0 Members and 90 Guests are viewing this topic.
In this video I cover Google's new Agent2Agent Protocol, what it can do, who is on board and who isn't.⏱️Time Stamps:00:00 Intro00:06 Google Cloud Next Developer Keynote02:11 Announcing the Agent2Agent Protocol (A2A)02:41 Anthropic: Model Context Protocol (MCP)03:07 Agent2Agent Protocol (A2A) Diagram04:52 A2A Protocol Github08:24 A2A Conceptual Overview09:56 50 Partners Contributing to the Agent2Agent Protocol
In this episode, recorded at the 2025 Abundance Summit, Joshua Xu dives into HeyGen, the future of AI avatars, and Steve Brown displays a use case for AI clones. Recorded on March 10th, 2025Views are my own thoughts; not Financial, Medical, or Legal Advice.Joshua Xu is the co-founder and CEO of HeyGen, an AI-powered video creation platform revolutionizing how businesses produce content by making video production significantly faster, cheaper, and scalable across languages. With a background in software engineering at Meta and Bloomberg LP, Xu brings deep technical expertise to his role, driving HeyGen?s rapid growth to over $35 million in annual recurring revenue. A graduate of Duke University with a degree in Electrical Engineering and Computer Science, he?s passionate about leveraging AI to democratize high-quality video communication for companies around the world.Steve Brown is a technologist and filmmaker passionate about media and innovation that strengthen human connection and sustainability. With a Physics degree from Stanford, he?s founded and sold two tech startups and developed award-winning documentaries. As Chief AI Officer at Abundance360, he builds tools that help people harness AI and exponential tech for creativity and impact.Learn more about HeyGen: https://www.heygen.com/ Learn more about Abundance360: https://bit.ly/ABUNDANCE360 For free access to the Abundance Summit Summary click: https://bit.ly/Diamandisbreakthroughs --------------------------------------------Chapters00:00 - The Future of Digital Avatars02:48 - HeyGen's Journey and Technology05:54 - The Role of AI in Business Communication09:00 - Creating Engaging Content with HeyGen11:58 - Personalization and User Experience15:10 - Trust and Safety in Avatar Creation17:56 - The Intersection of AI Agents and Digital Twins21:00 - The Future of Video Communication24:11 - Real-World Applications and Use Cases27:06 - The Importance of Storytelling and Legacy29:51 - Health and AI: A Personal Journey33:00 - The Power of Collaborative AI Agents
At Google DeepMind, researchers are chasing what?s called artificial general intelligence: a silicon intellect as versatile as a human's, but with superhuman speed and knowledge.
Google?s New AI Mode Can Explain What You?re Seeing Even if You Can?tWhat if your phone could see the world for you?and explain it in real time? Google just unveiled a groundbreaking AI feature that does exactly that. Whether you're struggling to recognize an object, translate a sign, or navigate an unfamiliar place, this new tool acts like a real-time visual interpreter. In this video, we test its limits and show you how it could change the way we interact with the world.From identifying obscure plants to reading complex street signs in foreign languages, we put Google?s AI through real-world scenarios. Imagine pointing your camera at a menu in Tokyo and getting instant translations, or having the AI describe a landmark?s history just by looking at it. The implications for travelers, students, and even people with visual impairments are huge.But how does it actually work? We?ll break down the neural networks and sensor fusion that power this feature?and why it?s different from existing image recognition tools. Plus, we?ll explore the privacy considerations: What happens to all those camera images? Could this technology eventually recognize faces or sensitive information?Can Google?s AI really explain anything you see? How accurate is it compared to human vision? What languages and objects does it support? Could this replace tour guides and translators? This video answers all these questions. Watch to the end for our live demo?you won?t believe some of the things it can do!
Even chance that communication with dolphins requires a mix of body language and sound (like humans).May need a robot dolphin model to complete the communication.
The human brain has different regions for various things. Would make sense that the first proper AGI will not just be a mixture of experts but a mixture of architectures.Exactly. Integrating LLMs with Reinforcement Learning models has been proven to be incredibly useful. A next step that many are beginning to propose is somehow extracting the "knowledge" of video generators and integrating it into LLMs, because a good video generator must already "understand" somehow the 3D world and the physical laws in order to produce believable videos, something which a model based purely on text lacks. Better holistic models also need 2 things that are currently lacking: 1. Real time few-shot learning, that is, being able to teach something to the model via direct interaction with it and the model being able to integrate that new knowledge immediately if finds it coherent and consistent without any further need for data or retraining. 2. A "deductive" tool: all current models are based on inductive reasoning and probability, which makes them prone to error when dealing with hard logic. This is ironic given that our computers are deterministic in nature, but DNN don't use any axioms or conjectures as a starting ground to ensure epistemological consistency like we do.
