About This Week’s Covers

The category cover theme for this week’s AI newsletter is The White Lotus Season 3. I had Claude 3.7 automatically write prompts for 40 categories and send them to Black Forest Labs’ Flux Ultra as a batch. One Python command generates all the images. Keep in mind that if I simply wanted to create a great image with AI, I could iterate through prompts and achieve solid results within a few runs. This is purposefully done the hard way.

Every week, my category images are first tries, no redos. They are complicated and even include their own text in the prompt. For example, MidJourney and DALL·E 3 simply can’t handle them (total disaster). Ideogram and Flux are the two best for this type of prompt. The Thai tapestry theme based on the White Lotus open credits turned out pretty well. With a few refinements in the prompts and touch-ups in Photoshop, these would be solid.

Here are my favorite six of the forty category covers this week:

This Week By The Numbers

Total Organized Headlines: 453

This Week’s Executive Summaries

This week was especially busy because Nvidia had their big conference which included a lot of product announcements. There was also a lot of agent news across the board. I organized over 400 headlines and 21 of them merited executive summaries.

I’m noticing a theme of Nvidia trying to ensure that they remain the computing chip of choice by giving away a ton of software and AI models that are optimized for their chips. It makes sense that Nvidia would open source as much as they can to get a lock on the ecosystem so people are, for lack of a better term, addicted to their chips.

Nvidia released two open models that allow AI agents to reason. Nvidia also released two major simulators for robotic training in virtual reality. They also introduced a new foundation model for robots to understand their environment through vision, including interacting with things they’ve never seen before. Nvidia also released a data platform to help companies maximize the integration between their data and AI tools through a special Nvidia data architecture. Lastly, Nvidia released two open-sourced multilingual speech-to-text and text-to-speech models, which also support translation.

Anthropic has given its AI the ability to search the web, which is a big improvement; however, it’s a little bit disappointing that you can’t paste in a web address and tell it to browse or refer to it. Seems a bit limited.

Manu, the extensively hyped Chinese reasoning agent, has entered its third week of attention with many people getting accepted into the beta group and coming back with stories of incredible capabilities. For example, this week someone ran a professional-level stock analysis in one hour that would’ve taken about two weeks for a human to accomplish.

A new study came out that shows that AI agents are doubling their abilities every seven months. This is kind of a new Moore’s Law for agents. It implies that within five years, AI agents will be able to quickly accomplish things that used to take humans weeks.

OpenAI has continued evolving their agent API with a new SDK that allows anyone to integrate their service with language models through structured interfaces. This is potentially a huge deal to allow plain language interaction with agents that can complete complicated transactional tasks. It’s a direct counterpoint to Anthropic’s MCP service. It’s only going to be a matter of weeks until we see integrations where agents can go out and purchase things effectively or make reservations, etc.

OpenAI’s chief product officer predicts that this year will mark the moment that AI permanently outperforms humanity at coding.

xAI has purchased a video foundation model company, which implies that it’s going after video companies like Runway and Sora, while also potentially looking towards building world models for robot embodiment and autonomous vehicles.

Payment processing provider Stripe has created an incredible template to show how companies can give discrete documentation to AI models through a markdown document. I highly recommend opening the document and reading how well Stripe has articulated what they can do so that a prompt engineer can simply build tools integrating with Stripe without knowing any technical details. It’s really a pivotal moment that is going unsung.

Meta’s open-source model Llama has hit 1 billion downloads.

OpenAI has released three advanced audio models, which include text-to-speech and speech-to-text with better transcription than the previous leader, Whisper. The text-to-speech is impressive as well because it can be manipulated and adjusted for inflections and tone, depending on the use case. The goal is to give agents voices for interaction with people.

There are a few more executive summaries in addition to these and also 15 separate compelling visuals of the week to check out. For the real gluttons for punishment, I also selected the remaining top 50 headlines of the week at the bottom. There’s no sign of slowing down.

Anthropic’s Claude Can Search The Web
Claude now searches the internet to provide timely information with direct source citations. This combines Claude’s knowledge base with real-time data, improving accuracy for queries requiring current information. Publisher fears come true, as users receive sources in conversation format rather than separate search results. Web search is currently available to paid US users. The feature aims to help sales teams analyze industry trends, financial analysts assess market data, researchers build stronger proposals, and shoppers compare products across multiple sources.

