- NATURAL 20
- Mark Zuckerberg: "We're currently training LLaMA 3..." | Sam Altman talks GPT-5 and Training Data
Mark Zuckerberg: "We're currently training LLaMA 3..." | Sam Altman talks GPT-5 and Training Data
PLUS: OpenAI partners with ASU, Microsoft's free AI Reading Coach and more.
Mark Zuckerberg: "We're currently training LLaMA 3..." | Sam Altman talks GPT-5 and Training Data
Meta is merging two AI research efforts to achieve general intelligence, aiming to make it open-source and practical for everyday use. They're focusing on enhancing AI in various areas like reasoning, planning, and memory. Meta plans to use significant computational resources, including 350,000 Nvidia H100 GPUs, and is currently developing the llama 3 AI model.
Meanwhile, Sam Altman of OpenAI discussed potential future AI models, emphasizing the continuous improvement in generalized intelligence. He anticipates that AI will increasingly integrate into daily workflows, shifting how we interact with technology.
OpenAI partners with Arizona State University to use ChatGPT in classrooms
Arizona State University (ASU) teamed up with OpenAI to use ChatGPT in their classrooms. ASU's aiming to boost student success, research, and smoothen their operations using ChatGPT Enterprise. Kyle Bowen from ASU said they joined forces with OpenAI after ChatGPT Enterprise addressed their security worries. ASU's faculty, including AI experts, will guide ChatGPT's campus use.
Starting February, they're taking ideas from faculty and students on how to use ChatGPT. Anne Jones, from ASU, mentioned that some classes already use AI for writing and multimedia stories, and chatbots might even become personalized tutors. This partnership, a first for OpenAI with a school, is part of their move to work with public organizations, including a recent deal with Pennsylvania's government.
Microsoft makes its AI-powered reading tutor free
Microsoft just rolled out Reading Coach, their new, free AI tool that helps kids get better at reading. It's on the web now, with a Windows app coming soon. It's part of Microsoft's plan to make reading fun and personalized, using the latest tech.
Reading Coach lets kids work on words they find tough and gives teachers a peek at their progress. There's even a cool new feature where kids can create their own story using AI, picking characters and settings, and the AI fills in the rest.
BMW will deploy Figure’s humanoid robot at South Carolina plant
BMW is teaming up with Figure to bring humanoid robots to their South Carolina factory, the company's only U.S. plant and a major production hub. The plan is to start with five specific tasks, rolling them out one by one. These robots aren't all-purpose; they'll grow their abilities over time, like adding apps to a smartphone. Initial jobs are likely simple repetitive tasks, like moving boxes or loading pallets.
Figure aims to ship its first robot within a year. The robots, unique for their human-like hands, will learn on the job, combining different learning methods. BMW's commitment depends on the robots' performance. Figure is renting these robots to BMW, a strategy they plan to continue.
The Rabbit R1 will receive live info from Perplexity’s AI ‘answer engine’
Rabbit, the hot topic at CES 2024, is teaming up with Perplexity to beef up its latest gizmo, the R1. It's got a cool design by Teenage Engineering and has already racked up 50,000 preorders at $199 a pop.
It links to Perplexity's answer engine, giving users fresh, real-time info without hitting a knowledge wall. First 100K buyers also get a year of Perplexity Pro, usually 20 bucks a month, which includes access to the latest AI smarts like GPT-4.
The R1 is a handy little thing, with a touchscreen and camera, perfect for controlling apps and more. Perplexity's tech mashes up search and AI like GPT-3.5, stepping into the ring with big players like Google and Bing.
Self-Rewarding Language Models by Meta discusses a new approach in language model training. It suggests using the language model itself to generate its own rewards during training. This method, called "Self-Rewarding", aims to surpass human-level performance by avoiding limitations of human-based feedback. The study demonstrates improved instruction following and self-rewarding capabilities in models like Llama 2 70B, outperforming others in AlpacaEval 2.0. This is a preliminary exploration into models improving themselves continuously.
