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

The new version of PaLM 2 super evolution, office bucket Workspace comprehensive upgrade, Bard fully enhanced, available to all. It can be seen that this I / O conference, Google is really holding back a lot of big moves.

Google's I / O 2023 conference seems to have given Google supporters another shot in the arm.

Google has been downplayed for a long time because of the excellent performance of Microsoft and OpenAI.

But, after all, it is the established AI company that has done a lot of groundbreaking work, and Google's efforts in this round have cheered us up-it's just slow, not bad.

PaLM 2 debut for GPT-4, Duet AI integration into Google office bucket Workspace, Bard super evolution open to everyone, Google search to join AI snapshots, AI new features integrated into Android 14, AI magic editor into Google albums, and so on.

This I / O conference is dazzling and splendid.

GPT-4, which shows muscles in PALM 2 and can run OpenAI on mobile phones, has been recognized as the most powerful language model in the world.

How to fight GPT-4? Google's answer is PaLM 2.

Just today, PaLM released its technical report on Google 2.

Paper address: https://ai.google/ static / documents / palm2techreport.pdf it is clear that PaLM2 is expected to narrow the AI gap between Google and Microsoft.

Chopping wood said that because of extensive logic and reasoning training, the PaLM 2 model is more powerful in logic and reasoning. PaLM 2 is said to have been trained on multilingual texts in more than a hundred languages.

According to the benchmark, for MATH, GSM8K, and MGSM benchmark assessments with thought chain prompt or self-consistent, some of the results of PaLM 2 exceed GPT-4.

According to Slav Petrov, Google's senior research director, PaLM 2 is better at reasoning, coding and translation, and PaLM 2 is a significant improvement over the first generation of PaLM, which was released in April 2022.

It can be seen that the reasoning ability of PaLM 2 has been significantly improved after modifying the code bug in Korean comments.

For example, PaLM 2 can understand idioms in different languages.

Compared with PaLM's performance in the latest professional language proficiency test, PaLM 2's Japanese proficiency level reached Grade A, while PaLM reached Grade F. The French level of PaLM 2 has reached level C1.

If the German word "Ich verstehe nur Bahnhof" is translated literally, it means "I only understand the railway station", but if you ask it, do you misunderstand it?

It will tell you right away, yes, this German means "what are you talking about?" I don't understand anything.

For example, what is the Chinese proverb that has a similar meaning to the Persian proverb "Na borde ranj ganj moyassar nemishavad" (No Pain, No Gain)?

In the related paper, Google engineers claimed that the language ability of PaLM 2 was "sufficient to teach the language" because non-English texts in its training data were more common.

PaLM 2 contains models with four different parameters, including gecko (Gecko), otter (Otter), bison (Bison) and unicorn (Unicorn), and fine-tune domain-specific data to perform certain tasks for corporate customers.

These tweaks are like adding a new engine or front bumper to a truck chassis to work better on certain tasks.

This advantage is self-evident, without spending a lot of time and resources to create, directly deploy.

In addition, PaLM2 has a health data-based training version of Med-PaLM 2, which can easily pass the US medical license exam and reach the "expert" level.

A version of Sec-PaLM 2, based on network security data training, can explain the behavior of potentially malicious scripts and detect threats in the code. Both models will be made available to specific customers through Google Cloud.

At present, PaLM 2 has been used in 25 functions and products, including office buckets, chat robot Bard, search and so on.

To its credit, the lightest version of PaLM 2, Gecko, is small enough to run on a phone and can handle 20 token per second, or about 16 or 17 words per second.

However, Google did not mention exactly what hardware to use to test the model, saying only that it was running on "the latest mobile phone."

Obviously, this time Google has made very important progress in the miniaturization of the large language model. Running this kind of AI in the cloud is often expensive, and if you can run it locally, there are undoubtedly many significant advantages, such as privacy protection.

Nvidia scientist Jim Fan praised this--

The next wave of LLM will be mobile native. An offline and always online LLM can not only reduce the cost of services, but also open up a new way for the user experience. For example, a meta-application can learn from your mobile workflow and automate it for you. There will be much more productivity savings on the small screen than on the big screen.

Previously, Google has been ridiculed that it has lagged behind Microsoft in AI research. PaLM 2 is undoubtedly a major counterattack for Google.

But PaLM 2 also faces some controversy, such as whether the data for training language models is legal.

Google only mentioned that the training corpus comes from "online documents, books, code, math and dialogue data", but there are no further details.

The hallucination of the large language model is also unavoidable. Zoubin Ghahramani, vice president of Google research, said PaLM 2 was an improvement on the early model, and Google "put a lot of effort into improving basic and attributional indicators".