Claude can now search the web \ Anthropic https://www.anthropic.com/news/web-search

“Web search is now available in claude dot ai. Claude can finally search the internet! https://x.com/alexalbert__/status/1902765482727645667?s=46

NVIDIA Introduces Open Reasoning AI Models for Building AI Agents
NVIDIA launched the Llama Nemotron family of open reasoning models designed to help developers build AI agents that can work independently or in teams to solve complex tasks. Post-trained by NVIDIA, these models improve multistep math, coding, reasoning and decision-making with up to 20% better accuracy and 5x faster inference than competing open models. Available in Nano, Super, and Ultra sizes as NVIDIA NIM microservices, the models are optimized for different deployment scenarios from single PCs to multi-GPU servers. Major companies including Microsoft, SAP, ServiceNow, Accenture, and Deloitte are already integrating these models to enhance their enterprise AI offerings. NVIDIA is also releasing supporting tools including the AI-Q Blueprint and AgentIQ toolkit to help businesses connect knowledge to AI agents that can perceive, reason, and act.

NVIDIA Launches Family of Open Reasoning AI Models for Developers and Enterprises to Build Agentic AI Platforms | NVIDIA Newsroom https://nvidianews.nvidia.com/news/nvidia-launches-family-of-open-reasoning-ai-models-for-developers-and-enterprises-to-build-agentic-ai-platforms

“NVIDIA has announced their first reasoning models, a new family of open weights Llama Nemotron models: Nano (8B), Super (49B) and Ultra (249B) From our early testing, @nvidia’s Nemotron Super 49B scores 64% on GPQA Diamond in reasoning mode and 54% in non-reasoning mode – we https://x.com/ArtificialAnlys/status/1902386178206429434

NVIDIA Releases Two Major Simulation Tools for AI and Robotics
NVIDIA Launches Cosmos-Transfer for Advanced World Simulation NVIDIA released Cosmos-Transfer, a world generation model that creates simulations based on multiple spatial inputs like segmentation, depth, and edge data. The system allows users to assign different weights to various inputs across locations, enabling highly controllable simulations. Applications include robotics training and autonomous vehicle data enrichment, with NVIDIA demonstrating real-time generation capabilities using their GB200 NVL72 hardware. NVIDIA Partners with DeepMind and Disney on “Newton” Physics Engine NVIDIA, Google DeepMind, and Disney Research introduced Newton, an open-source physics engine for robotics simulation built on NVIDIA Warp. The engine offers GPU acceleration, works with existing frameworks like MuJoCo, and includes differentiable physics for more accurate simulations. Disney is already implementing the technology to develop entertainment robots, including Star Wars-inspired BDX droids, while the system helps address the gap between simulation and real-world robot performance.

“Jensen just announced Newton, an open-source physics engine for robotics simulation developed by NVIDIA, Google DeepMind, and Disney Research. ⦿ Built on NVIDIA Warp, it offers GPU-accelerated performance and compatibility with frameworks like MuJoCo and Isaac Lab. ⦿ Features https://x.com/TheHumanoidHub/status/1902091038342484239

“Nvidia just released Cosmos-Transfer1 on Hugging Face Conditional World Generation with Adaptive Multimodal Control https://x.com/_akhaliq/status/1902187161841000938

“21/ @sundarpichai: “I’m really excited about the next phase of our partnership (with NVIDIA) as we work together on agentic AI, robotics and bringing the benefits of AI to more people around the world.”” / X https://x.com/AtomSilverman/status/1902087731100209624

NVIDIA Introduces GR00T N1 Foundation Model for Humanoid Robots
NVIDIA released GR00T N1, an open foundation model designed to simplify humanoid robot development. This cross-embodiment AI system processes images and language to perform manipulation tasks across different environments and within varying robot forms. Trained on extensive humanoid datasets and synthetic data, GR00T N1 enables robots like the Fourier GR-1 to grasp objects, transfer items between arms, and complete multi-step tasks. The model features a dual architecture combining a vision-language system for reasoning with a diffusion transformer that translates plans into physical movements. NVIDIA’s training approach combines internet-scale video data with synthetic simulations (generating 750,000 trajectories in just 11 hours) and real robot demonstrations, improving performance by 40% compared to using only real-world data. The 2B parameter version available today is the first in a planned series of customizable models.

