About This Week’s Covers
This week’s cover combines a few of the top stories: First, X unveiled the world’s most powerful AI training cluster, second, 1X released a humanoid robot, third, music producer Timbaland was blown away by AI’s potential to finish his unreleased tracks, and fourth, Apple announced its new iPhone with AI features. The cover shows a robot lounging on a couch at a data center, checking its phone and listening to music while Timbaland’s viral video plays on a TV. The image was created using Flux Dev and Photoshop. Prompt was “Medium shot. A humanoid robot relaxes on a couch. He’s in the middle of a sprawling sterile data center. He’s alone. The couch is an 80s couch with oranges and browns. A radio is playing on the floor. The robot is laying on the couch with its feet up. It’s checking an iphone. A sign reads “AI NEWS 49: 2024/09/06”.
The category covers were all created with Ideogram using “one-shot” prompting (no second tries or post-production). Each cover reflects a theme from the week’s links. Here are a few of my favorites:
Executive Summaries
X AI Unveils Colossus, The World’s Most Powerful AI Training Cluster
Elon Musk’s xAI team has launched “Colossus,” a powerful AI training system featuring 100,000 Nvidia H100 GPUs. Built in just 122 days, Colossus is now the most powerful AI training infrastructure in the world, and plans are underway to double its capacity with 200,000 GPUs in the coming months. This rapid progress has sparked concern among competitors like OpenAI, with CEO Sam Altman reportedly worried that xAI’s growing computational power could soon surpass OpenAI’s resources.
Elonmusk | xDaily
OpenAI Claims Upcoming GPT Will Be 100x More Powerful Than GPT-4
At the KDDI Summit 2024, OpenAI Japan’s CEO announced that the upcoming GPT-Next will be 100 times more powerful than GPT-4. The dramatic performance leap is attributed to advancements in model architecture rather than increased computing resources. GPT-Next will incorporate a smaller version of OpenAI’s secretive Project Strawberry.
rowancheung
Ilya Sutskever’s Safe Superintelligence Raises $1 Billion to “Build Safer AI”
Safe Superintelligence (SSI), a new AI startup co-founded by Ilya Sutskever, former chief scientist at OpenAI, has raised $1 billion from top investors, including Andreessen Horowitz, Sequoia Capital, and NFDG. Valued at $5 billion, SSI aims to develop AI systems that surpass human capabilities while prioritizing safety. The company, founded in June 2024, plans to use the funds to acquire computing power and hire a highly vetted team of researchers split between Palo Alto and Tel Aviv. SSI’s mission is to ensure AI safety through rigorous R&D before bringing its products to market, focusing on creating superintelligence that aligns with human values. Sutskever’s departure from OpenAI, following disagreements over leadership, set the stage for this new venture. It’s worth noting that Sutskever is one of the most beloved AI pioneers, and he’s 100% worth paying attention to no matter what he’s doing.
Twitter | reuters | wsj
1X Technologies Unveils New Humanoid AI Robot for Home Use
In two new videos, 1X Technologies introduces its latest humanoid robot, NEO Beta, designed to function in homes. 1X plans to combine physical presence with cognitive capabilities, performing tasks like caregiving, household chores, and companionship. They are competing most directly with Figure and Optimus.
1x.tech | youtube | youtube
Timbaland Blown Away by AI Music Tool Suno
Grammy-winning producer Timbaland recently shared his excitement over using Suno, an AI-powered music tool, to add vocals to an unfinished song from 2021. In a video, he marvels at how accurately the AI matched the song’s cadence, spaces, and musicality, finishing tracks he had shelved years ago. Timbaland sees huge potential in AI helping producers tap into their archives of unfinished projects, adding melodies and ideas to works that might otherwise go unused. Unlike many artists who view AI with skepticism or fear, Timbaland embraces it as a creative partner, calling it “next level” and a glimpse into the future of music production.
Emollick | Timbaland
Replit’s AI Coding Agent Is Shocking People By Building Prompt-based Apps
Replit’s new AI Agent is making waves by drastically lowering the barrier to building applications. Users with little to no coding experience have reported building fully functional web apps, games, and more, in just minutes. The AI handles everything—from choosing the tech stack to writing, debugging, and deploying code. Users have shared impressive results, like creating apps on their phones in under five minutes or building complex systems like sentiment analyzers or A/B test generators without typing a single line of code. Replit’s official overview highlights how the agent takes natural language prompts and turns them into working software, offering an accessible entry point for all skill levels. The tool excels in web app prototyping, where it handles development and deployment, transforming ideas into working products almost instantly. Andrej Karpathy (an AI pioneer) writes, “Very cool, place well under “feel the AGI” category. As mentioned in the post, making actual apps is a lot more than code, you have to set up the entire environment, deploy it, etc. Automating all of this other infra will allow anyone to quickly build and deploy entire web apps.” Here are a few examples of what people have been doing:
Emollick | amasad | sandersaar | DrLancaster | frankismartinez | martinbowling | victorpaycro | karpathy | replit
Apple Unveils AI-Driven iPhone 16 Amidst Slowing Sales
Apple has introduced the iPhone 16, headlined by a significant focus on artificial intelligence with its new “Apple Intelligence” system. The upgraded Siri, enhanced AI text-editing, and advanced photo and video recognition tools aim to rejuvenate sales, which have been declining as consumers hold onto their phones longer. The AI features include sorting messages, providing writing suggestions, and offering a smarter Siri that can handle personal requests securely on-device. Apple’s push into AI comes after years of watching competitors like Google and Microsoft, but the company hopes its privacy-first approach and deeper integration will set it apart. Despite these upgrades, Wall Street remains cautious, with questions about AI’s viability and whether these features will drive significant iPhone upgrades. Analysts predict a potential boost, with forecasts of up to 240 million units sold in the coming year.