DeepSeek has introduced a powerful new AI system called DeepSeek-GRM that teaches itself how to think, critique, and improve its own answers using a method called Self-Principled Critique Tuning (SPCT). This approach allows their 27B model to outperform even massive models like GPT-4o in several benchmarks by using repeated sampling and meta reward models. Meanwhile, OpenAI is upgrading ChatGPT with enhanced memory features and preparing to release new models like GPT-4.1, showing how fast self-improving AI is evolving.🔍 Key Topics:DeepSeek unveils DeepSeek-GRM, a 27B self-teaching AI model using SPCT Outperforms GPT-4o and Nemotron-4-340B in benchmarks like Reward Bench and PPE Introduces meta reward models and repeated sampling for smarter, more accurate outputs 🎥 What You?ll Learn:How SPCT trains AI to critique and improve its own answers without human feedback Why repeated sampling and meta RM filtering boost accuracy and flexibility What this means for smaller models, real-world applications, and future AI development 📊 Why It Matters:This video breaks down how DeepSeek-GRM is changing the AI game by proving smaller, self-improving models can match or beat giants like GPT-4o?pushing AI toward more adaptable, efficient, and intelligent systems. DISCLAIMER:This video explores DeepSeek-GRM?s architecture, training method, and benchmark results, showing its growing impact on the AI landscape and how it stacks up against top-tier models.
ChatGPT got a huge upgrade this week that lets it remember all of your conversations. Useful or creepy or both? Plus NotebookLM got a big upgrade, Midjourney v7 finally came out, and way more. Chapters:0:00 What?s New?0:58 ChatGPT Memory Upgrade5:35 HubSpot7:01 Midjourney V.79:43 Genspark11:29 Vapi AI14:10 NotebookLM Update15:40 Cove AI18:33 AIA Public Challenge19:31 Google Cloud Next20:25 Project Astra22:19 Higgsfield25:05 Copilot Update
AI has become increasingly integrated into our lives. But progress in its technological development has slowed down in the past year. A lot. Let?s take a look at why that is, and where the AI industry might be headed next.
GPT was an obscure paper 8y ago. I was lucky to see a YouTube video about it.
One thing a lot of people are missing is how much more efficient these models are becoming. They started out as total money sinks, and maybe they still are, but it's not as deep. It's not that easy to see this when the model is hidden behind a web-interface, running on mystery hardware. GPT-4 had 1.8 trillion parameters, and used radical amounts of power, o3-mini is significantly more powerful than GPT-4, but it has ~1/9th of the parameters, and that should roughly correlates to its relative power usage. I bought a 3090 2.5 years ago to run local models, both for images and code. The best it could handle at the time was maybe a 640x640 image, before things lost consistency. Today, on the exact same hardware, I can generate minutes of video, more convincing than those early images, at higher resolutions.
"Learn to learn" is probably the fastest way to say it. Perhaps "epistemologize to exercise" is a more targeted phrase.
Parroting based on probability will never be a sufficient approach for AGI. The world model you mention will be required. They need to sense and remember.
Anytime I hear someone say an LLM is on the verge of becoming an AGI I imagine some guy saying "So we've repicated a good portion of the language center of a brain. We will have a full brain any day now."
What is CUDA and why do we need it? An Nvidia invention, its used in many aspects of parallel computing. We spoke to Stephen Jones, one of the architects of CUDA at the recent GTC conference.
CUDA does NOT just "WORK" all the time in practice. As someone who builds with these systems daily, you need to align OS versions, with GPU hardware models, with GPU driver versions, with container versions, with CUDA versions, with pytorch versions, with tensorflow versions, and any other libraries you happen to be working with. It is a tremendous amount of effort. But these interoperability concerns are definitely outside the CUDA team's scope. From their perspective, they try (and succeed) at being as compatible as possible.
One part about CUDA he forgot to mention, is that many years ago there was a company called Ageia that made SDKs and cards that'd accelerate physics. Nvidia saw that, thought "hey, they're doing something similar to us" and bought it up. That technology was called PhysX.
