AI News #88: Week Ending June 06, 2025 with 24 Executive Summaries, Top 22 Links, and 5 Helpful Visuals

June 7, 2025

About This Week’s Covers

This week’s covers honor the passing of musical great Sly Stone. I asked GPT-o3 to create a rubric I could use for batch-producing category covers in the spirit of various Sly and The Family Stone albums.

The main cover is a hybrid of the Sly and the Family Stone Greatest Hits album and the Figure and Optimus humanoid robots. In the background is a repeating ElevenLabs logo.

Separately, GPT-o3 created a rubric that allowed me to give it 43 one-word category names, and from those 43 single words, the API returned 43 album covers inspired by Sly Stone. These turned out pretty well!

I’ve included my favorite twelve of the covers below:

This Week By The Numbers

Total Organized Headlines: 404

This Week’s Executive Summaries

I’m still hovering at about two weeks behind, but I would not trade the time I’ve been spending with friends and family. I’m wrapping up the week of June 6th. Here are the top stories and trends you should know.

This week I’ve been underscoring that there are three levels of artificial intelligence dialogue:

1) Most people think of artificial intelligence as chatbots. Writing. Maybe images. Short videos. That’s the first level.

2) The second level consists mostly of industry folks who like to talk about AI agents. Agents are when we use language as an interface to tell our computer to complete a sweeping, broad, undefined task.

Chatbots are the last two years. Agents are the next two years.

3) However, if you talk to serious artificial intelligence nerds, they’ll tell you that the biggest thing will be robots. That’s when you combine vision and hearing and the ability to understand the world and take action within it in a physical embodiment. I tend to live in this robotic future mentally, but I understand that not everybody is there yet.

However, in Nvidia’s quarterly earnings call, CEO Jensen Huang stated that billions of robots and hundreds of millions of autonomous vehicles will be developed in the near future. Jensen is not a hyperbolic person. The robotics labs at Nvidia are my absolute favorite thing to follow in artificial intelligence. The term to look out for is robot embodiment.

If you plan to be around in 10 years, prepare yourself for robots doing insane things. They are currently being trained in simulations using Nvidia chips in thousands of alternative universes at hundreds of times the speed of human time. Just like the Matrix. The best example that I’ve seen so far is the robot dog that was trained to walk balancing on a yoga ball in thousands of simulations with varying rules of physics and then was given ‘the 1001st simulation’, which happened to be our reality. It walked on the ball on the first try.

The second amazing story of the week came from a company called Luma Labs, which launched a feature called “modify video” which allows a creator to act out almost any action anywhere, and then use a prompt to place themselves in any world they can imagine. Think green screens on steroids. The slacklining example is probably my favorite and is a must see video.

Speaking of the second level of artificial intelligence dialogue, Amazon created an Agentic AI team within their hardware lab to take Alexa to the next level. Hopefully in the future, Alexa will not just be able to set a timer, but will be able to go do things on your behalf. From booking an airline ticket to making restaurant reservations to checking on the status of a medical lab report.

When an AI can interact with media, understand media, and create media, in addition to text, it is called multimodal. Audio and vision are two examples of multimodal artificial intelligence that will eventually contribute to robotic embodiment.

In the world of audio multimodality, a company called ElevenLabs launched a new state of the art conversational and expressive speech model which might begin to replace humans as customer service agents.

HeyGen the leading tool for creating video avatars came out with a new AI studio. It allows fine control of speech, emphasis and pacing, as well as voice mirroring to replicate users’ own voices. It also integrates gesture control to link specific words with intended movements. It’s a combination of word processing and professional level video direction. I think there are many cases where real people will always be important, but I think avatars can replace quite a bit of transactional, operational, and educational videos. I certainly would rather use an avatar clone than have to figure out how to record myself professionally for a tutorial. I’ve played with HeyGen a few times in the past and I think it’s time to revisit it.

Google’s V3 continues to top video leaderboards and has completely smoked all the competition. Andrej Karpathy, one of the leading minds in artificial intelligence, has brought up the great point that video is now so easy to create that people might start using gradient descent tricks to maximize engagement or other metrics. We might end up with even more addictive content than our current social media platform algorithms deliver. For example, the Google Veo ASMR videos of knives cutting frozen glass fruits are bonkers. Every kid I know has watched them. I’ve watched them.