Ycombinator | nytimes | wsj | youtube
Reflection 70B: Open-Source AI Model Sparks Excitement and Skepticism
Reflection 70B, developed by Matt Shumer and HyperWrite, has been unveiled as the world’s most powerful open-source AI model, claiming to outperform top-tier closed-source models like GPT-4o and Claude Opus on multiple benchmarks. Powered by “Reflection-Tuning,” a novel technique enabling the model to detect and correct its own errors, Reflection 70B has shown impressive results in benchmark tests like MMLU and GSM8K. However, critics question its performance, citing concerns about overfitting, inconsistencies on other benchmarks, and exaggerated claims of accuracy. If proven true, Reflection 70B represents a major leap in open-source AI, with even larger models like Reflection 405B expected soon. Updates to follow in the coming weeks as the model is peer-reviewed.
MatthewBerman | reddit | JJitsev | hughbzhang | mattshumer_ | venturebeat | gazorp5 | paulgauthier | reddit | terryyuezhuo | AIandDesign | gauravpathak
AI Companies Are Moving Away from User Data for Training
Recent discussions suggest that user data, including large amounts of content from amateur creators, is becoming increasingly irrelevant for training AI models. The scale and quality of such data no longer make a significant impact, which is why companies like Anthropic have shifted away from relying on it. The idea that AI companies are desperate to absorb every piece of text is also overstated. With the focus moving toward higher-quality, purpose-built datasets, the perceived value of random user data for AI training is being questioned.
Emollick | simonw
Top 53 Links of The Week – Organized by Category
Agents and Copilots
“I used Cursor AI to build a Web Scrapping AI Agent with RAG from scratch in less than 2 minutes. I built this Python app just by using English prompts. NoCode.
“Altera’s Project Sid made 1,000 fully autonomous agents play Minecraft for days without human intervention. The AI agents ended up building economies, cultures, governments, and religions. It’s AI agents all the way down.
“Benchmarking >80 LLMs shows: The best model is not necessarily the best for your programming language 😱 – Best overall: Anthropic’s Sonnet 3.5 – Best for Go: Meta’s Llama 3.1 405B – Best for Java: OpenAI’s GPT-4 Turbo – Best for Ruby: OpenAI’s GPT-4o Good models for one
“AI Agents Are Slowly Coming Alive Autonomous AI agents will soon perform most tasks. Our AI agents can automate various activities such as document generation, on-demand dashboards, report generation, and technical image generation. Here’s how AI Agents in Abacus AI works:
Salesforce to acquire AI voice agent firm Tenyx, joining AI talent race
“xLAM: A Family of Large Action Models to Empower AI Agent Systems abs:
Exclusive: Marc Benioff has declared a ‘hard pivot’ to autonomous AI agents. Will it be enough for Salesforce to thrive in the generative AI era? | Fortune
Anthropic
“Announcing the Anthropic Quickstarts repo, a collection of projects designed to help developers quickly get started with building deployable applications using the Anthropic API. Featuring our first quickstart – a Claude-powered customer support agent app:
“Anthropic announced the full release of Artifacts for all Claude users, including on mobile Users can now create Artifacts directly on their phones, and I’ve seen a ton of cool use cases already
“Google can’t compete with this. After a few hours (yes, hours!) trying to figure out this problem using Google, I decided to ask Claude. 5 minutes later, I had step-by-step instructions on what to do. I love 2024 so much.
Anthropic launches Claude Enterprise plan to compete with OpenAI | TechCrunch
Apple
Apple Pay at 10: a tap-to-pay revolution that is about to be upended – The Verge
Apple is thinking about a rival to Meta Ray-Ban glasses
Augmented and Virtual Reality (AR/VR)
“This AI paper is pretty mind-blowing: OmniRe reconstructs 3D urban scenes from driving data, handling both static and dynamic elements like vehicles, pedestrians, and cyclists. And of course the best way to illustrate that understanding is to reskin all the pedestrians 🙂 Why
Autonomous Vehicles
Tesla Releases ‘Actually Smart Summon’: Features & Videos [Now Includes HW3]
Chips, Hardware, and Infrastructure
“🚀@LumenOrbit (YC S24) is building a network of megawatt-scale data centers in space, scalable to gigawatt capacity. Congrats on the launch, @johnstonphil, @ezrafeilden, and @adioltean!
