Shulou(Shulou.com)11/24 Report--

The birth of ChatGPT made Pichai really afraid. Is it Google's strategic mistake not to vigorously develop chatbots?

This month, OpenAI's ChatGPT was born, immediately setting off a big storm in the AI circle.

Clear, intelligent, persuasive, and even able to arrange papers and code. At that time, many people predicted that Google might take the pill.

This is not empty talk, now, Google is really scared.

It is reported that with the ChatGPT fire, Google management has recently released a "red code".

Previously, Google management had vowed that ChatGPT could not replace us. Now this is the speed of light hit in the face?

What does "code red" mean by chopping wood and shivering? In Silicon Valley, this means sounding the "fire alarm".

The largest company in Silicon Valley is really scared now-maybe ChatGPT is really causing a huge technological change that upends the entire industry.

The change was so rapid that Google didn't even respond.

ChatGPT can answer people's questions in clear, simple sentences, rather than just providing a list of links. It interprets concepts in a way that is easy for people to understand.

It can even write an article from beginning to end, including business planning, Christmas gift advice, blog posts and holiday plans.

More and more people are asking: will search engines be replaced? Now, Google is not sure.

So far, Google CEO Pichai has participated in several meetings around Google's AI strategy and highlighted the need for Google teams to focus on addressing the threat posed by ChatGPT to the company's search engine business.

According to the New York Times, Pichai has told some teams to change direction and work on developing AI products. According to the Times, teams from Google's research, trust and security departments and other departments have been instructed to shift gears to assist in the development and release of artificial intelligence prototypes and products.

Some of these employees are tasked with building AI products that generate art and graphics, similar to OpenAI's DALL-E, which is already used by millions of people and is a fairly successful product.

Why is Google so afraid of ChatGPT? Because, if it does replace the search engine, Google's advertising revenue business model will be greatly affected, or even fundamentally subverted.

This is from the root of the rice bowl.

Sridhar Ramaswamy, a former head of Google's advertising team, said that if ChatGPT continues to be so popular, Internet users will no longer click on Google links with ads. Advertising, which earned Google $208 billion in 2021, accounts for 81% of Alphabet's total revenue.

Five days after its public release, ChatGPT has accumulated more than 1 million users, and in response to user queries, it can collect information from millions of websites and generate unique answers in a conversational and humane way. It can write papers, write code, and even act as people's therapists.

Of course, it is often full of mistakes because it cannot verify what it says and even fabricates answers.

In fact, Google has had the opportunity to go this way. Google's Lamda chat robot was famous long before ChatGPT came out.

But Zoubin Ghahramani, head of Google Brain at Google's AI Lab, told the Times that chatbots are "not something that people can reliably use every day."

Instead, experts say, over time, Google should focus on its search engine rather than dismantling it.

Ten days later, the attitude changed sharply. Google's attitude towards ChatGPT was not the same before.

More than a decade ago, Google executives said that Google would not launch a competitor to ChatGPT because it had a greater "reputational risk" than startups such as OpenAI.

And they believe that chatbots are not to the extent that they can replace search engines.

Google's AI chat robot LaMDA has made a stunning debut, but Google is reluctant to release it to the public because it has a high error rate and is easy to carry poison.

More than a decade ago, Jeff Dean, head of Google's AI, told a plenary meeting that while Google had the technology and ability to make AI products, they had to make decisions "more conservatively" than "small startups".

At that time, at the meeting, an employee of Google asked: did Google miss the opportunity without launching a chat robot like OpenAI? Especially considering that Google already has its own conversation technology, such as LaMDA, as well as language models for other dialogue applications.

Dean's answer is that while Google is fully capable of being chatbots, chatbots are vulnerable to biases and false information, and Google, a large company with more than a billion users, is even less likely to get rid of this influence.

"We certainly want to use these technologies in real products, especially in things that highlight language models, rather than hiding behind the scenes, like this," Dean said of Google's AI. "but, more importantly, we have to do the right thing. "

In addition to LaMDA, the BERT and MUM AI language models that Google uses to improve its search engine can also be pulled out to compete with ChatGPT.

However, Morgan Stanley also released a report saying that ChatGPT does not pose a major threat to Google's position, as Google will continue to improve its search engine and language model.