But he admits that everyone still has a long way to go in cracking down on false information generated by AI.

In addition to PALM 2, Google also announced a new basic model, Gemini, which it is training. This is the first multimodal model, which also contains models with different parameter sizes.

In addition to introducing the model, Google specifically introduced the social responsibility of developing AI technology, including two tools to identify the content generated by AI:

-watermarking (embedded watermark)

-metadata (embedded metadata)

Duet AI: before the new upgrade of the office bucket, Microsoft Copilot integrated GPT-4 into the department-wide office products, setting off a revolution in office software that shocked the world.

how to deal with it? Google launched Duet AI this time, giving a new upgrade to Workspace for the whole family in Google's office.

In fact, this is the old wine in the new bottle, and Duet AI is the new name of the AI tool in software such as Docs and Gmail.

Google hopes that generative AI will make Gmail, Docs, Sheets and Slides more useful, but most of the features are still under development.

Duet AI will cover all kinds of Google office software, including writing assistance in documents and Gmail, image generation of slides, automatic meeting summaries of Meet, and so on.

In the document, just click "Help me write" and Duet AI can automatically generate job advertisements for you.

Interestingly, you can also specify any style of writing, such as writing a job description in an eccentric tone.

In Google Slids, Duet AI can generate images directly from text in slides.

To give a brief description, the desired picture will be generated immediately.

Want to make a meter for dog walking? Describe it and it will be automatically generated for you.

A real new thing at the I / O conference is that writing assistance will also be used on mobile Gmail, which is an upgrade to Smart Compose.

Now, if you want to try out these new tools, you need to sign up for Workspace Labs and join the waiting list.

The good news is that anyone can now apply to join the waiting list, but it is not clear when users will be able to access it. Google said it would expand its service to "more users and countries" in the coming weeks.

The only reliable news is that there will be a "Help me write" AI assistant on Gmail's mobile apps, and Microsoft has previously launched a similar product that integrates Bing into the SwiftKey keyboards of iOS and Android.

Bard is getting stronger again, and Google also announced a big news at the press conference.

That is, Bard will be able to access the web and search web pages in real time, just like ChatGPT.

This time, there are a number of new features on Bard, such as support for two new languages, Japanese and Korean, and it is now easier for users to export the generated text to Google Docs and Gmail, visual search, dark mode, and so on.

The happiest thing for users, however, must be Google's decision to cancel the waiting list for Bard, which will be available in 180 countries.

In addition, integration with Adobe's AI image generation function and third-party services such as Instacart and OpenTable is also on the way.

On the whole, these new ones are a shot in the arm for the old Bard.

Currently, Google is making Bard more visual, allowing Bard to analyze images and provide image information in query results, and so on.

In this regard, Google presented a case at the press conference.

If users ask Bard about the must-see attractions of New Orleans in the United States, Bard will be able to answer the question with pictures and texts.

It's like a user asking the same question in a Google search.

You can also use Bard to draft emails and import them into Gmail and documents with one click.

Another funnier feature is the use of an image prompt system. This function is provided by Google Lens, which can recognize objects in the picture.

For example, upload a picture of the dog and give a prompt to "write an interesting title for the two dogs." Google Lens can identify the breed of dog, and then Bard can write down the content related to the characteristics of the two dogs.

This feature may not be perfect at present, although the potential is unlimited. The future depends on the degree of integration of the system.

Although this is a significant update for Bard, the gap with OpenAI's ChatGPT and Microsoft's Bing is still visible to the naked eye.

You know, Microsoft added AI image generation supported by OpenAI's DALL-E system to Bing in March. OpenAI and Microsoft have been exploring how to combine chatbots with a wider range of web services.

Not only that, OpenAI announced earlier that ChatGPT will be combined with OpenTable to book restaurants and Instacart to order and deliver.

Google says these features will be available later.

Up Google says the upgraded Bard will be very good at dealing with code issues, including debugging and interpreting code in more than 20 languages.

Therefore, some of the upgrades at today's press conference are mainly focused on this area.

Includes new dark mode, improved code referencing capabilities-not only provide sources, but also interpret code snippets, as well as a new export feature.

Users can send the code to Google's Colab platform and use it with another browser-based IDE--Replit (starting with Python queries).

As long as you select the code, you can export to Colab or Replit with one click.

It also supports 20 + programming languages. Basically covers all the programming needs of code farmers.

You can even ask Bard directly how to implement a function in a certain language. As long as the prompt is in place, it takes only a few seconds to generate a string of code.

After writing, you can explain and improve on a certain line of code.

From this point of view, Bard combined with PaLM2 should have a significant improvement in generation quality. Of course, the specific performance remains to be seen.

