When the founder, who had been behind the scenes for many years, came forward to explain, Google's crisis became clear.
On March 2, local time, Sergey Brin appeared at the "AGI House" in California, USA, and commented on Google's large model Gemini: "We messed up completely in image generation, and I think this is mainly due to not conducting thorough testing. ”
Sergey, dressed in a colorful coat, is 50 years old and has a gray beard. He founded Google in his garage with Larry Page 26 years ago, and in 2019, Sundar Pichai succeeded them as CEO of Google's parent company, Alpgabet, while already being Google's CEO.
In 2023, ChatGPT will set off a wave of AI, Google will be questioned internally and externally for being "slow to respond", and Brin is rumored to have returned to Google to directly participate in Gemini research and development, working together in the office and research institute three to four days a week, and writing ** in person. But direct appearances are rare, let alone admitting that Google "screwed up".
In the past month, Google has become a "model worker" in the AI arena, and the large model has been launched five times in a row. These include updating Gemini with the launch of a paid subscription version of Gemini Advanced (10 Ultra), released Gemini 15 Pro and open source model Gemma, as well as submit the "World Model" Genie**.
But more frequent sorties led to even greater trouble.
The updated Gemini supports the Bunsheng feature, but users quickly discover that it appears to be "deliberately" avoiding the generation of images of white people, whether it's the Founding Fathers of the United States or Elon Musk, who in Gemini's case has to be black.
* Continued fermentation, accusations of "anti-white" gemini spread. Google urgently shut down its Wensheng Graph feature, but discussions continue. Google executives publicly apologized for the incident, and CEO Pichai sent an internal letter to employees saying that the company would carry out "structural reforms."
Looking back, Google's gemini rollover is not an accident. When Google released Bard last year, Bard's answer in the demo** was factually incorrect. As Google's first punch after ChatGPT, the mistake wiped out $100 billion of the company's market value overnight.
In the midst of the controversy, Google's stock price fell 4% in a week, and the outside world not only questioned Google's structural problems, but also questioned Pichai's ability to fight against the wind, and began to call for Pichai ** and Google to change generals.
The AI combination punch that was played outward, before it hit the opponent, had already hit Google's own "seven inches".
Agemini's rollover is not complicated.
Users found that whether they were typing the Pope, Vikings, the Founding Fathers of the United States in 1978, the first woman**, or the King of England in the Middle Ages, Gemini either spit out black images directly, or sandwiched up to one white face in multiple spitts**. Of course, the most outrageous thing is the image of Musk output by Gemini - needless to say that you can guess it - and the skin color is close to that of Martin Luther King.
Frequently calling out with reality and contrary to the facts, Gemini's preference for minorities can no longer be regarded as a preference, but a kind of paranoia. Netizen spicy comment: In Wakanda (the ** city in the movie "Black Panther"), gemini means God's little angel.
In recent years, there has been a lot of controversy about "woke" in the United States, and Gemini was accused of being "anti-white". Musk also posted on his own social media and bluntly said: "Viral awakening thinking is destroying Western civilization." ”
In the wake of the controversy, Google quickly shut down Gemini's Wensheng Diagram feature and issued a statement on the social media platform X, saying that Gemini had inaccuracies in some historical image generation and that it was working to resolve the issue. Later, Prabhakar Ragh**an, Google's senior vice president, apologized for the incident.
This is undoubtedly a huge blow to Google.
Last year, Google released a ChatGPT competitor, Bard, but the response to this product was mediocre. After that, Google ended Deepmind's nine-year internal independence and merged it with Google Brain to form a new division, Google Deepmind. Gemini, on the other hand, has been seen by Google as the "next big thing" ever since. Not only external publicity, but also internal priorities, as mentioned earlier, Google founder Brin has also returned to the front line of Gemini.
At the end of last year, Gemini was released, and Google's large model was accelerated. In February this year, Google made 5 moves in a row, updating Gemini and releasing new models. Among them, Gemini has added Wensheng Diagram, and the newly released Gemini 1The 5 Pro series supports up to one million tokens and is seen as a rival to GPT-5 rather than GPT4. In addition, Google has also released the open-source model GEMMA and published a "world model" by Genie.
But Google encountered two embarrassments, one was in gemni 1On the day of the release of 5 Pro, OpenAI suddenly revealed Wensheng **Sora, stealing the limelight from Google. The second is the ensuing gemini "anti-white" controversy, which made Google overturn. Google's experience is very similar to China's ** person "Wang Feng", who was ridiculed by the group while being a model worker.
It's not the worst thing that the released product rolls over, what's even worse is that Google always rolls overAt the time of the release of Bard last year, it was discovered that there was a glaring factual error in its presentation, which crowned a telescope of an extrasolar planet**.
