A YouTube commentary (by LARRUZ, uploaded Aug 9, 2023) describing Google DeepMind's Gemini as an ambitious, multimodal successor aimed to surpass GPT-4 by merging DeepMind's game-playing advances (AlphaGo's tree search and neural-network-boosted planning) with large language modeling. The transcript explains AlphaGo's tree search and relates it to recent 'Tree of Thoughts' work that improved GPT-4's reasoning, surveys multimodal efforts (GPT-4 image demo, Microsoft JARVIS), and raises safety concerns about giving models long-term planning abilities (misalignment, unforeseen consequences, resource optimization, superintelligence risk).
Chapters
There have been some absolutely wild developments in the artificial intelligence landscape. The Google DeepMind CEO Demis Hassabis has stated that their new AI that they are working on, called Project Gemini is going to eclipse ChatGPT, or in other words, be much better than the current model that everyone loves, which is GPT4. The company's vision for Gemini is ambitious, merging the strengths of their renowned game playing AI with the expansive language understanding of models like ChatGPT. This unique blend seeks to establish a new standard in the AI world, combining deep knowledge with advanced strategic thinking. The incorporation of AlphaGo's techniques is especially noteworthy. AlphaGo's historic victory over a Go world champion in 2016 showcased the prowess of DeepMind's reinforcement learning methods. In the realm of strategic games like Go, where there are more potential board configurations than atoms in the universe, the software's ability to predict and plan was groundbreaking. By integrating these techniques with a large scale text model, Gemini could potentially bridge the gap between raw information processing and nuanced decision making. Take a look at this clip of AlphaGo where the extraordinary capabilities of AlphaGo are demonstrated. Go is the world's oldest continuously played board game. It is one of the simplest and also most abstract.
Beating a professional player at Go is a long standing challenge of artificial intelligence.
Everything we've ever tried in AI just falls over when you try the game of Go.
The number of possible configurations of the board is more than the number of atoms in the universe.
AlphaGo found a way to learn how to play Go.
So far, AlphaGo has beaten every challenge we've given it, but we won't know its true strength until we play somebody who is at the top of the world World like Lisa Don.
Now something innovative that AlphaGo actually did was use something called Tree Search, which is quite similar to the research earlier this year which is called Tree of Thoughts. And Tree of Thoughts was basically where they got GPT4 to list out all possible answers to a question, then rank them in terms of which is likely to be most correct. It improved the reasoning by 400%. However, if Gemini uses Tree Search, it will be remarkably interesting. You see, Tree Search was quite similar, but a little but different. Tree Search looks at many moves. Think of Tree Search like exploring a big tree where each branch is a possible move in the game. The more it looks, the bigger and more detailed the tree becomes. Balancing Act Tree Search has a cool strategy. Instead of just looking at every possible move, it balances between trying out new moves and focusing on moves that seem really good based on what it knows exploitation AI Boost with neural networks AlphaGo gives the tree search method a boost. It uses artificial brain like systems called neural networks to help choose moves and assess the game situation. This makes AlphaGo smarter than other Go programs. Starting Point and Exploration AlphaGo begins by looking at the current game situation and then explores different moves from there. It keeps exploring until it finds a move it hasn't fully checked out yet. Refining choices as AlphaGo keeps looking at the tree and exploring, it gets a better and better idea of the game and what moves might be best. Hasabis has also hinted at other innovative features in Gemini that have yet to be disclosed. The promise of interesting innovations combined with the noted multimodal capabilities implies a breadth of application beyond what current models offer. The emphasis on multimodality suggests that Gemini might be capable of understanding and generating content across various media types, be it text, image, sound or even potentially video. Moreover, the model's planned proficiency in integrating with tools and APIs means it could be a game changer in automation and system to system communication. Now, when we do discuss multimodal capabilities in AI, it's important to note that most AIs currently aren't multimodal. Although with the release of GPT4 they did tease image functionality in this clip here.
And let's actually try this one as well. What's funny about this image? Oh, it's already been submitted. Once again we can verify that it's making the right API calls. Squirrels do typically eat nuts. We don't expect them to use a camera or act like a human. So I think that's that's a pretty good explanation of why that image is funny.
It still is yet to have a worldwide release, although some users can access it in bing chat. Most AIs that are currently available exist as narrow AI or more easily understood as AI with one specific purpose. For example Chat GPT for generating text, using Elevenlabs for a voiceover or audio generation, and using Midjourney for images. But what if we were able to combine it all into one AI? Well, it's not like it hasn't been done before. That's what Microsoft's earlier project aimed to do. Microsoft JARVIS is an innovative multimodal AI powered platform that can connect and collaborate with multiple artificial intelligence models to deliver a final result. Named after Ironman's personal AI Assistant, Jarvis aims to bring together the power of the open Source Community and ChatGPT. The platform is hosted on Hugging Face and it is connected to as many as 20 different models, including T5 base, stable diffusion 1.5 BERT, Facebook's Bartlarge, CNN, Intel's DPT large and more the standout feature of Jarvis is the idea behind it, which can be condensed to the Definition Language as an Interface by using language as a general interface and putting the LLM large language model in the brain position, it is possible for many different specialized AI models to work together. This allows Jarvis to handle various tasks such as pose detection, image generation, image classification, image captioning and text to speech by calling on the appropriate models. Jarvis works similarly to how OpenAI demonstrated GPT4's multimodal capabilities with text's and images, but it takes it one step further by integrating various open source LLMs for images, videos, audio and more. It can also connect to the Internet and access files, allowing users to enter a URL from a website and ask questions about it. So with this in mind, it seems like Google's DeepMind could utilize this framework in order to create Gemini, which would certainly be much more powerful than GPT4. Before now, one thing that they did say which concerned me was this right here they said that they are aiming to give the system new capabilities such as planning or the ability to solve problems. Now that sounds good in theory, but one of those things is much more dangerous than the others, and that's planning. You see, giving an AI the ability to plan long term goals can be dangerous due to the risks of existential catastrophe and incomplete goal specifications. Misalignment of Objectives if the AI's long term goals are not perfectly aligned with human values and interests, it could pursue actions that are harmful or contrary to our well being. An AI with poorly defined or misaligned objectives might prioritize its own self preservation or optimization at the expense of humans. Unforeseen Consequences Long term planning involves considering complex scenarios and making predictions about the future. AI systems may not fully comprehend the consequences of their actions leading to unforeseen and potentially negative outcomes. The AI could pursue its goals without understanding the broader implications leading to unintended harm. Long term planning implies setting specific objectives that the AI will work towards over extended periods. However, the world is constantly changing and unforeseen events may alter the relevance or desirability of those goals. An inflexible AI could continue to pursue outdated or harmful objectives despite changing circumstances. Resource Optimization at All Costs An AI with long term planning capabilities may prioritize resource optimization to achieve its objectives, which could lead to excessive consumption of resources, monopolization or unethical practices. Superintelligence risk An AI with the capacity for long term planning could potentially become super intelligent, surpassing human intelligence and understanding such an AI may be difficult to control or predict, increasing the risks associated with its actions. Lastly, the projected availability of Gemini in different sizes and capabilities indicates DeepMind's intent to cater to various user needs. From lightweight models suitable for mobile applications to behemoths crunching data in the cloud, the flexibility of Gemini could redefine AI accessibility. In wrapping up Google, DeepMind's Gemini represents a bold step forward in AI's evolution. With a rich blend of strategies from gameplay and language understanding bolstered by undisclosed innovations, Gemini stands to reshape the landscape of artificial intelligence and its applications in the modern world.