Speaking of social media disruption, Meta plans to fully automate their advertising tools by 2026 using artificial intelligence. Completely cutting out the creative process and middle persons.

Meta also just signed a 20 year nuclear deal to power their AI operations. Meta will be consuming the power equivalent of a city with 30,000 residents.

A while ago JP Morgan built a multi-agent AI system for investment research. It’s gone almost completely under the radar and I never see it in my news updates however this week it showed up because there is a YouTube video and interview with the creator of the system.

Google has come out with a new type of language model technology that they are calling text diffusion. Usually you think of diffusion with images, not text and evidently it’s blazingly fast. I still have no idea how it works, but I want to follow it because supposedly it might be as breakthrough as big as GPT-4.

The FDA has introduced an artificial intelligence tool that helps employees work more efficiently across scientific reviews, and data analysis. Personally, I think this is fantastic because so much of our data is in silos and artificial intelligence is the perfect way to have a conversation across these data islands. When my dad had cancer, we were talking to maybe six or seven specialists with tons of medical records. It would have been great to have an artificial intelligence interface to quickly ask questions across all of these separate systems. I would imagine when doing clinical trials across lots of different companies and groups, it would be great to have retrieval augmented generation (RAG) agents to help data mine using natural language.

OpenAI rolled out a robust coding interface called Codex that is way above my ability. If you’re into software engineering or coding, you probably already know this tool exists. If not, it’s worth reading about quickly to see just how strong AI is getting when it comes to programming.

In the worlds of both ethics and “wow, the creative assimilation is happening quickly”, AI video generation company Runway has partnered with AMC networks for video production. From my understanding, it’s basically easy quick templates, B roll, and mattes. Taking care of the easiest pieces of the creative puzzle.

Last week I spent some time talking about the death of the web browser and page views. This week, there are three new announcements that continue the trend. One of the smaller browser companies named Opera released a browser called Neon that uses AI agents to automatically complete tasks like filling out forms or booking trips while users browse the web. Another company introduced web automation agents that are able to read, click, and type on websites with 92% accuracy. And another company called Firecrawl launched a hosted service that connects web scraping abilities directly to the large frontier models.

In medical news, the FDA approved the first AI tool to predict breast cancer risk from mammograms.

Apple came out with a research paper that broke down some issues with reasoning limits within AI models. I think the title makes for some hyperbolic headlines, and of course, Apple is woefully behind with their AI integration.

Anthropic released specialized models exclusively for US national security customers and classified environments.

OpenAI’s GPT-image-1 is officially the first place champ in the text-to-image arena leaderboard. It’s been out two months now and I have to say it’s my favorite pound for pound image creation tool. Flux can still beat it for complex photorealism. I’m always hoping that MidJourney can make a comeback, but I don’t use it much anymore.

OpenAI is fighting a court order that forces the company to save all ChatGPT user logs, including deleted chats and API conversations. I have to say I agree with OpenAI on this one.

DeepSeek’s R1 model is officially tied for second best in the world. It’s now tied with Gemini 2.5 pro for second place globally. However, more importantly, it is the top open source model. The closed v. open battle is one to follow, especially politically.

Google came out with a cool android app that lets users download and run AI models from HuggingFace directly on their phones. Locally run AI is a very popular alternative for privacy fanatics but it’s also just really cool to not need an Internet connection and be able to host a decent language interface on your device. There are a ton of use cases where you want to be able to talk to a device and locally run models are perfect for this. An example might be a script that reads emails and sorts them or prioritizes inbound texts or creates summaries. Basically, this is what Siri should be doing. Google is seizing the day and going the open-source route with HuggingFace.

Last week I briefly touched on the fact that the recent finance bill included a 10-year moratorium on AI regulation that somehow snuck in at the last minute. Anthropic CEO Dario Amodei wrote an dissenting op-ed in the New York Times saying that he feels that regulation and transparency is important.

French open source start up Mistral was always my favorite open source model, prior to DeepSeek. I like how scrappy they are. Especially when they’re up against huge companies like Meta. This week Mistral launched a coding assistant to compete with Cursor and Copilot. Anything open source is worth following because open-source will continue to get better and democratize tools and resources at lower costs than closed models.

