Note to our subscribers: Last Week in AI did no go out last week and is coming late this week due to some personal matters. You can expect the newsletter and podcast to resume their normal schedule next week – apologies for the interruption and thank you for being a subscriber.
Top News
Apple launches iPhone 16 with Apple Intelligence
Apple has unveiled its iPhone 16 line, which includes the iPhone 16, iPhone 16 Plus, iPhone 16 Pro, and iPhone 16 Pro Max, all designed with the Apple Intelligence mind. However, Apple Intelligence will not begin rolling out next month with updates to iOS 18.1, iPadOS 18.1, and macOS Sequoia 15.1, and “with more features launching in the coming months.†The key features it includes are:
Systemwide Writing Tools that allow users to rewrite, proofread, and summarize text in apps like Mail, Notes, Pages, and third-party platforms
In Photos, users can search for specific images using natural language, remove unwanted objects from pictures with the Clean Up tool, and create personalized movies by simply typing a description.
In apps like Notes and Phone, users can record, transcribe, and summarize audio, helping them capture and recall important information from conversations and calls.
All of this is implemented with an emphasis on user privacy. Apple Intelligence will debut in U.S. English, expanding to other English dialects and languages over time.
Apple Intelligence Arrives in October: Here’s What You Get First
Apple Intelligence comes to iPhone, iPad, and Mac starting next month
AI-powered visual search comes to the iPhone
California Legislature Approves Bill Proposing Sweeping A.I. Restrictions
The controversial California bill SB 1047, aimed at preventing AI disasters, has passed the state’s Senate and is now awaiting Governor Gavin Newsom’s decision. The bill, introduced by state senator Scott Wiener, targets large AI models that could potentially cause catastrophic events, such as loss of life or cyberattacks costing over $500 million. It proposes making AI developers liable for any harm caused by their models, similar to gun manufacturers being held accountable for mass shootings. The bill also grants California’s attorney general the power to sue AI companies for hefty penalties if their technology is used in a catastrophic event. While some industry figures, including Elon Musk, are cautiously optimistic about the bill, others fear it could stifle innovation in California’s booming AI industry. Governor Newsom has until September 30 to sign or veto the bill.
OpenAI plans to build its own AI chips on TSMC’s forthcoming 1.6 nm A16 process node
OpenAI is reportedly planning to build its own AI chips using TSMC’s forthcoming 1.6nm A16 process node, according to United Daily News. This move implies OpenAI may be interested in manufacturing their own computing chips, which would be a big change of strategy for them and a means to reduce the high costs associated with running ChatGPT on Nvidia’s AI servers. The A16 process node, which is still in development, will be the first TSMC node to use backside power delivery, known as Super Power Rail. However, the future of this project is uncertain.
Ilya Sutskever’s startup, Safe Superintelligence, raises $1B
Safe Superintelligence (SSI), an AI startup co-founded by Ilya Sutskever, former chief scientist at OpenAI, has successfully raised over $1 billion in funding. Key investors include NFDG, a16z, Sequoia, DST Global, and SV Angel, bringing SSI’s valuation to an estimated $5 billion. The funds will be utilized to acquire computing power and expand the team of researchers and engineers based in Palo Alto and Tel Aviv. Sutskever, who previously led the Superalignment team at OpenAI, left the organization following a publicized disagreement with several board members and the CEO, Sam Altman.
Other News
Tools
Adobe says video generation is coming to Firefly this year – Adobe’s AI model for video generation, Firefly, will introduce features like Generative Extend, Text to Video, and Image to Video, aiming to complement or accelerate existing workflows, while also implementing safeguards to restrict certain content.
DeepSeek-V2.5 wins praise as the new, true open source AI model leader – DeepSeek-V2.5, an open-source AI model, is praised for its enhanced language processing and coding capabilities, outperforming its predecessors and other models in various benchmarks, and is made accessible for commercial usage.
