Google I/O 2023: PaLM 2 Announced With Gecko, Otter, Bison, Unicorn & Med-PaLM 2, Gemini & DeepMind

[FurkanGozukara]

Google I/O 2023: PaLM 2 Announced With Gecko, Otter, Bison, Unicorn & Med-PaLM 2, Gemini & DeepMind

Full tutorial: https://www.youtube.com/watch?v=1UvUjTaJRz0

In today's #GoogleIO2023, we are given a sneak peek into the future of AI technology. Google unveiled their latest AI model, Palm 2, and showcased how it improves knowledge and learning, boosts creativity and productivity, enables developers and businesses to build transformative products and services, and importantly, how it ensures the responsible use of AI.

Palm 2 is an advanced model, offering excellent foundational capabilities and is designed to be highly efficient. They are launching over 25 products and features powered by Palm 2. The new model is stronger in logic and reasoning, trained on a broad range of scientific and mathematical topics and also understands and generates nuanced results in over 100 languages.

Special mention goes to the specialized versions of Palm 2, such as SecPalm for security use cases and MedPalm 2 for medical applications. MedPalm 2, for instance, has been demonstrated to perform at the expert level on medical licensing exam-style questions.

Google has also announced their next-generation foundation model, Gemini, which is currently in training. The company is investing heavily in AI responsibility, including the tools to identify synthetically generated content.

Finally, they showcased Bard, their experiment for conversational AI, now fully running on Palm 2.

Stay tuned for more updates and let's delve into the future of AI together.

Source Google. Full Event (10 May 2023) ⤵️

https://www.youtube.com/watch?v=cNfINi5CNbY

Our Discord server ⤵️

https://bit.ly/SECoursesDiscord

If I have been of assistance to you and you would like to show your support for my work, please consider becoming a patron on 🥰 ⤵️

https://www.patreon.com/SECourses

Technology & Science: News, Tips, Tutorials, Tricks, Best Applications, Guides, Reviews ⤵️

https://www.youtube.com/playlist?list=PL_pbwdIyffsnkay6X91BWb9rrfLATUMr3

Playlist of StableDiffusion Tutorials, Automatic1111 and Google Colab Guides, DreamBooth, Textual Inversion / Embedding, LoRA, AI Upscaling, Pix2Pix, Img2Img ⤵️

https://www.youtube.com/playlist?list=PL_pbwdIyffsmclLl0O144nQRnezKlNdx3

#Google #GoogleIO #ArtificialIntelligence #Palm2 #AI #DeepMind #MachineLearning #AIResponsibility #SecPalm #MedPalm #Bard #Gemini

Video Transcription

• 00:00:00 Looking ahead, making AI helpful for everyone is  the most profound way we will advance our mission. 

• 00:00:08 We are doing this in four important ways. 

• 00:00:11 First, by improving your knowledge and learning  and deepening your understanding of the world. 

• 00:00:17 Second, by boosting creativity and productivity  so you can express yourself and get things done. 

• 00:00:24 Third, by enabling developers  and businesses to build their  

• 00:00:27 own transformative products and services. And finally, by building and deploying AI  

• 00:00:33 responsibly so that everyone can benefit equally. We are so excited by the opportunities ahead. 

• 00:00:40 Our ability to make AI helpful for everyone relies  on continuously advancing our foundation models. 

• 00:00:49 So I want to take a moment to  share how we are approaching them. 

• 00:00:54 Last year, you heard us talk about Palm, which  led to many improvements across our products. 

• 00:01:00 Today, we are ready to announce our  latest Palm model in production,  

• 00:01:05 Palm 2. Palm  

• 00:01:13 2 builds on our fundamental research  and our latest infrastructure. 

• 00:01:18 It's highly capable at a wide  range of tasks and easy to deploy. 

• 00:01:24 We are announcing over 25 products  and features powered by Palm 2 today. 

• 00:01:29 Palm 2 models deliver excellent foundational  capabilities across a wide range of sizes. 

• 00:01:35 We have affectionately named them  Gecko, Otter, Bison and Unicorn. 

• 00:01:42 Gecko is so lightweight that  it can work on mobile devices. 

• 00:01:47 Fast enough for great interactive  applications on device, even when offline. 

• 00:01:53 Palm 2 models are stronger in logic and reasoning,  

• 00:01:56 thanks to broad training on  scientific and mathematical topics. 

• 00:02:01 It's also trained on multilingual  text spanning over 100 languages,  

• 00:02:06 so it understands and generates nuanced results. Combined with powerful coding capabilities,  

• 00:02:13 Palm 2 can also help developers  collaborating around the world. 