Google: my AI is stronger, but advertising is more important. Google's entanglement is easy to understand, because Google is not at a disadvantage in the field of chatbots.

Google unveiled its latest artificial intelligence system, LaMDA, at the I / O conference in May 2021.

Google said that LaMDA can make the answer to the question more "make sense", make the conversation more natural, and these responses are not predetermined (unlike those chatbots), and even the same answers will not be used a second time.

At the I / O conference in May 2022, Google unveiled its upgraded second-generation model, LaMDA 2.

During the demonstration, AI asked at the beginning, "have you ever wondered why dogs like to play fetching games so much?" In response to simple follow-up questions such as "Why is this?" AI gives more information about dogs and their sense of smell.

What makes LaMDA completely out of the circle is that in June this year, Google engineer Blake Lemoine actually had a relationship with LaMDA, believing that it not only had the intelligence of an eight-year-old, but also was "conscious".

The next month, Google fired the "obsessed" Lemoine.

Although the whole thing seems ridiculous, it does show that chatbots have come a long way in technology.

Since Google itself is working on chat AI, why not deploy it for so long?

First of all, a considerable amount of the data on which the AI model is trained comes from the Internet, which is full of prejudice, hatred and abuse.

Given that Google itself is very popular, if the model outputs negative results, public opinion can easily ferment to the point where it will not end. And the image Google has taken decades to build will be destroyed.

This is not an unfounded worry, as can be seen from netizens' collective wisdom and ways to circumvent the security restrictions of ChatGPT.

Second, even if Google perfects the chat robot, there is still a question to be solved-will the technology erode the company's lucrative search advertising?

After all, for Google, digital advertising accounts for more than 80% of all revenue.

Obviously, if chatbots can respond to queries with tight sentences, there will be no reason for people to click on advertising links.

So, before ChatGPT came out, Google's strategy was to use chatbot technology to enhance the search engine experience.

ChatGPT actual competition Google search so, after all, can ChatGPT replace Google?

Let's start with a wave of actual combat exercises to see who is the winner.

CNBC columnist Sofia Pitt wrote an article entitled "about the day I switched from Google to ChatGPT" and tested and compared the sense of use between the two:

I recently bought a second violin leaf fern plant, but the new plant died in a few days, and I usually ask Google what to do.

This time I tried to find the answer on ChatGPT, and I typed "how can I keep my pteridophyte alive?" The result is consistent with the proposal I received from the botanical company Easy Plant.

Compared to Google, I don't have to travel through thousands of entries and websites, and I don't have to put up with all kinds of pop-up ads.

When I used Google to search for the same question, the top results gave me a detailed explanation in an article, including ads and links to buy new soil.

In the first round, ChatGPT wins.

Next, I would like to prepare a gift for my husband's Hanukkah. I asked ChatGPT for some advice.

I didn't give it any information about my husband, except that these gifts are for Hanukkah.

Then ChatGPT assumes that my husband is a kosher and that he likes technology and watches. There is nothing wrong with this, but this is not the answer I hope.

When I Google this question, I can access hundreds of articles, including different gift ideas and website links.

I received advertisements from Lululemon and Bombas, which I used to go to when I bought gifts.

In this case, Google used its personalized knowledge of me to find more targeted results on the Internet.

Google wins the second round.

Can we kill Google or not? After all that has been said, the question is: can ChatGPT, or the upcoming GPT4, replace Google?

The answer of Zhihu respondent Zhang Junlin is, not yet.

Based on the powerful GPT 3.5 large-scale language model (LLM,Large Language Model), ChatGPT introduces "manual tagging data + reinforcement learning" to continuously fine-tune the pre-training language model, the main purpose is to make the LLM model learn to understand the meaning of human command instructions.

First of all, for many types of knowledge questions, ChatGPT will give what seems to be reasonable, but in fact the wrong answer (you can see all kinds of bizarre answers).

Secondly, ChatGPT's current mode of tagging data training based on the GPT large model is very unfriendly to the LLM model to absorb new knowledge.

New knowledge is always emerging, and it is unrealistic to retrain the GPT model with the emergence of some new knowledge, whether it is the training time cost or the money cost.

If we adopt the Fine-tune model for new knowledge, it seems feasible and the cost is relatively low, but it is easy to lead to the disaster forgetting of the original knowledge due to the introduction of new data.