Reflection is a new AI model that can run on a good laptop and still beat GPT-4o in tests – Reflection 70B, an open-source AI model developed by HyperWrite, introduces a novel “reflection” mechanism to enhance reasoning capabilities and accuracy, outperforming GPT-4o and demonstrating exceptional performance across various benchmarks.
Yi-Coder Released by 01.AI : A Powerful Small-Scale Code LLM Series, Delivering Exceptional Performance in Code Generation, Editing, and Long-Context Comprehension – Yi-Coder by 01.AI is a powerful small-scale code LLM series that delivers exceptional performance in code generation, editing, and long-context comprehension, surpassing larger models in competitive programming and standard code generation benchmarks.
OLMoE: Open Mixture-of-Experts Language Models – OLMoE is a fully open, state-of-the-art language model with 7 billion parameters, pre-trained on 5 trillion tokens, and outperforming other models with similar active parameters.
Open-MAGVIT2: An Open-Source Project Toward Democratizing Auto-regressive Visual Generation – Open-MAGVIT2 is an open-source project that aims to democratize auto-regressive visual generation by providing replication of Google’s MAGVIT-v2 tokenizer and exploring its application in plain auto-regressive models.
Tell Replit’s AI Agent Your App Idea, and It’ll Code and Deploy It for You – Replit has launched an AI agent capable of building entire applications from scratch, operating with a higher degree of independence and making decisions and executing complex tasks without constant user input.
Google tests its ‘Ask Photos’ AI assistant that understands what’s in your pictures – Google is testing a new “Ask Photos” feature that uses AI to understand the content of your pictures and allows you to explore your photo library in new ways.
Roblox Builds Open-Source 3D AI Model, Adds Tech for Faster Game Loading – Roblox is developing an open-source generative AI tool, the 3D Foundational Model, which can understand various prompts and aims to speed up game development, while also making technical improvements for faster game loading.
Business
You.com ‘refocuses’ from AI search to deeper productivity agents with new $50M round – You.com is shifting its focus from AI search to deeper productivity agents, aiming to excel in answering complex questions and providing a productivity engine for knowledge work.
Musk’s xAI Supercomputer Goes Online With 100,000 Nvidia GPUs – Elon Musk assembles a new supercomputer running 100,000 Nvidia GPUs in just four months, aiming to accelerate training for AI projects and unlock new capabilities.
Amazon-backed Anthropic rolls out Claude AI for big business – Anthropic, an Amazon-backed AI startup, has launched Claude Enterprise, a new product designed for businesses looking to integrate Anthropic’s artificial intelligence, with features such as larger context windows, activity feeds, and company privacy.
Canva says its AI features are worth the 300 percent price increase – Canva’s aggressive rollout of generative AI features has led to a 300 percent price increase for some subscriptions, causing backlash from users.
Oracle shares jump as AI push perks up cloud demand – Oracle’s shares surged over 10% due to its AI push in cloud services, leading to a 21% rise in cloud product revenue and a narrowing of the gap with market leaders.
Intel Just Scored a Big AI Chip Customer – Intel’s Gaudi 3 AI chips are gaining traction as IBM collaborates to integrate them into its cloud data centers, marking a significant win for Intel in the AI accelerator market.
OpenAI Hits 1 Million Paid Users For Business Versions of ChatGPT – OpenAI’s corporate versions of ChatGPT have surpassed 1 million paid users, indicating a growing demand from businesses for its chatbot despite competition.
BlackRock-backed SambaNova launches ‘world’s fastest AI inference’ service – SambaNova Systems launches the world’s fastest AI inference service, enabling developers to run AI models at unrivaled speeds with low latency, exceeding inference speeds of systems run by vendors like OpenAI, Anthropic, and Google.
Glean Raises $260 Million to Expand Work AI Platform – Glean raises $260 million in Series E funding to accelerate its AI innovation and global expansion of its Work AI platform for enterprises, doubling its valuation to $4.6 billion.
Japan’s Sakana AI partners Nvidia for research, raises $100M – Japan’s Sakana AI partners with Nvidia to develop AI community, raises $100M for talent and infrastructure development.