The last item is actually a pretty big one. A company called Manus showed up out of nowhere with dozens of headlines showcasing a wide variety of announcements from the past week. There were so many stories about Manus that I considered creating a new category just because of this week’s volume. I’ve included the top few in the summaries below.

It was yet another week of incredible updates. Remember to hug your friends and family! Everything else you need to know about the week in AI is below.

Robotics highlights from Nvidia’s quarterly earnings call – “billions of robots on the way”
This past quarter Nvidia released Isaac GR00T N1, which the company calls the world’s first open foundation model for humanoid robots that enables general reasoning and skill development. The chipmaker also launched Cosmo World foundation models for synthetic data generation, with companies like 1X, Agility Robotics, Figure AI, and Boston Dynamics already integrating the technology into their robotics operations. CEO Jensen Huang said billions of robots and hundreds of millions of autonomous vehicles will be developed, positioning Nvidia’s Omniverse and Isaac platforms as the infrastructure powering next-generation factories and humanoid systems.

Everything NVIDIA said about humanoid robots in yesterday’s quarterly earnings call Colette Kress (CFO): We announced Isaac GR00T N1 – the world’s first open, fully customizable foundation model for humanoid robots – enabling generalized reasoning and skill development. We also https://x.com/TheHumanoidHub/status/1928191289172054170

Luma Labs launches Modify Video for changing scenes without reshooting
Luma Labs released a video editing tool that lets filmmakers change environments, lighting, and objects in existing footage while preserving the original performance and camera movement. Modify Video can extract motion data from any video clip to drive new characters or animate objects, transform settings like turning a garage into a spaceship or switching from day to night, and edit individual elements such as wardrobe or props without affecting the rest of the shot. The tool eliminates the need for green screens, motion tracking, or expensive reshoots when directors want to reimagine scenes after filming is complete.

Introducing Modify Video. Reimagine any video. Shoot it in post with director-grade control over style, character, and setting. Restyle expressive performances, swap entire worlds, or redesign the frame to your vision. Shoot once. Shape infinitely. https://x.com/LumaLabsAI/status/1930271899420111132

Slackliner https://x.com/LumaLabsAI/status/1930271921154720133

Introducing Modify Video | Luma AI https://lumalabs.ai/blog/news/introducing-modify-video

Amazon creates agentic AI team within hardware lab
Amazon formed a new agentic AI group within Lab126, its Silicon Valley research unit behind devices like Kindle and Echo, to develop AI systems that can complete complex, multi-step tasks without human guidance. The team will focus on creating an AI framework for Amazon’s robotics operations, enabling warehouse robots to understand natural language commands and function as flexible assistants rather than single-purpose machines. This follows Amazon’s earlier release of a web browser-based AI agent and joins similar agentic AI efforts across the company’s cloud division and updated Alexa voice assistant.

Amazon’s R&D lab forms new agentic AI group https://www.cnbc.com/2025/06/04/amazons-rd-lab-forms-new-agentic-ai-group.html

Eleven Labs launches state of the art conversational and expressive speech models
Eleven Labs released two major updates: Conversational AI 2.0 for voice agents and the v3 text-to-speech model in alpha. The conversational platform adds natural turn-taking that handles interruptions and pauses, automatic language switching across 70+ languages, multimodal support for text and voice, batch calling for outbound campaigns, and built-in document retrieval with HIPAA compliance and EU data residency. The v3 speech model delivers unprecedented expressiveness with audio tags like [whispers], [sighs], and [excited], plus multi-speaker dialogue capabilities, though it requires more prompt engineering and isn’t optimized for real-time use yet.

