Original link: https://www.latepost.com/news/dj_detail?id=1645
User 1: Are you using the PaLM 2 model?
Google Bard: I switched to PaLM 2 on March 31, 2023. This was announced by Google CEO Sundar Pichai on the New York Times Hard Fork podcast.
User 2: Are you using the PaLM 2 model?
Google Bard: I’m not using PaLM 2, I’m using a smaller model that’s in development. PaLM 2 is a more powerful, but still research model, I’m not sure when it will be released, and open to the public…
After the Google I/O developer conference held in the early hours of this morning, throw the same question to Google’s chatbot Bard, and you will see a completely different answer. According to information released by Google, Bard has switched to the new generation model PaLM 2, but did not complete the switch on March 31 (but the podcast it mentioned does exist). So the first answer is partly correct.
Google didn’t solve the fundamental problem of inaccurate answers from large models—probably no one ever did. But the new PaLM 2 large language model, and a series of radical product updates announced at the same time, show how the Internet giants can promote the practical use of artificial intelligence within the current technical constraints.
Smaller models can be used better, and even run offline on mobile phones
Google released a new generation of large language model PaLM 2 at the meeting, not one, but four. Name the animals on a parametric scale from small to large: Gecko, Otter, Bison, and Unicorn.
Pichai said Gecko could run on a phone and be fast enough to work offline.
Previously, there were developers running large open-source models on mobile phones, but the speed was very slow, and it took ten minutes or more to wait for a result. It can’t compare with Google, which controls the Andriod system and has the ability to develop large models.
While larger AI models tend to work better, they also consume more computing power. Access to GPT-4’s New Bing and ChatGPT Plus once made Microsoft’s computing resources stretched. OpenAI temporarily suspended users’ paid subscriptions to ChatGPT Plus and limited the number of times paid users can use GPT-4 per day.
For individual users, the most reliable computing power is the mobile phone in hand or the computer at home. If you don’t play games, the computing power in these devices is redundant. Larger models that are smaller and consume fewer computing resources can make these redundant computing resources useful.
In the technical report, Google evaluated PaLM 2 in three sizes of S, M, and L – which did not correspond to the four models mentioned at the Google press conference – in some tasks, the smallest PaLM 2 will be 540 billion faster than the previous generation. parameter PaLM performed well.
Nvidia AI scientist Jim Fan calls Gecko the “biggest highlight” of the PaLM 2 model: “There’s a lot more productivity gain on the small screen than on the big screen.”
According to Google’s technical report, even the largest “unicorn” version, PaLM 2 has fewer parameters than the previous generation model, but is stronger in multiple tasks such as natural language generation, translation, and reasoning. “This shows that increasing the size of the model is not the only way to enhance the capacity of the model.”
This technological breakthrough is critical to the continued advancement of artificial intelligence. In the past few years, companies such as Google and OpenAI have confirmed the law of “the larger the model parameters, the stronger the capabilities”, and the large model competition of technology companies has pushed the model parameters to trillions. When training the models, they basically use up all the text data on the planet.
In April of this year, OpenAI CEO Sam Altman (Sam Altman) said at the Massachusetts Institute of Technology that “we are at the end of the era of giant models” and that the progress of the model will not come from making the model bigger.
Now Google is the first to submit the answer, but the problem-solving process: “slightly”.
In the same way that OpenAI introduced GPT-4, Google also released a technical report of more than 90 pages when introducing PaLM 2 – following the arXiv paper format commonly used in academia. The format is close to the paper, but the signed author of the article has become Google, and the list of researchers has been moved to the final appendix.
Similar to the release of GPT-4, Google also hides how PaLM 2 trains the model and how big the model parameters are. Artificial intelligence research is closely related to academia, but for large companies, it is ultimately a fierce commercial competition.
Compared with technical details, Google is more willing to talk about PaLM 2’s ability to learn across languages. According to the technical documentation, Google used data from 100 languages when training PaLM 2. Among the main language data, there are texts corresponding to English. Pichai said that PaLM 2 can understand subtle differences between languages and generate results that exceed expectations.
He gave an example of PaLM 2 helping people from different countries to write code together. Relying on PaLM 2, developers in South Korea can comment codes and make suggestions for revisions in Korean, and developers in the United States can understand them as well. PaLM 2 will also help developers in the US write comments in Korean.
Pichai did not directly mention competitor GPT-4 in the press conference. In the PaLM 2 technical report, GPT-4 only appeared a few times as a reference object. For example, when evaluating reasoning ability, Google says that PaLM 2 performs as well as GPT-4 on some data sets that test reasoning ability.
Some users on Reddit forums and social media said that the new Bard is “lightning fast”, but writing code is still not as good as GPT-4. Website designer Mike Hancock said that he gave GPT-4 and Bard the same code test questions. GPT-4 has not finished writing one answer, and Bard has given three complete answers, but the final result is that GPT-4 is better. .
Google also showed the results of Fine-tuning PaLM 2 with different data:
- Sec-PaLM, fine-tuned with secure data. It detects malicious scripts, helping security professionals understand and address threats.
- Med-PaLM 2, fine-tuned with medical data. It answers patient questions like a clinician. Accuracy is close to that of clinicians. It was the first language model to reach the “expert” level on the medical licensing exam, and it is the most advanced. Google said it will add the ability to view X-ray film in the future.