Third, another limiting factor is cost.

At present, the cost of AI training and producing results is still not cheap. If there are hundreds of millions of user requests for real search engines, assuming that the free strategy continues, OpenAI cannot afford it.

However, if the charging strategy is adopted, it will greatly reduce the user base, so whether to charge or not is a dilemma.

In addition, the emergence of ChatGPT does have an essential impact on Google's own search engine business.

Google's biggest source of revenue is advertising, where users type questions into the search box to get relevant content and advertising links.

Google's business is to solve users in the vast sea of information, more quickly to find the correct information that can solve users' problems. But the problem is that if it is too fast and easy to solve, there is no way to sell more ads, which is a somewhat self-contradictory weakness of search engine advertising itself.

Of course, LaMDA also allows users to type questions and plug ads next to or below the content, but the experience is very different from that of search engines, so this conversational AI may have an impact on Google's business nature.

But the advent of ChatGPT, which is also part of the competition, is likely to accelerate the pace at which Google is forced to launch LaMDA, at least not allowing ChatGPT to monopolize a large number of users to erode the demand for search engines.

And ChatGPT is currently not connected to the Internet, so it is impossible to optimize his answer with the latest data, but this will not be a limiting factor in the future.

Marcus, who has always watched the hustle and bustle, is sure to be cynical when it comes to whether ChatGPT can replace Google when it comes to ChatGPT himself and the big shot in the self-aware circle.

Black Lemonie, an engineer who was previously fired by Google, believes that ChatGPT is still a few years behind LaMDA.

OpenAI's CEO Sam Altman himself mentioned that ChatGPT still has a lot of limitations and a lot of enhancements in terms of robustness and reliability, and it's too early to rely entirely on it.

The following are the main limitations of ChatGPT at this stage, which may vary depending on the specific language model and the tasks it is intended to perform:

Language models are trained on specific data sets and may not be able to answer questions outside the training area (such as language models trained in news articles that may not answer questions about medical topics or technical details)

The language model does not always understand the context or expected meaning of the question, so it may provide inaccurate or less relevant answers

Language models may not be able to handle complex multi-step reasoning tasks, such as solving mathematical problems

The language model cannot browse the web or access external information sources, and can only provide answers based on its training data.

Language models may not provide accurate answers to sensitive and negative topics every time.

To sum up, the value of ChatGPT is that it can explain complex problems and accomplish simple writing tasks like talking to a real person.

For example, ask ChatGPT to explain concepts such as the stock market and the Internet, and we get the same answer as talking to experts.

Compared to Google, we don't have to spend time filtering unwanted results.

We can also ask it to write a letter to the landlord asking for early termination of the lease, and we can almost send him the ChatGPT results directly, saving time and effort.

On the other hand, Google knows us better.

It will customize our needs and provide diversified suggestions.

For example, if you search for Apple stock ticker or Cheap flights to Aruba, it will directly display the market chart containing the latest price information, or the itinerary with the most likely cheapest flight date, and connect to multiple websites so that we can buy tickets directly.

ChatGPT, on the other hand, does not scan the Internet for real-time information, and it has only received data training before 2021.

Google is also very reliable because of the company's huge operating budget and years of expertise. By contrast, ChatGPT is still being tested and sometimes breaks down.

Can ChatGPT replace Google? It can be guessed that most people will continue to use Google to search for most of the questions, but if we are not satisfied with the results, we have a thoughtful alternative.

According to reports, at a recent plenary meeting, Pichai said that Google has many plans for AI in 2023.

At Google's annual developer conference I / O in May next year, we should be able to see what Google has done after giving up the chatbot route.

Reference:

Https://zhuanlan.zhihu.com/p/589533490

Https://www.businessinsider.com/google-management-issues-code-red-over-ChatGPT-report-2022-12

Https://www.nytimes.com/2022/12/21/technology/ai-ChatGPT-google-search.html

Https://www.cnbc.com/2022/12/15/google-vs-chatgpt-what-happened-when-i-swapped-services-for-a-day.html

Https://vocus.cc/article/6391887afd89780001827331

This article comes from the official account of Wechat: Xin Zhiyuan (ID:AI_era), Editor: editorial Department

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.