAI and the temporary success or failure may not be much in the future. Being qualified to go to the top track is in itself a proof of Google's AI strength.
Compared with the results of winning or losing, it is more worth discussing that Google has become the "king of volumes" for a whole year, why has it been left behind by openai, and it can't keep up with it?
Tian Ji horse racing, losing once is a tactical mistake, losing every time, you may be able to pay attention to whether there is already a problem from the source of horse breeds, horse rings, forage, etc.
Back to the source, Google and OpenAI can be said to have the same end and different paths.
In the same end, both sides aim to achieve general artificial intelligence and take off the holy grail of AGI;
The difference is that the two sides choose different technical routes. OpenAI uses more general language capabilities as the basis for the implementation of AGI, so it adopts the Transformer architecture, which is crucial to the NLP field, to create a series of GPT models, which has led to the stunning appearance of ChatGPT.
This is not the case with Google. Over the years, Google's AI R&D organization DeepMind has used reinforcement learning and deep learning to solve various artificial intelligence problems, and has accumulated a wide range of technologies. For example, the earth-shattering AlphaGo, the radically changing biology AlphaFold, and NLP technologies such as Transformer.
This is equivalent to two drivers preparing for the race, and OpenAI chooses a venue for AGI, such as "formula car", and then develops and manufactures the model with language as the core, and optimizes (engineering) the structure, length and width, engine, cylinder, etc. of the car (model). Google's Deepmind, on the other hand, was not sure which car would end the AGI race, and had a lot of technical tools at its disposal, so it built formula cars, sports cars, and motorcycles.
There is no superiority or disadvantage between the two routes. However, with the "intelligent emergence" of large language models, it proves that the technical route chosen by OpenAI is more promising to achieve AGI, and Google's Deepmind's technical route has exposed obvious shortcomings:
1.The direction is scattered and costly. The pan-innovation invested in various technical directions consumes a lot of money, and the contradiction between Deepmind and Google's parent company alphabeta in commercialization is deepening. At a time when OpenAI's massive financing is accelerating, Google is saving costs by laying off employees in order to increase investment in AI.
2.There are too many options to focus. Google pioneered many technologies, but the importance and intensity of each technology was also dispersed, and wells were drilled everywhere without water. The most typical is the Transformer architecture, which was invented by Google but carried forward by OpenAI. The emergence of ChatGPT is also after researchers discovered it in Google, but it was not taken seriously and moved forward after leaving OpenAI.
3.The landing is slow, and the results are too slow. Google is also notoriously conservative about AI, resulting in inefficient translation of results, even with advanced technology. A former Google employee once complained that Google's projects are generally bragged about for a while, then nothing is released, and then the project is cut a year later. This can be seen in the explosion of SORA, Google has the corresponding technical reserves and achievements for the diffusion model and Wensheng graph model used to train SORA, but it has not been able to make a product like SORA first.
It can be seen that due to betting on the wrong track at the beginning, when the large language model becomes the most potential implementation path for AGI, OpenAI's leading trend has become a trend. At this time, Google is going to return to the technical track where OpenAI is located, of course, it will be in a disadvantageous position.
One wrong step, one wrong step in the "life", standing up means everything
Frankly, Google is already actively solving problems, including the wrong choice of technology strategy, internal management efficiency, personnel redundancy, and the outflow of AI technical talent.
In April last year, Google merged its two AI "handle" teams, Google Brain and DeepMind, to jointly develop Gemini. In terms of the final result, Gemini's performance is excellent, 1Version 5 is currently one of the most advanced large models in the world. Internal resources are also heavily tilted towards the AI field, and some outflow of AI talents has returned to Google.
The actual actions show that Google's determination and speed to catch up with OpenAI are on the same level after clarifying the track.
But the reality of continuous backwardness also fully explains one point: one's own failure is terrible, and the success of friends is even more worrying.
Although Google has tried its best to solve its own shortcomings and make every effort to promote large models, it cannot stand the acceleration of OpenAI to be stronger.
On the one hand, OpenAI's R&D team can be said to be going all out, while Google's newly merged team still needs to run in. Bill Peebles, the core R&D staff of SORA, once revealed that the team basically did not sleep every day and worked intensively for a year. After the merger of Google Brain and Deepmind, many employees had to give up the software they were familiar with and the original project to develop Gemini.
In addition, compared with Google's remedial recruitment of talents, OpenAI siphons the momentum of the world's top AI talents. As recently as February, Altman publicly stated on social media that "all key resources are in place and very focused on AGI," and that he is looking for talent. In the final analysis, the competition of AI is the competition of talents, because the most important thing in AGI is intellectual resources, and there are only so many top and best talents, which also makes people sweat whether Google can catch up with OpenAI.
In the movie "Flying Life 2", after the protagonist tried racing again and overturned, he did not continue to pursue winning on the field, but as a driver who loves motorsports, he stepped on the track just to prove himself.
Google's matchup with OpenAI cannot be attributed to a simple win or lose. As Google says in Why We Focus on AI (and to What End): We believe that AI can be a foundational technology that will revolutionize the lives of people around the world – and that's what we're all about, and that's what we're passionate about!
All the AI "racers" who have the courage to get off the field and stand on the track deserve applause. And this agi game, which is full of speed and passion, will definitely bring more shocks to us in the audience.