OpenAI has made it clear that it will not develop the next-generation model (GPT-5) in the short term, but will find other ways to make GPT-4 better. But Google isn’t about to stop there. As soon as PaLM 2 was opened to the public, Google has been developing a new generation of large-scale model Gemini from scratch.
Gemini will be multi-modal-able to process data such as language and pictures at the same time, and integrate various tools and APIs. “Although it is still in the early stages, we have seen impressive features that were not in the previous model.”
From Search to Maps to Gmail, big mockups are crammed into Google’s core products
“We’re at an exciting inflection point,” Pichai said. “Through generative AI, we’re reimagining all of our products. That includes search.”
Google is the Internet company with the most users – 15 products with more than 500 million users and 6 products with more than 2 billion users. Outside of China, most people use Google’s search, map, email, and video products.
During the keynote at Google I/O, there were new AI-powered versions of these products. Most notable is search advertising, which generates tens of billions of dollars in profits for Google each year. Any small adjustment may affect the foundation.
Microsoft’s New Bing directly made ChatGPT a “chat” interface, encouraging people to leave the search results-and leave the advertisements. Google has chosen a more balanced approach, embedding a new module called AI Snapshot on the search page to show the results generated by the large model.
When you search for “Bluetooth speakers suitable for pool parties”, the traditional search results appear first on the page – ten blue text links. After a few seconds, the content generated by artificial intelligence appears and is placed on the top, telling you the precautions for buying Bluetooth speakers (battery) longevity, water resistance, sound quality), and gives a buying guide (on the right), a link to the product (below) and a brief description of the product. You can also set a price (eg, under $100) and let it regenerate the results.
However, in mobile devices such as mobile phones, AI Snapshot will fill the entire screen. This will drastically reduce the frequency of clicks on traditional search results. Coupled with the fact that Google directly gives results instead of links, it will be more difficult for various websites to get traffic from Google search. Now, Google is experimenting with how to add ads to AI Snapshot.
Not all searches will trigger AI-generated results, and the former will only appear when Google’s algorithms deem AI-generated content superior to standard search results. When users search for sensitive topics such as health, finance, and safety hazards, AI Snapshot will not appear.
Google plans to test AI Snapshot in the United States first, with a limited number of places open in the next few weeks. On the application page, Google also plans to test other search functions, such as Code Tips, which directly generates code by entering programming questions in the search box.
During the keynote, Google executives showed how AI could transform other core products:
- A more immersive Google Maps. Google has synthesized more than 1 billion panoramic maps with artificial intelligence technology. When you select the navigation route and determine the departure time, the artificial intelligence will generate an immersive 3D route map with a bird’s-eye view, calculate the weather and traffic conditions in advance, and display them in the animation simultaneously. Google plans to launch this feature this summer, and then expand to 15 cities including London, New York, Tokyo, and San Francisco.
- Google Photos with automatic photo retouching. You only need to make a request, such as increasing the brightness, cutting out or completing the objects in the picture, moving the position of the character, changing the dark clouds in the background to blue sky, etc., and the new Magic Editor function can be automatically completed.
- Automatically compose emails for Gmail. You only need to enter the requirements in Gmail’s “Help me write” tool, such as an email requesting a full refund, click Create, and it will combine the information in the previous email to write a complete email. This feature will be rolled out as part of an update to Workspace. Other functions of Workspace include automatically writing speeches based on PPT content.
- More powerful chatbot Bard. The underlying model of Bard is replaced by PaLM 2; Korean and Japanese are added in addition to English, and Chinese is expected to be supported in July; Bard’s reply can be transferred to Gmail and Docs with one click; help you add titles and descriptions to photos taken with your mobile phone; and support Adobe Firefly, a picture generation tool without copyright disputes, is not a drawing application trained by itself with public data sets. Here’s Bard’s response when asked to create an invitation image for his daughter’s birthday party (required to include unicorns and a birthday cake):
Google has been in a unique position in the AI wave.
It is one of the earliest companies to study artificial intelligence and has the strongest technical strength. It has many patents on the underlying technologies of artificial intelligence, such as the current large-scale model infrastructure Transformer.
It is also the largest Internet company in the world, serving billions of people every day. It has Andriod that can directly affect the mobile phones used by more than 3 billion people around the world. It is more capable of deploying large models to more scenarios than other companies.
Therefore, Google is closely watched by users and regulators. The development and launch of each artificial intelligence product must consider legal and public opinion risks. Many Google employees believe that this is the reason why OpenAI was able to launch ChatGPT first.
Under the continuous impact of OpenAI and Microsoft, Google moved quickly to integrate Google AI and DeepMind, two top artificial intelligence teams that belong to the company but have not cooperated, postponed the disclosure of the latest research results, and actively developed artificial intelligence-based Search engine Magi. This new product, which is being tested internally at Google, also uses PaLM 2.
Not long ago, Pichai was asked “What did you miss if you didn’t release Bard before ChatGPT”. He gave a standard answer from a CEO of a large company. Google is not the first to make a search engine, nor is it the first to make a browser. “Sometimes it is important to be the first, but sometimes it doesn’t matter.” In his opinion, As long as the product is continuously improved to achieve better functions, the latecomers can also come first.
This developer conference demonstrated Google’s artificial intelligence technology accumulation and product iteration capabilities.
Title map source: Visual China
This article is transferred from: https://www.latepost.com/news/dj_detail?id=1645
This site is only for collection, and the copyright belongs to the original author.