Accelerate Generalist Humanoid Robot Development with NVIDIA Isaac GR00T N1 | NVIDIA Technical Blog https://developer.nvidia.com/blog/accelerate-generalist-humanoid-robot-development-with-nvidia-isaac-gr00t-n1/

“Nvidia just released GR00T-N1-2B on Hugging Face open foundation model for generalized humanoid robot reasoning and skills https://x.com/_akhaliq/status/1902124817228194289

“Excited to announce GR00T N1, the world’s first open foundation model for humanoid robots! We are on a mission to democratize Physical AI. The power of general robot brain, in the palm of your hand – with only 2B parameters, N1 learns from the most diverse physical action dataset https://x.com/DrJimFan/status/1902117478496616642

“Nvidia dropping a 2B open foundation model for humanoid robots on the hub!🔥 https://x.com/reach_vb/status/1902120408742080558

NVIDIA Isaac GR00T N1: An Open Foundation Model for Humanoid Robots | Research https://research.nvidia.com/publication/2025-03_nvidia-isaac-gr00t-n1-open-foundation-model-humanoid-robots

“NVIDIA announced GR00T N1, the first fully customizable open-source humanoid robot foundation model, designed to advance general-purpose robotics. ⦿ A dual-system AI inspired by human cognition: System-1 handles fast, intuitive actions, while System-2 enables methodical https://x.com/TheHumanoidHub/status/1902083623207211507

“NEO Gamma vacuuming the floor at NVIDIA GTC Would you rather have a vacuum cleaning bot or a versatile humanoid that can use your manual vacuum and other cleaning tools in addition to tackling many other home chores? You’ll have to wait for the latter. https://x.com/TheHumanoidHub/status/1902129279905001738

“Jim Fan at NVIDIA GTC: “This year is the year for humanoid robots.” Hyped for Jensen’s keynote tomorrow! [10am PT] https://x.com/TheHumanoidHub/status/1901738197333516555

“At GTC 2025, Jensen Huang made several AI and robotics reveals: —Next-gen GPUs: Blackwell Ultra (2025), Vera Rubin (2026), and Feynman (2028) —GR00T N1 open AI for humanoids —DGX Spark & DGX Station personal AI computers —Newton robotics physics engine https://x.com/rowancheung/status/1902249965780406666

Nvidia Partners with Storage Companies to Integrate Large Data Sets with AI Systems
NVIDIA announced a new system called the AI Data Platform that helps companies use their stored information more effectively with AI. Leading storage companies like Dell, IBM, and HP are using this design to build integrated and ‘smarter’ storage systems. These optimized architectures can quickly answer complex questions about company data without requiring technical expertise. The technology uses NVIDIA’s specialized computer chips to process information faster while using less electricity. This means businesses can get useful insights from their data almost instantly, whether that data is in documents, images, or videos. Companies plan to start offering these improved storage solutions this month. The trend here is that in NVIDIA is trying to become the dialtone for everything computing.

NVIDIA and Storage Industry Leaders Unveil New Class of Enterprise Infrastructure for the Age of AI | NVIDIA Newsroom https://nvidianews.nvidia.com/news/nvidia-and-storage-industry-leaders-unveil-new-class-of-enterprise-infrastructure-for-the-age-of-ai?ncid=ref-inpa-350226

“13/ NVIDIA AI Data Platform: New reference architecture to transform enterprise data systems from file storage to knowledge serving via AI query agents collaborating with data storage leaders to create a platform that serves knowledge via AI query agents rather than just files” / X https://x.com/AtomSilverman/status/1902087716462072234

NVIDIA Releases Two Open Source Multilingual Speech and Translation Models
NVIDIA has open-sourced Canary 1B and 180M Flash, high-performance speech recognition and translation models that rank second on the Open ASR Leaderboard. The models achieve over 1,000 times real-time processing speed and come in two sizes (880M and 180M parameters) optimized for on-device use. They support English, German, French, and Spanish with capabilities for word-level and segment-level timestamps. The models demonstrate improved accuracy with fewer hallucinations and are available under a CC-BY license that permits commercial applications.