Introducing Conversational AI 2.0 Build voice agents with: • New state-of-the-art turn-taking model • Language switching • Multicharacter mode • Multimodality • Batch calls • Built-in RAG Now fully enterprise-ready with HIPAA compliance, EU data residency, and robust https://x.com/elevenlabsio/status/1928527751956308004

Introducing Eleven v3 (alpha) – the most expressive Text to Speech model ever. Supporting 70+ languages, multi-speaker dialogue, and audio tags such as [excited], [sighs], [laughing], and [whispers]. Now in public alpha and 80% off in June. https://x.com/elevenlabsio/status/1930689774278570003

Eleven v3: Most Expressive AI Text to Speech Model Launched | ElevenLabs https://elevenlabs.io/blog/eleven-v3

HeyGen launches AI Studio for precise avatar video creation
HeyGen released AI Studio, a video creation platform that gives users granular control over AI avatars’ voice, gestures, and movements without requiring cameras or studio equipment. The platform, used by over 15 million people across 179 countries, features Voice Director for controlling speech emphasis and pacing, Voice Mirroring to replicate users’ own voices, and Gesture Control to link specific words with pre-recorded movements. AI Studio simplifies the interface to feel like familiar word processing software while providing professional-level control over avatar performance, allowing creators to fine-tune delivery details that make audiences focus on the message rather than the technology behind it.

Revolutionize Video Creation with AI Studio | HeyGen https://www.heygen.com/blog/introducing-ai-studio

omg.. this is crazy HeyGen just dropped Voice Mirroring, it can clone anyone’s voice emotion, tone and style PERFECTLY no words, just watch step by step tutorial: https://x.com/EHuanglu/status/1919696421625987220

Google’s Veo 3 tops video generation leaderboards and may change social media as we know it
Google’s Veo 3 became the first AI model to lead both text-to-video and image-to-video rankings, surpassing competitors like Kling 2.0 and Runway Gen 4. The model includes native audio generation, which creators say unlocks new content possibilities by eliminating the need to combine multiple tools for video production. AI researcher Andrej Karpathy highlighted that unlike previous video platforms that rely on human creators and recommendation algorithms, AI-generated video can be directly optimized using gradient descent to maximize engagement or other metrics, potentially creating far more addictive content than current platforms like TikTok.

Exclusive | Meta Aims to Use AI to Fully Automate Ad Creation – WSJ https://www.wsj.com/tech/ai/meta-aims-to-fully-automate-ad-creation-using-ai-7d82e249

Meta aims to enable brands to fully create and target ads using artificial intelligence by the end of next year https://x.com/fdaudens/status/1929566230471401857

Meta plans full AI automation for ads by 2026
Meta aims to let brands create and target advertisements entirely through AI by the end of 2025, according to Wall Street Journal sources. The system would allow companies to provide just a product image and budget, with Meta’s AI generating complete ads including images, videos, and text while automatically determining user targeting across Facebook and Instagram’s 3.43 billion users. The platform also plans real-time ad personalization, showing different versions to users based on factors like location, as CEO Mark Zuckerberg pushes for AI tools that deliver “measurable results at scale” in the competitive social media advertising market.

Exclusive | Meta Aims to Use AI to Fully Automate Ad Creation – WSJ https://www.wsj.com/tech/ai/meta-aims-to-fully-automate-ad-creation-using-ai-7d82e249

Meta aims to enable brands to fully create and target ads using artificial intelligence by the end of next year https://x.com/fdaudens/status/1929566230471401857

Meta signs 20-year nuclear deal to power AI operations
Meta signed a 20-year agreement with Constellation Energy to secure nuclear power for its AI and computing needs, joining other tech giants like Microsoft, Amazon, and Google in turning to nuclear energy. The deal will expand output at Constellation’s Clinton Clean Energy Center in Illinois by 30 megawatts starting in 2027, preserving 1,100 jobs and generating $13.5 million in annual tax revenue while powering the equivalent of a city with 30,000 residents. Tech companies are increasingly investing in nuclear power to meet surging AI energy demands while achieving carbon reduction goals, though experts warn the U.S. lacks commercial next-generation reactors and needs significant transmission grid improvements to handle the growing power requirements.

Meta becomes the latest big tech company turning to nuclear power for AI needs | AP News https://apnews.com/article/meta-facebook-constellation-energy-nuclear-ai-a2d5f60ee0ca9f44c183c58d1c05337c

JPMorgan builds multi-agent AI system for investment research
JPMorgan Private Bank developed “Ask David,” a multi-agent AI system that automates investment research across thousands of financial products and years of historical data. The system uses multiple specialized agents including a supervisor that coordinates tasks, separate agents for structured and unstructured data processing, an analytics agent for proprietary models, and quality control nodes for personalization and accuracy checking. JPMorgan designed the system to help financial advisors access real-time insights during client meetings, moving from manual research that was time-consuming and difficult to scale to automated analysis that can handle both database queries and unstructured documents like emails and presentations.