“BREAKING: Nvidia has been subpoenaed by the US Justice Department in an escalation of the agency’s antitrust investigation
Ethics/Legal/Security
AI-Generated Influencer Ads | Icon
“🚨🇺🇸HUGE LOSS FOR PUBLIC LIBRARIES: COURT RULES AGAINST INTERNET ARCHIVE | 500K BOOKS REMOVED A court has ruled that the Internet Archive, a key resource used by public libraries to lend books digitally, can no longer lend over 500,000 books like “The Catcher in the Rye” and “To” / X
Self-learning AI reveals NFL against the spread, over-under, money-line picks for each 2024 Week 1 game – CBSSports.com
Venezuela’s Newest News Agency Says AI Anchors Protect Reporters Amid Government Crackdown
Police officers are starting to use AI chatbots to write crime reports. Will they hold up in court?
The US Needs Deepfake Porn Laws. These States Are Leading the Way | WIRED
YouTube is making tools to detect face and voice deepfakes
“When it comes to the internet—telecos + big tech own all the pipes, but Starlink owns all of space 🛰️
YouTube is making new tools to protect creators from AI copycats – The Verge
NO FAKES – A Dream for Lawyers, a Nightmare for Everyone Else | Electronic Frontier Foundation
Republican voters less confident than Democrats in spotting AI deepfakes, poll shows | US politics | The Guardian
AI-Fakes Detection Is Failing Voters in the Global South | WIRED
A.I. Can Now Create Lifelike Videos. Can You Tell What’s Real? – The New York Times
North Carolina Man Charged With Using AI to Win Music Royalties – The New York Times
Philippines’ Call Centers Navigate AI Impact on Jobs – Bloomberg
Google Photos: Search improvements and early access to Ask Photos
Google’s AI-powered Ask Photos feature begins US rollout | TechCrunch
Locally Run
“@emollick I like telling those people that I have a USB stick with Llama 3.1 70B – a very capable model – that runs directly on my laptop That’s not going away You can grab a copy from here, packaged with the software needed to run it on any modern OS
Multimodality
“Those are some crispy and consistent depth maps! DepthCrafter looks like the new SOTA for video depth estimation tasks. Take any video, of any length, and get temporally coherent depth maps that you can use for VFX, or as an input to other AI models. Quick thread
OpenAI
OpenAI’s First In-House Chip Will Be Developed By TSMC On Its A16 Angstrom Process For Its Sora Video Applications, Move Will Also Boost Apple Device Sales
“OpenAI Japan CEO revealed that GPT-Next will be 100x more powerful than GPT-4 at KDDI Summit 2024 The model will reportedly use a smaller version of OpenAI’s mysterious Project Strawberry The boost in performance comes from better architecture, not a ton of computing resources
OpenAI plans to build its own AI chips on TSMC’s forthcoming 1.6 nm A16 process node
Podcasts/YouTube/Op-Eds
“11/ @sema4ai’s co-founder @aikarjal had a great podcast episode where he talked about how AI Agents are changing how enterprises operate.
What Is ‘Agentic’ AI? It Doesn’t Require a Human to Tell It What to Do – The New York Times
Michael Dell on AI: Talk of a slowdown is way overblown
Bill Gates has a good feeling about AI: a Q&A – The Verge
Publishing
Venezuela’s Newest News Agency Says AI Anchors Protect Reporters Amid Government Crackdown
Google AI Overviews rollout ramps up, hitting publisher visibility
(RAG) Retrieval-Augmented Generation
“Just coded a RAG pipeline in <1hr using only @cursor_ai composer, optimized via Hyde (hypothetical document embeddings) and @cohere reranker, without writing a single line of code. To make it more fun, I did the whole thing while just dictating to my computer with my voice, I
Robotics and Embodiment
“Just got the updated @Figure_robot tour from @adcock_brett. One word: INSANE
Science and Medicine
Alex Plays Counter-Strike – YouTube
“We’re presenting AlphaProteo: an AI system for designing novel proteins that bind more successfully to target molecules. 🧬 It could help scientists better understand how biological systems function, save time in research, advance drug design and more. 🧵
“Google DeepMind unveiled AlphaProteo, an AI system that designs custom proteins to bind with specific molecular targets This could potentially speed up drug discovery and cancer research, among many other things
AlphaProteo generates novel proteins for biology and health research – Google DeepMind
Video
“Lmao. FLUX LoRAs + Kling AI is a ton of fun. You can easily put yourselves into all sorts of cinematic scenarios. Prompt: “k**l it with fire” – inspired by Alien: Romulus
“The winners of the first-ever Reply AI Film Festival were announced for best original short films made with AI. Gisele Tong’s To Dear Me came in first place and tells a story about a woman tormented by the memories of her parents’ divorce. This is incredibly cool:
“Lego as your daily meal 🎥 hypnomotion
Twitter/X/Grok
“When it comes to the internet—telecos + big tech own all the pipes, but Starlink owns all of space 🛰️ https://twitter.com/bilawalsidhu/status/1830393388962631694
Leave a Reply