“NEW: Nvidia just open sourced Canary 1B & 180M Flash – multilingual speech recognition AND translation models 🔥 > Second on the Open ASR Leaderboard > Achieves greater than 1000 RTF 🤯 > 880M & 180M sizes – perfect for on-device > Supports word-level & segment-level timestamps https://x.com/reach_vb/status/1902730989811413250

AI Agent Tool, Manus, Contininues Hype Cycle into Third Week
Manus, the now viral agent combining research capabilities with computer operation skills, has continuted demonstrating impressive performance on complex tasks. In testing, it completed what would typically be two weeks of professional-level stock analysis in approximately one hour. The China/Singapore-based company reports using Claude Sonnet 3.5 for execution alongside a specially post-trained Qwen model for planning. Founder Yichao suggests that developing agentic capabilities may be more about alignment than foundational technology. Early users highlight both the tool’s functional capabilities and streamlined user experience, with examples including the ability to generate 3D games through simple prompts.

“Got access and it’s true… Manus is the most impressive AI tool I’ve ever tried. – The agentic capabilities are mind-blowing, redefining what’s possible. – The UX is what so many others promised… but this time it just works. prompt: “code a threejs game where you control a https://x.com/victormustar/status/1898505307896131708

“Manus, the new AI product that everyone’s talking about, is worth the hype. This is the AI agent we were promised. Deep Research+Operator+Computer Use+Lovable+memory. Asked it to “Do a professional analysis of Tesla stock ” and it did ~2wks of professional-level work in ~1hr! https://x.com/deedydas/status/1898444603071795378

AI Agents Are Doubling Their Capabilities Every 7 Months
AI Agents Getting Better at Complex Tasks, Still Fall Short on Reliability A recent study shows AI systems are improving at handling longer tasks, with agent capabilities doubling approximately every 7 months – similar to a “Moore’s Law for AI agents.” The research found that successful AI runs cost less than 10% of what human software engineers would charge for the same work. Despite this progress, these systems aren’t consistently reliable yet. Researchers predict that within five years, AI agents could independently complete software projects that currently take humans days or weeks to finish.

“A new paper shows that AI agents are improving rapidly at long tasks – but they aren’t reliable yet. That being said this feels significant: “more than 80% of successful runs cost less than 10% of what it would cost for a human [L4 software engineer] to perform the same task.” https://x.com/emollick/status/1902443733158609005

“When will AI systems be able to carry out long projects independently? In new research, we find a kind of “Moore’s Law for AI agents”: the length of tasks that AIs can do is doubling about every 7 months. https://x.com/METR_Evals/status/1902384481111322929

Measuring AI Ability to Complete Long Tasks – METR https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/

Open AI’s Agent API and SDK Mark The Start of A New Era
OpenAI introduced an Agents SDK and Responses API that streamlines AI workflow automation across tools. The platform combines chat capabilities with web search, file management, and computer interaction tools. In a demonstration, an AI agent located an invoice, analyzed billing data from a spreadsheet, and processed payment through Stripe. The system allows multiple agents to work together while monitoring performance. This development threatens many AI agent startups by eliminating the need for complex prompt engineering, custom orchestration, and extensive debugging. The move positions OpenAI as a comprehensive AI development platform but locks developers into their ecosystem, contrasting with Anthropic’s more open approach with Claude’s Open Model Context Protocol.