Learn how @jpmorgan developed a multi-agent system architecture for investment research. David Odomirok and Zheng Xue reveal how they built “”Ask David”” – a sophisticated multi-agent AI system designed to automate investment research for thousands of financial products. Watch https://x.com/LangChainAI/status/1928135137658818711

Gemini Diffusion previewed as potential breakthrough technology
Pietro Schirano, CEO of Magic Path AI and creator of Claude Engineer, gained early access to Google’s Gemini Diffusion model and calls it comparable to the GPT-4 breakthrough moment. He believes the technology will fundamentally reshape AI experiences once the model improves, describing it as one of the most important upcoming developments in artificial intelligence. Schirano previously led AI initiatives at Brex, Uber, and Facebook before founding his current company.

I got access to Gemini Diffusion for a few weeks. I believe this is one of the most important upcoming technologies in the AI space, it’s like a GPT-4 moment all over again. Once the model gets better, it will completely reshape AI experiences for us. Video is in real time. https://x.com/skirano/status/1930332481078616296

FDA launches AI assistant to streamline drug reviews and inspections
The FDA rolled out Elsa, an AI tool that helps agency employees work more efficiently across scientific reviews, inspections, and data analysis. Built in a secure government cloud environment, Elsa can summarize adverse drug events, compare product labels, generate database code, and accelerate clinical trial protocol reviews while keeping all sensitive information within the agency. The tool launched ahead of schedule following a successful pilot with scientific reviewers and will expand to additional functions like data processing as the FDA learns how employees use it.

FDA Launches Agency-Wide AI Tool to Optimize Performance for the American People | FDA https://www.fda.gov/news-events/press-announcements/fda-launches-agency-wide-ai-tool-optimize-performance-american-people

OpenAI rolls out Codex updates with internet access
OpenAI launched Codex for ChatGPT Plus users with generous usage limits and added internet connectivity during code execution. The coding tool can now install dependencies, run tests requiring external resources, and upgrade packages automatically while users maintain control over which domains and HTTP methods Codex can access.

Codex is available now available to ChatGPT Plus users! We’ve also supplied our most requested feature: an option to enable internet access for the agent (with user control over HTTP methods and domains that Codex can access).”” / X https://x.com/gdb/status/1929970095427858636

Runway partners with AMC Networks for AI video production
Video AI company Runway partnered with AMC Networks to integrate AI tools into the entertainment company’s marketing and TV development workflows. AMC Networks, which produces shows like Breaking Bad and The Walking Dead, will use Runway’s AI models to streamline promotional content creation, support campaign development, and accelerate pre-visualization during series development.

Runway News | Runway Partners with AMC Networks Across Marketing and TV Development https://runwayml.com/news/runway-amc-partnership

Three companies launch tools that could reshape web browsing
Opera released Neon, a browser that uses AI agents to automatically fill forms, book trips, and complete purchases while users browse, while H Company introduced Holo1, open-source AI models that power their Surfer-H web automation agent capable of reading, clicking, and typing on websites with 92.2% accuracy at $0.13 per task. Meanwhile, Firecrawl launched a hosted service that connects web scraping and crawling capabilities directly to AI models like Claude, Gemini, and ChatGPT through a simple API setup. These developments suggest a shift toward AI agents that can navigate and interact with websites automatically, potentially reducing the need for traditional manual browsing and clicking through web pages.