Audible to Start Generating AI Voice Replicas of Select Audiobook Narrators – Audible will use AI to replicate select audiobook narrators’ voices, aiming to quickly and affordably expand its audiobook offerings and integrate traditional narrators into the evolving world of audiobook automation.
Research
Google DeepMind Launches AlphaProteo , an AI Model for Generating Proteins – Google DeepMind has launched AlphaProteo, an AI system that generates novel proteins to accelerate research in drug design, disease understanding, and health applications, potentially leading to breakthroughs in treatments and diagnostics.
Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers – LLMs are capable of generating more novel research ideas than human experts, but are slightly weaker on feasibility, prompting the need for further study on their capabilities.
Fire-Flyer AI-HPC: A Cost-Effective Software-Hardware Co-Design for Deep Learning – A cost-effective software-hardware co-design for deep learning, the Fire-Flyer AI-HPC architecture reduces construction costs by half and energy consumption by 40% while achieving performance similar to the DGX-A100.
Hermes: Memory-Efficient Pipeline Inference for Large Models on Edge Devices – Efficient pipeline inference for large AI models on edge devices is achieved through the Hermes system, which optimizes memory usage.
MMEvol: Empowering Multimodal Large Language Models with Evol-Instruct – MMEvol is a novel multimodal instruction data evolution framework that enhances Multimodal Large Language Models with complex and diverse image-text instruction datasets, resulting in improved performance on vision-language tasks.
Improving Pretraining Data Using Perplexity Correlations – A framework for selecting high-quality pretraining data based on LLM losses and downstream benchmark performance correlations outperforms existing methods in controlled experiments.
Robot Utility Models: General Policies for Zero-Shot Deployment in New Environments – Robot Utility Models (RUMs) enable zero-shot deployment of robot policies in new environments, achieving a 90% success rate in unseen scenarios without the need for further data, training, or fine-tuning.
VFA: Vision Frequency Analysis of Foundation Models and Human – The article discusses the VFA (Vision Frequency Analysis) of foundation models and human, emphasizing the importance of openness, community, excellence, and user data privacy.
VILA-U: a Unified Foundation Model Integrating Visual Understanding and Generation – A unified foundation model, VILA-U, integrates visual understanding and generation, but the specific details of the article are not available.
Concerns
Music Producer Arrested Over Using AI And Bots To Boost His Streams On Platforms – Music producer arrested for using AI and bots to boost streams and generate AI music, facing charges of money laundering and wire fraud.
Taylor Swift says AI version of herself falsely endorsing Trump ‘conjured up my fears’ – Taylor Swift expresses concern about an AI version of herself falsely endorsing Trump, and officially endorses Kamala Harris and Tim Walz for the 2024 Presidential Election.
Policy
US, EU and UK sign world’s first international AI treaty – US, EU, and UK sign the world’s first international AI treaty, emphasizing human rights and democratic values as key to regulating public and private-sector AI models, with the treaty holding each signatory accountable for any harm or discrimination caused by AI systems.
X permanently stops Grok AI from using EU citizens’ tweets after court action by Irish data watchdog – Grok AI is permanently prohibited from using EU citizens’ tweets after the Irish Data Protection Commissioner took legal action, leading to the issue being referred to the European Data Protection Board for further adjudication.
Analysis
AI Image Statistics : How Much Content Was Created by AI – AI has led to the creation of over 15 billion images since 2022, with an average of 34 million images being created per day, surpassing the pace and scale of photography development.
Human drivers are to blame for most serious Waymo collisions – Waymo’s driverless cars have been involved in fewer injury-causing crashes per million miles of driving than human-driven vehicles, with most serious crashes being caused by human drivers.
Explainers
A.I. Can Now Create Lifelike Videos. Can You Tell What’s Real? – A.I. technology can now create lifelike videos, but it still has limitations in accurately depicting certain nuances and details.
Source: Read MoreÂ