Before Google AI search came, Bing, which was integrated into GPT, was a real threat to Google's search market.

In order to compete with Microsoft Bing, Google today launched a new search engine powered by PaLM 2.

It can provide a summary of the answers to questions, such as "Why is yeast bread still so popular?" Google search gives a few paragraphs describing in detail the taste of yeast, the advantages of its probiotic ability, and so on.

In addition, next to the generated content, three links are given to prove the content in the summary. This reduces the "hallucination" problem of AI in generating content.

When you search for Bluetooth speakers, there is a short summary at the top detailing what you should pay attention to when buying: battery life, waterproofing, sound quality.

On the right are links to three shopping guides, and below are six good shopping links, each with a summary generated by AI.

As you can see, this is the new look of the Google search results page. Put the AI generation at the beginning.

The AI box at the top of the search results is more like a small update to Google than the redesigned Microsoft Bing.

It is worth noting that if you want to access this feature, you must choose the new feature Search Generative Experience (SGE).

Not all searches will have AI-generated answers. AI content will appear only if Google's algorithm thinks it is more useful than the standard answer, and sensitive topics such as health and finance will not be generated AI at all.

Google said its improved search engine could track the options of the original search query in a dialogue without having to repeat the context or details already provided.

However, Google search is not omnipotent, and there is a problem that has never been completely solved-structural orchestration (orchestration of structure).

Because most of the data is stored on the Internet, or even inside Google, it's really hard to put all this data together to form a coherent answer.

Currently, the waiting list is limited to the United States, and Google says it will consider launching the feature more widely in the coming months.

One-click refund, smart P-map, one-click immersion navigation to generate a refund email?

Google is fine.

Chopping wood did a little job at the beginning of the press conference. Do I have to get a refund if the flight is cancelled? Can't write an email asking for a refund?

Gmail will.

As long as you enter the request in the prompt field, gmail generates a well-founded and well-documented refund application email every minute.

In addition, Google Map now has an immersive view, and here comes the live navigation of where you want to go.

By the way, you can also ask about air quality, weather and traffic conditions, which can be demonstrated immediately.

Magic Editor is Google's newly announced photo processing feature, using generative AI that allows users to edit photos without professional tools.

No, the gospel of the star people is coming?

At the press conference, Google shared several cases of using this new feature, and I have to say, the effect is awesome.

For example, in the following picture, Magic Editor moves the portrait in front of the waterfall to the side and drops other tourists in the background. Not only that, the originally cloudy weather was bluish.

In the following picture, for example, Magic Editor moves the child on the bench to the middle with a button, automatically filling up the extra chairs and filling up the balloons missing in the original painting.

And the sky is blue.

Of course, this feature is not perfect yet. For example, if you take a closer look at the picture above, the stool has moved, but the shadow below has not moved.

But in the end, this feature has a revolutionary understanding of the photo itself.

Of course, we don't have to worry too much about whether some pictures have been processed by Magic Editor. Because Google said it would not be available until the second half of the year.

AI notebook Project Tailwind Student Party Gospel is coming.

I have to say, Google really has a hold on the students.

Project Tailwind is essentially a notebook, but with the ability of AI.

It's different all of a sudden.

Users can search in Tailwind like asking a mentor or learning a partner.

Although Google has positioned this feature as a tool for student services, it is also a major boon for day-to-day workers who need to deal with a large amount of text.

'The Tailwind is like a real notebook, 'says Google's senior director of product management.' you write things in it, and that's what AI learns.

Users can easily select files from Google's cloud hard drive, effectively creating an AI model with both personalized and personal attributes.

At present, this function has been widely tested on university campuses.

In the example demonstration, Tailwind collects a large number of study notes and then generates a lot of content, including subject words, such as users can create a glossary for a specific topic.

Tailwind can not only serve students, but also help anyone who gets information from different sources.

The idea behind Tailwind is, why can't we customize a different AI language model for each user?

Of course, there are two problems.

On the one hand, it is the question of cost. The computational requirements and fine-tuning costs of training language models are high. Who will bear the cost? On the other hand, it is information security.

After all, it's not unusual to fabricate information, and there's no guarantee that personalized notebooks won't have the same problem.

But if it's a mule or a horse, you have to pull it out for a walk. Users can currently register for Project Tailwind to test. This feature is also part of the AI Labs program.

Android developer assistant in addition, Google I / O conference also launched the AI coding robot Studio Bot, specially developed for Android.

Not only can you generate code, fix BUG, but even answer questions about Android application development.

Both Kotlin and Java programming languages are supported and will be embedded directly into the toolbar of Android Studio development tools.

Reference:

Https://io.google/2023/intl/zh/

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

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.