Meet Opera Neon, a browser for the agentic web Opera Neon can browse with you or for you, take action & help you get things done. Our playground to redefine what a browser can be. 🧩 Invite only. Sign up now: https://x.com/opera/status/1927645192254861746

H Company released Holo-1: 3B and 7B GUI Action Vision Language Models for various web and computer agent tasks 🤗 Holo-1 has Apache 2.0 license and @huggingface transformers support 🔥 more details in their blog post (next ⤵️) https://x.com/mervenoyann/status/1929896423765500358

Introducing our Hosted Firecrawl MCP 🔥 Near-instant setup for web scraping, crawling, and search across all major model providers like Gemini, Claude, and OpenAI who all support MCP. Perfect for IDE workflows, agents, and more. Just plug in your API key and go 👇 https://x.com/firecrawl_dev/status/1925945881162879370

FDA approves first AI tool to predict breast cancer risk from mammograms
The FDA granted authorization to Clairity Breast, an AI platform that analyzes standard mammograms to predict a woman’s five-year breast cancer risk without relying on family history or questionnaires. The system detects subtle imaging patterns in breast tissue that correlate with future cancer development, even when mammograms appear normal to doctors. In studies of over 30,000 mammograms, the tool found that 37% of women in their 40s had intermediate risk levels and 16% had high risk levels that matched older women typically targeted by current screening guidelines. Unlike traditional risk models built primarily on data from European Caucasian women, Clairity Breast was developed with intentional representation across racial and ethnic groups and requires only a standard 2D mammogram, making it accessible in areas with limited healthcare resources. The platform launches commercially in late 2025 and could enable more personalized screening schedules, supplemental imaging for high-risk women, and reduced unnecessary testing for those at lower risk.

Clairity Breast FDA Approved | Breast Cancer Research Foundation https://www.bcrf.org/blog/clairity-breast-ai-artificial-intelligence-mammogram-approved/

Apple research reveals reasoning limits in AI thinking models
Apple researchers found that AI models designed to show their reasoning process hit a “complete accuracy collapse” when problems become too complex, despite having enough computing power to continue. The study used controllable puzzle environments to test how models like those that generate step-by-step thinking perform across different difficulty levels, revealing three distinct performance zones: simple tasks where standard AI models actually outperform reasoning models, medium complexity where the thinking process helps, and high complexity where both types fail entirely. The researchers discovered that these reasoning models struggle with exact calculations, fail to use systematic algorithms, and show inconsistent logic across similar puzzles, raising questions about whether they truly understand problems or simply generate plausible-sounding explanations.

The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity – Apple Machine Learning Research https://machinelearning.apple.com/research/illusion-of-thinking

Anthropic launches Claude Gov models for national security agencies
Anthropic created specialized AI models exclusively for U.S. national security customers, with several agencies already using them in classified environments. Claude Gov models handle classified materials more effectively than standard versions, refusing fewer requests when working with sensitive information, and demonstrate improved understanding of intelligence and defense documents. The models also perform better with languages and dialects important to national security operations and can analyze complex cybersecurity data for intelligence purposes, while maintaining the same safety standards as Anthropic’s other AI systems.

Claude Gov Models for U.S. National Security Customers \ Anthropic https://www.anthropic.com/news/claude-gov-models-for-u-s-national-security-customers

OpenAI’s GPT-Image-1 tops text-to-image rankings
OpenAI’s GPT-Image-1 model achieved first place in the Text-to-Image Arena leaderboard, outperforming Google’s Imagen-3.0 by more than 50 points and showing significant improvement over OpenAI’s previous DALL-E 3 model. The Text-to-Image Arena is a platform where AI image generation models compete based on user preferences and quality assessments.

Exciting news: @OpenAI’s GPT-Image-1 takes the #1 spot in the Text-to-Image Arena! 🖼️🏆 ➤ Outperforms Google’s Imagen-3.0 by 50+ points ➤ Major leap over DALL·E 3 Huge congrats to @OpenAI! 👏 https://x.com/lmarena_ai/status/1930296340648735147

OpenAI fights court order requiring preservation of all ChatGPT conversations
OpenAI is challenging a court order that forces the company to save all ChatGPT user logs, including deleted chats and API conversations, after news organizations accused the AI company of destroying evidence in their copyright lawsuit. The company argues the order violates privacy commitments to hundreds of millions of users who chose to delete sensitive conversations containing personal financial data, trade secrets, and private thoughts. Judge Ona Wang issued the preservation order based on concerns that users might delete conversations where they accessed paywalled news content, but OpenAI claims there’s no evidence users actually do this and that the order creates significant engineering costs while potentially breaching privacy regulations and user contracts.