“Multi-Agent workflows are the future of AI. OpenAI released new Agent APIs today, and Box built an Agent that combines documents from Box and web search tools to generate answers. Enterprise devs can grab sample code from our GitHub repo to customize with their data. https://x.com/levie/status/1899522600172507617

“With Agents SDK, @OpenAI has definitely stirred up the orchestration landscape where @crewAIInc and @LangChainAI have fought hard for the pole position. But the disruption doesn’t end at agent orchestration.👇” / X https://x.com/BetterSayAJ/status/1899708901299286300

“🔷 OpenAI Drops Powerful New Tools for AI Agents! 🚀 @OpenAI 🔹 OpenAI introduces new tools & APIs to streamline logic, orchestration & interactions. 🔹 Key Features: ✅ Responses API – Chat & tool use combined for seamless interactions ✅ Built-in Tools – Web search, file https://x.com/MervinPraison/status/1899528674845610432

“Great to see @AgentOpsAi trending on github after their partnership with @OpenAI was announced with the launch of the Agents SDK @AgentOpsAi is a Python SDK for AI agent monitoring, LLM cost tracking, benchmarking, and more. Integrates with most LLMs and agent frameworks https://x.com/atulkumarzz/status/1900436286928003255

“🤖 Agents SDK—our new open-source SDK for orchestrating multi-agent workflows, improving upon Swarm. Configure agents with built-in tools, hand off tasks, add safety guardrails, and visualize execution traces for debugging and optimizing performance. https://x.com/OpenAIDevs/status/1899531857143972051

“@OpenAI Just Wiped Out 1,000 AI Agent Startups with One Announcement The AI agent landscape just changed overnight. OpenAI unveiled its Agents SDK & Responses API, making AI-powered workflow automation ridiculously simple. The demo? A Stripe AI agent that: → Found an invoice https://x.com/adugbovictory/status/1899749168383770921

“Build a multi-agent AI research team using OpenAI Agents SDK in less than 100 lines of Python Code. It combines specialized agents that plan research, search the web, and compile reports with proper citations. 100% Opensource Code with step-by-step tutorials. https://x.com/unwind_ai_/status/1901815794670960947

“@OpenAI launched its Agents SDK yesterday, and I played around with it by creating a sales playbook agent! Here’s what it can do: – Listen to sales calls and provide real-time guidance – Recommend responses to objections based on your sales playbooks (1/2) https://x.com/danteocualesjr/status/1899942964526498042

“OpenAI released new tools for building custom enterprise agents, including a Responses API with tools for web search, computer use, and file search There’s also open-source Agents SDK to orchestrate single and multi-agent systems https://x.com/adcock_brett/status/1901303264924016784

“Today, @OpenAI released the agents SDK and computer use via their API. 3h later I built an agent with it that applies to hundreds of jobs for me 🤯🤯🤯 It was honestly not that hard and I did not write a single line of code, this was all vibecoded with Cursor. The year of https://x.com/know_u_self/status/1899567846591484359

API Reference – OpenAI API https://platform.openai.com/docs/api-reference/realtime-client-events/session

“Today, @OpenAI launched their Agents SDK developer framework. Until now, agents have mostly been chat-oriented (chat in, chat out), but agents will increasingly be action-oriented (data in, action out). We, @stripe, would love to show you a few financial agents we’ve built! ⤵️” / X https://x.com/jeff_weinstein/status/1899543525198614996

“We’ve been working with OpenAI for the past few weeks testing the Agents SDK and are launching as a tracing provider. At @AgentOpsAI, we will be sharing best practices for getting started with the SDK. Here is everything you need to get started (🧵): https://x.com/AgentOpsAI/status/1899870669397205125

“OpenAI just launched the Agents SDK – a simple yet powerful toolkit for building AI apps that can actually do things in the real world! I summarized everything you need to know about it and how it is going to be a game changer for anyone building agents. https://x.com/AtomSilverman/status/1899511053601698073

AI Code Models Expected to Surpass Top Human Programmers Next Year
OpenAI’s Chief Product Officer Kevin Weil predicts that 2025 will mark the moment AI permanently outperforms humans at programming. Weil points to AI’s rapidly improving competitive coding abilities, forecasting that an AI system will reach the number one position in coding competitions next year. He frames this shift as democratizing software creation, suggesting it will enable anyone to build whatever they imagine regardless of technical background.