• 00:02:17 Let's look at this example. Let's say you're working with a  

• 00:02:21 colleague in Seoul and you're debugging code. You can ask it to fix a bug and help out your  

• 00:02:28 teammate by adding comments in Korean to the code. It first recognizes the code is recursive,  

• 00:02:35 suggests effects, and even explains  the reasoning behind effects. 

• 00:02:40 And as you can see, it added comments  in Korean, just like you asked. 

• 00:02:51 While Palm 2 is highly capable, it really shines  when fine-tuned on domain-specific knowledge. 

• 00:02:59 We recently released SecPalm, a version of  Palm 2 fine-tuned for security use cases. 

• 00:03:05 It uses AI to better detect malicious  scripts and can help security experts  

• 00:03:11 understand and resolve threats. Another example is MedPalm 2. 

• 00:03:16 In this case, it's fine-tuned  on medical knowledge. 

• 00:03:19 This fine-tuning achieved a 9x reduction in  inaccurate reasoning when compared to the model. 

• 00:03:26 Approaching the performance of clinician  experts who answered the same set of questions. 

• 00:03:32 In fact, MedPalm 2 was the first language  model to perform at expert level on medical  

• 00:03:38 licensing exam-style questions and  is currently the state of the art. 

• 00:03:43 We are also working to add capabilities to MedPalm  2 so that it can synthesize information from  

• 00:03:49 medical imaging like plain films and mammograms. You can imagine an AI collaborator that helps  

• 00:03:56 radiologists interpret images  and communicate the results. 

• 00:04:01 These are some examples of Palm 2  being used in specialized domains. 

• 00:04:05 We can't wait to see it used in more. That's why I'm pleased to announce that  

• 00:04:10 it is now available in preview  and I'll let Thomas share more. 

• 00:04:22 Palm 2 is the latest step in our  decade-long journey to bring AI  

• 00:04:26 in responsible ways to billions of people. It builds on progress made by two world-class  

• 00:04:32 teams, the Brain team and DeepMind. Looking back at the defining AI  

• 00:04:37 breakthroughs over the last decade, these teams  have contributed to a significant number of them. 

• 00:04:42 AlphaGo, transformers,  sequence-to-sequence models, and so on. 

• 00:04:48 All this helps set the stage for the  inflection point we are at today. 

• 00:04:52 We recently brought these two teams  together into a single unit Google DeepMind. 

• 00:04:58 Using the computational resources of Google,  

• 00:05:01 they have focused on building more  capable systems safely and responsibly. 

• 00:05:07 This includes our next generation foundation  model, Gemini, which is still in training. 

• 00:05:13 Gemini was created from the  ground up to be multimodal,  

• 00:05:18 highly efficient at tool and API integrations, and built to enable future innovations like  

• 00:05:25 memory and planning. While still early,  

• 00:05:28 we are already seeing impressive multimodal  capabilities not seen in prior models. 

• 00:05:34 Once fine-tuned and rigorously tested for  safety, Gemini will be available at various  

• 00:05:40 sizes and capabilities, just like Palm2. As we invest in more advanced models,  

• 00:05:47 we are also deeply investing in AI responsibility. This includes having the tools to identify  

• 00:05:54 synthetically generated content  whenever you encounter it. 

• 00:05:59 Two important approaches are  watermarking and metadata. 

• 00:06:04 Watermarking embeds information directly  into content in ways that are maintained  

• 00:06:10 even through modest image editing. Moving forward, we are building our  

• 00:06:15 models to include watermarking and  other techniques from the start. 

• 00:06:21 If you look at the synthetic image, it's  impressive how real it looks, so you can imagine  

• 00:06:26 how important this is going to be in the future. Metadata allows content creators to associate  

• 00:06:33 additional context with original files, giving you  more information whenever you encounter an image. 

• 00:06:41 We'll ensure every one of our  AI-generated images has that metadata. 

• 00:06:47 James will talk about our  responsible approach to AI later. 

• 00:06:53 As models get better and more capable,  one of the most exciting opportunities  

• 00:06:59 is making them available for  people to engage with directly. 

• 00:07:04 That's the opportunity we have at Bard,  our experiment for conversational AI. 

• 00:07:10 We are a rapidly evolving board. It now supports a wide range of  

• 00:07:14 programming capabilities, and it's gotten  much smarter at reasoning and math problems. 

• 00:07:21 And as of today, it is now  fully running on Palm 2. 

• 00:07:26 To share more about what's coming,  let me turn it over to Sissy.