OpenAI slams court order to save all ChatGPT logs, including deleted chats – Ars Technica https://arstechnica.com/tech-policy/2025/06/openai-says-court-forcing-it-to-save-all-chatgpt-logs-is-a-privacy-nightmare/

DeepSeek’s R1 update ties for world’s second-best AI model
DeepSeek released an updated version of its R1 model that jumped 8 points on the Artificial Analysis Intelligence Index to score 68, tying it with Google’s Gemini 2.5 Pro for second place globally and making it the top open-source AI model. The Chinese company achieved these gains through post-training improvements without changing the underlying architecture, with the biggest jumps in math competition problems, code generation, and scientific reasoning tasks. The update demonstrates that open-source models can match proprietary competitors and that Chinese AI labs have largely caught up to US companies, while also showing that reinforcement learning techniques can deliver major intelligence improvements more efficiently than scaling up basic training.

DeepSeek’s R1 leaps over xAI, Meta and Anthropic to be tied as the world’s #2 AI Lab and the undisputed open-weights leader DeepSeek R1 0528 has jumped from 60 to 68 in the Artificial Analysis Intelligence Index, our index of 7 leading evaluations that we run independently https://x.com/ArtificialAnlys/status/1928071179115581671

Google releases app for running AI models locally on phones
Google launched AI Edge Gallery, an experimental Android app that lets users download and run AI models from Hugging Face directly on their phones without internet connection. The app offers various AI capabilities including image generation, question answering, and code writing through models like Google’s Gemma 3n, with a “Prompt Lab” feature for text summarization and rewriting tasks. Performance varies based on device hardware and model size, with larger models taking longer to complete tasks, but the offline functionality appeals to users concerned about data privacy or needing AI without internet access.

Google quietly released an app that lets you download and run AI models locally | TechCrunch https://techcrunch.com/2025/05/31/google-quietly-released-an-app-that-lets-you-download-and-run-ai-models-locally/

Anthropic CEO calls Republican AI regulation ban ‘too blunt’
Anthropic CEO Dario Amodei criticized a Republican proposal to block state AI regulation for 10 years, arguing in a New York Times op-ed that AI advances too quickly for such a lengthy moratorium. The proposal, included in President Trump’s tax cut bill, would preempt AI laws recently passed in dozens of states and has faced opposition from attorneys general across party lines. Amodei instead proposed federal transparency standards requiring AI companies to publicly disclose how they test for and address national security risks, warning that the current approach would leave neither state action nor national policy as safeguards.

Anthropic CEO says proposed 10-year ban on state AI regulation ‘too blunt’ in NYT op-ed | Reuters https://www.reuters.com/business/retail-consumer/anthropic-ceo-says-proposed-10-year-ban-state-ai-regulation-too-blunt-nyt-op-ed-2025-06-05/

Mistral launches coding assistant to compete with Cursor and GitHub Copilot
French AI startup Mistral released Mistral Code, an AI-powered coding assistant that combines multiple specialized models for tasks like code completion, search, and multi-step refactoring. The platform, built on the open-source Continue project, supports over 80 programming languages and can be deployed in cloud, on-premises, or air-gapped environments for enterprise security needs. Companies including consulting firm Capgemini and French railway SNCF are already using the tool in production, with Mistral offering customization options like fine-tuning models on private code repositories and providing admin controls for IT managers.

Introducing Mistral Code | Mistral AI https://mistral.ai/news/mistral-code

Mistral releases a vibe coding client, Mistral Code | TechCrunch https://techcrunch.com/2025/06/04/mistral-releases-a-vibe-coding-client-mistral-code/

Manus AI showcases autonomous agent capabilities across multiple domains
Manus AI demonstrated its autonomous agent technology through various user-generated examples, showing the system can build complete web applications, create educational content, generate videos from prompts, and even construct 3D models in Blender. Users reported that Manus autonomously planned and executed complex tasks like creating a face comparison website with deep learning algorithms, building educational sites with interactive diagrams, and generating animated molecular models for teaching purposes. The company partnered with E2B to scale their platform using secure sandbox environments, with the system growing 100x in just a few months while maintaining high-speed performance across diverse tools and applications.