“”This is the year that AI gets better than humans at programming forever. And there’s no going back.” OpenAI CPO Kevin Weil highlights the rapid improvement of AI models in competitive coding, predicting one will reach the number 1 spot in 2025. He says AI surpassing humans in https://x.com/vitrupo/status/1901258363364897016

xAI Acquires Video AI Startup Hotshot
Elon Musk’s xAI has acquired Hotshot, a startup that developed three video foundation models over the past two years. The acquisition positions xAI to compete directly with established video AI companies like Runway and Sora. This move expands Musk’s AI strategy beyond text-based systems (where xAI already competes with OpenAI and Google) and image generation into comprehensive video capabilities. The acquisition suggests xAI is building toward a complete multimodal AI platform that can handle text, images, and video generation.

“xAI acquired an AI video startup. Elon about to start gunning for Runway, Sora, Veo too. Damn. Isn’t he already gunning for OpenAI, Google and Perplexity with deep search and deep think? Bro also dropping a Midjourney class image generator while gunning for Character AI and” / X https://x.com/bilawalsidhu/status/1901723862720778267

Stripe Adds Machine-Readable Documentation for AI Tools
Stripe has integrated its documentation with large language models by adding /llms.txt and Markdown formats to its developer resources. This update allows developers to easily import Stripe’s technical information directly into AI systems. The new machine-readable format enables more efficient transfer of payment processing knowledge into developers’ preferred AI tools, streamlining implementation processes.

Stripe Developers on X: “We’ve added /llms.txt and Markdown to @stripe docs: https://t.co/mG9FIaXLTS Use the .md pages to quickly move Stripe knowledge into your LLM of choice. 📄 https://t.co/qkxhlLtHcL” / X https://x.com/StripeDev/status/1901662810897101256

Google Launches Gemini Robotics to Bridge AI and Physical World
Google DeepMind has introduced Gemini Robotics, a model built on Gemini 2.0 that can handle tasks it hasn’t encountered during training. This AI system adds physical actions as an output capability, allowing direct robot control. The company also unveiled Gemini Robotics-ER, which provides advanced spatial understanding for roboticists. As part of this initiative, Google is partnering with Apptronik to develop humanoid robots powered by Gemini 2.0, while working with selected testers to refine the technology. These models aim to enable robots to perform a wider range of real-world tasks than previously possible.

“Gemini Robotics model is able to generalize to tasks it has never seen before in training. https://x.com/TheHumanoidHub/status/1900262977582096576

“Meet Gemini Robotics: our latest AI models designed for a new generation of helpful robots. 🤖 Based on Gemini 2.0, they bring capabilities such as better reasoning, interactivity, dexterity and generalization into the physical world. 🧵 https://x.com/GoogleDeepMind/status/1899839624068907335

Perplexity Launches Model Context Protocol
Perplexity has released its Model Context Protocol (MCP) server for Sonar, enabling AI assistants to perform real-time web searches. This integration allows Claude and other AI systems to access current information from across the internet, delivering more timely and accurate responses. Developers can now implement this capability through the Perplexity API, giving their preferred AI applications access to up-to-date information.

“The Perplexity API Model Context Protocol (MCP) is now available. We’ve built an MCP server for Sonar, giving AI assistants real-time, web search research capabilities. Powered by Perplexity, Claude can now search the web and deliver real-time and accurate insights on demand. https://x.com/perplexity_ai/status/1899849114583765356

“Perplexity API now supports MCP. You can use it give any of your favorite AIs real time information, eg: Claude. https://x.com/AravSrinivas/status/1899850017546129445

AskPerplexity can now ingest videos and offer explanations
I don’t have a lot of context for this one, but it looks like you can tag perplexity in a tweet and have it watch a video and tell you what is in it. If so, that’s pretty cool!

“AskPerplexity can now ingest videos and offer explanations! https://x.com/AravSrinivas/status/1901840001866023146

Meta’s Llama AI Models Hit 1 Billion Downloads
Meta announced that its Llama AI models have reached the milestone of 1 billion downloads. The company credits the open source AI community for contributing significantly to Llama’s development and improvement over the past two years. Meta shared a recap highlighting collaborative achievements and expressed gratitude for the partnership and community enthusiasm that continues to enhance Llama’s capabilities.