Too bad all this is closed source proprietary
In this video I explore how Sam Altman?s hints about GPT-4.5 and GPT-5 are reshaping the AI landscape. You?ll discover the key differences between GPT-4.5 and GPT-5, learn why OpenAI?s next release could unify ?fast? and ?thoughtful? AI models, and find out how chain-of-thought reasoning could change everything from creative writing to complex problem-solving. I also discuss the biggest challenges OpenAI has faced during GPT-5?s development?from massive data requirements to persistent engineering snags?and why this model might feel closer to AGI than any chatbot yet.Chapters:0:00 - Introduction0:24 - Timeline3:20 - Release date3:38 - Why the Wait?5:01 - More problems7:25 - The Big & The Bigger8:41 - A solution9:38 - New Architecture11:32 - Deeply Multimodal Interaction12:34 - Built-in Operator & Scheduling Features13:23 - Personalization & Persistent Memory14:11 - Larger Context Windows14:56 - Visual Planning & Collaboration (Canvas)16:13 - Will GPT-5 Feel Like AGI?17:31 - Why GPT-5 Matters18:35 - It's Almost Here
When I first started using ChatGPT I assumed it would be most useful for big, complicated tasks. Instead, I?ve found its real strength lies in handling the small but frustrating problems that might crop up every day ? the things that usually slip down the to-do list.From summarizing documents to organizing scattered thoughts or speeding up decision-making, these simple prompts have made a noticeable difference to how I work. They aren?t flashy, but they solve real problems faster and more efficiently than I could on my own.After months of trial and error, these five prompts are the ones I come back to again and again ? because they actually make life easier.
In this video we extrapolate the future of AI progress, following a timeline that starts from today?s chatbots to future AI that?s vastly smarter than all of humanity combined?with God-like capabilities. Such AIs will pose a significant extinction risk to humanity.This video was made in partnership with ControlAI, a nonprofit that cares deeply about humanity surviving the coming intelligence explosion. They are mobilizing experts, politicians, and concerned citizens like you to keep humanity in control. We need you: every voice matters, every action counts, and we?re running out of time.
Stanford's Jeremy Utley reveals that "most people are not fully utilizing AI's potential." Why is that? He explains that it lies in how we approach AI. He said a simple mindset shift could be what you've been missing in the AI revolution. How are you collaborating with AI in this new era? With so many LLM tools emerging, are you truly leveraging them to enhance your creativity and productivity? As an expert in creativity and AI, Jeremy shares profound insights on how AI is transforming our creative potential.Key Insights:📌How treating AI as a teammate rather than just a tool can dramatically improve outcomes📌Why you should have AI ask you questions instead of just answering yours📌How non-technical professionals can leverage AI to achieve extraordinary results📌The difference between treating AI as a tool versus as a teammate00:00 Intro01:26 Who is Jeremy Utley?03:05 Do not Ask AI, Let It Ask You03:57 The 10X Creativity Hack06:03 I Don't USE AI06:46 Why Do Some People Produce More Creative Results Using the Same AI Tools?08:23 Treat AI As a Teammate09:30 Inspiration is a Discipline11:15 The Definition of Creativity in the Age of AI
Wow, may be the most significant paper of 2025!A team at Tsinghua has figured out how to get an AI to generate its own training data, and surpassed the performance of models trained on expert human-curated data.We may not hit another data wall between here and ASI.
While the idea of a model teaching itself with zero external data is intellectually appealing, there?s reason for caution. First, the notion of ?zero data? is misleading?these models are still built on weights trained from massive datasets, so this is more like zero new data, not zero data overall. Second, self-generated tasks can drift into triviality or circular reasoning; a model might ?succeed? by solving problems it implicitly made easier. Without external benchmarks or human oversight, it?s difficult to ensure that learning progress translates to real-world utility or generalizable intelligence. Lastly, verifiable reward functions work well in narrow domains like code or math, but extending this to ambiguous, real-world reasoning (like law, ethics, or medicine) could be far more difficult. The hype may outpace the practicality.
Generating your own training data is the basis of religion. Look how much good that has done.
We?ve always thought large language models (LLMs) like Claude, GPT-4, and Gemini were just next-word predictors?but new research from Anthropic tells a very different story. In this video, I break down their blog post ?Tracing the Thoughts of a Large Language Model? and explore what?s really happening under the hood.00:00 How LLMs work02:49 Next Word Prediction vs. Planning Ahead03:28 Interpreting LLM Reasoning04:14 Comparing LLMs to Computer Vision Models05:20 The Biology of a Large Language Model05:53 Universal Language of Thought12:00 LLMs and Mathematical Reasoning15:23 Faithfulness in Chain of Thought20:05 LLMs and Hallucinations22:06 Understanding Jailbreaks