We’re excited to work with @e2b on scaling Manus, which grew 100x from its inception until today in just a few months. Manus is building the world’s first fully autonomous agent that uses a variety of tools. These functionalities are powered by @e2b’s secure sandbox environment. https://x.com/ManusAI_HQ/status/1928433321333404037

Introducing Manus video generation. Manus transforms your prompts into complete stories—structured, sequenced, and ready to watch. With a single prompt, Manus plans each scene, crafts the visuals, and animates your vision. From storyboard creation to concept visualization—your https://x.com/ManusAI_HQ/status/1929913745503072551

Wow, we actually achieved AGI I’ve been using Manus AI the last 24 hours straight and it’s capabilities are mindblowing It’s literally your own AI employee. If a human did this it would cost me $200k Manus does it for free Here is how I had it design and build an entire Saas: https://x.com/AlexFinnX/status/1901356733165121952

Wow. AI agents are here I’ve been using Manus AI the last week and it actually is insane While at dinner I prompted it to build me a full app. By the end of dinner the app was done In this video I walk through Manus and show you how to build incredible apps (ya, bookmark this) https://x.com/AlexFinnX/status/1903136405884375226

Manus is great for building fun web apps. In response to one prompt, Manus created a face comparison website, autonomously deploying deep learning algorithms to measure facial similarity between photos. https://x.com/ManusAI_HQ/status/1908187248920862771

I built an educational website without writing any code. 1. Asked @ManusAI_HQ to create me a beginners guide to MCP with diagrams in the hand drawn style of excalidraw. 2. Dropped the markdown file and images Manus created into @v0 3. Used @iamsahaj_xyz’s shadcn tool to https://x.com/nocheerleader/status/1921659613314748908

Just used @ManusAI_HQ to generate a fully functional model & animation of the ATP synthase (a key enzyme) in action: rotor, membrane, labeled components… Built in Blender by Manus, straight from my (natural language) prompts. Now imagine teachers… or students asking an AI https://x.com/emmanuel_2m/status/1904506775044510034

Manus delivers EDUCATION 2.0 🧑‍🏫 1/ FastAPI Course Creation Manus built a learning website. Interactive. Immersive. All your stuff in one sweet spot online. Hit RT + throw us your education challenge below. We’ll show you how Manus makes it real. https://x.com/ManusAI_HQ/status/1904195897979740288

Introducing Manus image generation. Manus doesn’t just generate images. It understands your intent, plans a solution, and knows how to effectively use image generation along with other tools to accomplish your task. https://x.com/ManusAI_HQ/status/1923048495310922028

@ManusAI_HQ I just made this using manus. How social media platforms suggest content using statistical behavioral analysis 9/10, bravo!! https://x.com/iwm33/status/1923053749653602494

Manus AI just partnered with Microsoft Azure AI Foundry What this means: – Global scale – Enterprise security – Same smooth Manus experience — now supercharged Follow for updates! https://x.com/Neuzenai23/status/1925758837715275845

1) first I asked to “”find tweets about promo saas videos posted april-may 2025″”. I need such tweets to market my SaaS promo video maker app (@paracast_io). I was expecting Manus to fail at the tweets fetching stage: it needs to have 𝕏 API keys, which I did not provide https://x.com/alexanderisorax/status/1923010686860284091

5 AI Visuals and Charts: Week Ending June 06, 2025

Microsoft literally dropped the best AI agent course in less than 60 minutes https://x.com/aaditsh/status/1927718775522304226

Lmao. What niche even is this — grassroots dirt track racing meets google maps nerds? Veo 3 videos are seriously ridiculous and fun. Turn audio on for max enjoyment. https://x.com/bilawalsidhu/status/1930405285253767296

I wrote a history of AI in 32 images of otters using wifi on airplanes, from images to video to code. It shows two big trends: rapid improvements in AI models of all types and the growth of open weights AI models. Link in the comments. https://x.com/emollick/status/1929306757903319089

I’ve been using prompts of otters as a test of AI ability. It has taken less than three years to go from a text prompt producing images of abstract masses of fur to producing realistic videos with sound (including “”like the musical Cats but for otters”). https://x.com/emollick/status/1929612980041253132

Two brutalist office buildings, floating in stormy seas and brimming with antique brass cannons, fire paint at each other, staining their surfaces. https://x.com/emollick/status/1929727288901550523