“Llama has surpassed 1 billion downloads 🚀 And we would not be where we are today without the contributions of the entire open source AI community – thanks to each of you who helped make Llama better. We threw together a short recap of some of the amazing things we have built https://x.com/Ahmad_Al_Dahle/status/1902001873781194956

Creative Industries Unite Against AI Copyright Exemption Push
OpenAI and Google are urging the US Administration to remove copyright protections for AI training data, prompting strong opposition from entertainment industry professionals. In a formal response to the Administration’s AI Action Plan, representatives from film, television, music and other creative fields argue that America’s global AI leadership shouldn’t undermine the $229 billion creative sector that supports 2.3 million jobs. They emphasize that tech companies valued in the trillions could simply license content legally rather than seeking special exemptions. The group warns this issue extends beyond entertainment to all knowledge industries, potentially threatening America’s intellectual property leadership across multiple fields.

Hollywood’s response to AI Action Plan https://s3.documentcloud.org/documents/25590860/hollywoods-response-to-ai-action-plan.pdf

OpenAI Launches Advanced Audio Models for Voice Applications
OpenAI has introduced three audio models in its API to enhance voice interaction. Two new speech-to-text models outperform OpenAI’s previous Whisper model, improving transcription accuracy in challenging conditions like accents and noisy environments. The text-to-speech model allows developers to control not just content but delivery style and emotion with instructions like “talk like a sympathetic customer service agent.” These models are fully integrated with OpenAI’s Agents SDK, enabling developers to build voice agents with minimal code. The release builds on OpenAI’s recent focus on agent capabilities, extending their utility beyond text-based interactions to more natural spoken communication.

Introducing next-generation audio models in the API | OpenAI https://openai.com/index/introducing-our-next-generation-audio-models/

“🔊 Three new audio models for you today! * A new text to speech model that gives you control over timing and emotion—not just what to say, but how to say it * Two speech to text models that meaningfully outperform Whisper All available in the API + fully integrated into our” / X https://x.com/kevinweil/status/1902769861484335437

“We’re also holding a radio contest. 📻 Tweet out your https://t.co/1kI0L2V1jW TTS creations (hit “share”). The top three most creative ones will win a Teenage Engineering OB-4. Keep it to ~30 seconds, and be creative with voice, delivery, pronunciation, or even tone changes” / X https://x.com/OpenAIDevs/status/1902773659497885936

“Three new state-of-the-art audio models in the API: 🗣️ Two speech-to-text models—outperforming Whisper 💬 A new TTS model—you can instruct it *how* to speak 🤖 And the Agents SDK now supports audio, making it easy to build voice agents. Try TTS now at https://x.com/OpenAIDevs/status/1902773579323674710

Open Source Speech Model Outperforms Major Tech Companies
OPEN SOURCE SPEECH MODEL OUTPERFORMS MAJOR TECH COMPANIES Kokoro-82M has become the highest-ranked open weights text-to-speech model in the Artificial Analysis Speech Arena, scoring 1086 ELO points. This places it ahead of solutions from major tech companies while trailing only proprietary models from OpenAI, Cartesia, and ElevenLabs. The 82M-parameter model uses StyleTTS 2 architecture, supports multiple languages including English, French, and Hindi, and only required $1,000 and several hundred hours of audio to train. At $0.63 per million characters when run on Replicate, it costs approximately 100 times less than some leading proprietary alternatives.

“Kokoro-82M v1.0 is now the leading open weights Text to Speech Model in Artificial Analysis Speech Arena! Kokoro currently scores a Speech Arena ELO of 1086, just behind proprietary speech models from OpenAI, Cartesia, and ElevenLabs. For the first time, this puts an open https://x.com/ArtificialAnlys/status/1902762871106441703

Adobe Introduces Horribly Named “Experience Platform Agent Orchestrator”
Adobe’s Experience Platform Agent Orchestrator deploys specialized AI agents that handle specific marketing tasks such as audience targeting, content creation, and site optimization. These agents operate within Adobe’s enterprise applications and can integrate with third-party systems, streamlining workflow automation across different marketing platforms. Buzzword bingo unlocked! But you get the idea.