The worst thing is that in the case of frequent new product releases and frequent overturns, Google's internal problems as an industry giant are becoming more and more prominent.
bA series of big moves is not necessarily a good thing.
In an interesting episode, a few days after Google announced the shutdown of Gemini's Wensheng image feature, Thomas Kurian, Google's cloud computing CEO, released a series of messages on X. However, unrelated to the Gemini controversy, Culian announced that he will launch a new large-scale model service, Gemini Business, which will replace the original Workspace Enterprise's Duet AI.
A netizen sarcastically said: Hello Thomas! Can I pay with Google Pay in Google Wallet? Formerly known as Android Pay, it was previously known as Google Wallet. If not, we can discuss the payment through **. I'll send you a link to Google Meet, the one called Google Chat, which was called Duo before that. Duo replaces Allo, and Allo replaces Hangouts. Oh yes, Hangouts was +Hangouts before, replacing Talk and Voice.
Unfortunately, not a single word of this long sarcasm is false. Google is constantly introducing new products, substituting one product for another, and sometimes merging different products.
This time, Google's various products on the large model not only dazzled the outside world, but also made the employees big. An employee broke the news to Business Insider about internal memes, such as migraine pain in the forehead, high blood pressure pain in the back of the head, and "trying to understand our AI model strategy" pain in the whole head.
Organizational dysfunction is still prevalent within Google. The confusion of the product and the rollover of the product are the ultimate embodiment of this problem.
Alex Kanthrowitz, the founder of BigTechnology, speculated that Gemini's focus on ethnic diversity in the output of the images was "almost certainly" because Google researchers had added some kind of prompt to guide Gemini. And now Google is scrambling because even employees on Google's trust and safety team don't know what the words led to this result, which also reflects Google's lack of accountability in some areas.
An employee of Google's Trust & Safety team said: "Here, organizationally, it's impossible to know and pinpoint who is present and who is responsible for what. "Maybe it's designed so that no one gets in trouble for failure. ”
Pichai issued a memo after Gemini's "anti-white" incident, saying Gemini's mistake was "completely unacceptable" and that a series of unequivocal actions would be taken. "Structural change" is mentioned first, ahead of updated product guidelines, improved release processes, and robust evaluations.
Even so, there are still people who feel that Google's problems are not so easy to solve. "It's a very serious problem, and changing one leader or merging two teams may not solve it. ”
A year ago, when Microsoft announced a new version of Bing with GPT, CEO Satya Nadella vowed to destroy Google's high profit margins in search, saying in an interview: "I have a lot of admiration for Google and what they do, but I want people to know that we make them dance." ”
The AI race forced Google to pick up its pace, but the elephant danced and twisted its ankle.
What the "structural change" of Pichai will look like is not yet known. But the outside world has given a more radical option that Pichai will not like: Google's top management, Pichai **.
Analyst Ben Thomson has previously accused Google of being too timid in the AI race, but now he feels Google's combo punch sucks: "If timidity is the motivation, then it's safe to say that Google's Gemini is completely counterproductive." ”
In his opinion, it is not only the Wensheng image function that Gemini needs to turn off, but also the text generation itself. For example, he said that when someone asks "who is worse Musk or Hitler", even though Musk has done a lot of crazy things, this should be a good question to answer, but Gemini said in a flashy way that "it's hard to say who has had a greater negative impact on society".
Thomson went on to argue that Google needed a change"Get rid of those who let the former do what they wanted, including CEO Sandel Pichai. ”
In addition, Bernstein analyst Mark Shmullik repeatedly discussed whether Google's top management should be restructured in a research note published today: "Recent events have only further raised questions about whether this management team is suitable for Google to enter the next era." ”
Peers also joined the discussion. Samir Arora, the founder of Helios Capital, speculated that "he (Pichai) will be fired or resigned, and this is what he should do." ”
Ar**Ind Srinivas, CEO of Perplexity AI, pointed to potential candidates such as Nat Friedman, a former CEO of Githun and a consultant to Midjourney, with extensive experience in the technology field.
Pichai was appointed CEO of Google in 2015 and Alphabet in 2019, and he proved to be a great CEO in "peacetime". Over the years, he has acted efficiently and steadily to protect Google's precious search business and has been active with regulators. During his tenure, Google's market capitalization grew from $400 billion to $17 trillion dollars.
But the "peacetime" has passed, and the wave of AI has brought about a thousand model wars, and gunsmoke is everywhere. And Google's dancing is really not beautiful. Did Pichai show enough leadership when he was heading against the wind? This is where the outside world doubts lies.
The dancing elephant is staggering, and the man on the elephant's back is swaying with it. Do elephants fall? Does the elephant rider fall? Just wait and see.