Adobe Experience Platform Agent Orchestrator is designed to unlock your full potential, making personalization at scale easier than ever. https://x.com/AdobeExpCloud/status/1902042855520305191

Stability.ai Creates 3D Videos From 2D Images
Stability.ai released Stable Virtual Camera, a diffusion model that transforms still images into 3D videos with realistic depth and camera movement. The system, currently in research preview, can generate videos following user-defined camera paths or choose from 14 preset camera movements including 360° rotations, spirals, and dolly zooms. Unlike traditional approaches, Stable Virtual Camera doesn’t require complex 3D reconstruction techniques and can work with as few as one or as many as 32 input images.

Introducing Stable Virtual Camera: Multi-View Video Generation with 3D Camera Control — Stability AI https://stability.ai/news/introducing-stable-virtual-camera-multi-view-video-generation-with-3d-camera-control

15 AI Visuals and Charts: Week Ending March 21, 2025

“An introduction to the Model Context Protocol (MCP), a new standard enabling AI agents to interact with diverse APIs in a unified way. I created the first version for a commercetools MCP server. https://x.com/lgiavedoni/status/1899778198642069916

“🔴Excited to share my TED AI talk! How we advance AI into an era of superintelligence, with true *practical* impact in our world. ❓I ask: For all the groundbreaking milestones we’ve reached, all the intelligence benchmarks we’ve shattered, for the $53B invested in generative https://x.com/stephzhan/status/1902377745335947580

NVIDIA Isaac GR00T N1 for Complex Manipulation Tasks – YouTube https://www.youtube.com/watch?v=H2wcr8UYsnI

NVIDIA Isaac GR00T N1: An Open Foundation Model for Humanoid Robots – YouTube https://www.youtube.com/watch?v=m1CH-mgpdYg

“Stability AI unveiled Stable Virtual Camera, a new diffusion model It transforms single images into 3D videos with 14 dynamic camera paths Currently in research preview under a non-commercial license! https://x.com/rowancheung/status/1902250078913360345

“google’s ai is finally based again prompt: “edit the image to add massive flames coming out of the front two nozzles of the flame thrower, having nice layers of glow, cinematic lighting, billowing fiery tendrils. turn the person in the background into a storm trooper.” https://x.com/bilawalsidhu/status/1901731029800608031

“Gosh y’all know I love 3d gaussian splats — but I love it even more when you line up your exterior & interior 3d scan and make a killer transition like this https://x.com/bilawalsidhu/status/1901156838483104013

“the level of scene understanding the new gemini model exhibits is wild. you’d need multiple controlnets to get this result, and now it’s just a text prompt. prompt: edit this image to create a 3d wireframe representation of every unique object and subject in this scene. it https://x.com/bilawalsidhu/status/1901413678202892369

Thoughts on Claude Code / AI workflows https://docs.google.com/presentation/d/1XRIflchTrZR2aqvxd7dROu8-76mf5N4Priwk8l9wNns/edit?slide=id.g3412c9db504_2_108#slide=id.g3412c9db504_2_108

“Photoshopping will never be the same. Gemini 2.0 Flash in a @Gradio app = 🤯 https://x.com/fdaudens/status/1901704690598842864

“I find playing with Gemini multimodal image generation to be really fun. Took a pic: “turn the bottles into a Saturn V complete with tiny ground crew. Add a neon sign to the cups saying ‘moon’ with an up arrow” “Make the rocket out of legos. Make the crew ducklings on stilts” https://x.com/emollick/status/1901370982557794658

“Using Gemini Flash Experimental to ruin art by adding ice cream. https://x.com/emollick/status/1900056829683462234

GTC March 2025 Keynote with NVIDIA CEO Jensen Huang – YouTube https://www.youtube.com/watch?v=_waPvOwL9Z8&t=1885s

DiffPortrait360: Consistent Portrait Diffusion for 360 View Synthesis https://freedomgu.github.io/DiffPortrait360/

“Watch for a 14min demo of me using Manus for the 1st time. It’s *shockingly* good. Now imagine this in 2-3 years when: – it has >180 IQ – never stops working – is 10x faster – and runs in swarms by the 1000s AGI is coming – expect rapid progress. https://x.com/mckaywrigley/status/1898756745545252866