Rethinking FID: Towards a Better Evaluation Metric for Image Generation by Google critiques the popular Frechet Inception Distance (FID) metric used in image generation. It highlights FID's drawbacks, like poor representation of varied content, incorrect assumptions, and inconsistent results. The authors propose a new metric, CMMD, using CLIP embeddings and maximum mean discrepancy, offering a more robust and reliable assessment of image quality. This research suggests that current FID evaluations might be unreliable and that CMMD could be a better alternative.
DiffusionGPT: LLM-Driven Text-to-Image Generation System by ByteDance introduces DiffusionGPT, a novel system blending Large Language Models (LLM) with diffusion models for image generation. It efficiently handles diverse prompts and integrates expert models for superior performance. This system features domain-specific model selection guided by LLM-generated Trees-of-Thought and enriched with human feedback. The result is a more flexible, high-quality image synthesis across various domains.
Improving fine-grained understanding in image-text pre-training by Google Deepmind presents SPARC (SPARse Fine-grained Contrastive Alignment), a method for pre-training multimodal representations from image-text pairs. It focuses on learning fine-grained connections between image patches and language tokens, using a sparse similarity metric. This approach allows for a more detailed and computationally efficient learning process. SPARC enhances both coarse-grained tasks like image classification and fine-grained tasks such as retrieval and object detection, outperforming existing methods in both areas.
ChatQA: Building GPT-4 Level Conversational QA Models by NVIDIA introduces ChatQA, a family of conversational question answering models that achieve GPT-4 level accuracy. It outlines a two-stage instruction tuning method enhancing zero-shot conversational QA performance in large language models. A key feature is the fine-tuned dense retriever for multi-turn QA, offering comparable results to state-of-the-art models but at a reduced cost. ChatQA-70B notably outperforms GPT-4 in average scores across 10 conversational QA datasets.
The AI4Math/MathVista dataset is a collection of math problems designed for visual and textual question answering. It includes a range of math questions, from simple arithmetic to complex problems, accompanied by images and diagrams. The dataset is multilingual, covering English, Chinese, and Persian, and is aimed at various educational levels. The problems are categorized into multiple-choice, open-domain, and closed-domain questions, and the dataset is particularly useful for training and evaluating AI in solving math problems using both text and visual information.
The paper presents the Text-to-Video Score (T2VScore), a novel evaluation metric for text-to-video generation. Addressing the limitations of existing metrics, T2VScore assesses both the alignment between the video and the text description (Text-Video Alignment) and the overall production quality of the video (Video Quality). The evaluation is enhanced with a new dataset, TVGE, which includes human judgments of over 2,500 text-to-video generated clips. This approach offers a more comprehensive and accurate assessment of video generation models.
Spiritme - AI technology platform that specializes in creating AI avatars for video generation. It allows users to generate videos with customizable AI avatars, offering a unique and innovative approach to video creation.
Kindllm - distraction-free chat app optimized for Kindle, particularly Kindle Paperwhites. It's designed for readers who want to look up information without leaving their book.
Book & Bot - interactive storyworld where you can read a book and chat with the character.
CodeMonk - offers a comprehensive solution for optimizing tech teams. It includes talent sourcing, onboarding, payment, and team optimization.
Ubique - AI tool that personalizes video messages using voice and face cloning technology.
OpenAI must defend ChatGPT fabrications after failing to defeat libel suit
OpenAI's in a pickle with ChatGPT's made-up stories. A Georgia judge said "nope" to tossing out a radio guy's lawsuit against them. The radio host, big on gun rights, got falsely accused by ChatGPT of swindling money. OpenAI tried saying ChatGPT's blabber isn't legally harmful, but the court wasn't buying it. Now, they've got to prove their case, and everyone's watching to see if a chatbot's fibs can really mess someone up. ARSTECHNICA
AI Certification Program Verifies Systems Are ‘Fairly Trained’
A new certification program aims to make sure AI systems play fair. Think of it as an "organic" or "fair trade" sticker, but for artificial intelligence. This badge of honor shows that an AI has been checked out and follows the rules, especially about not stealing other people's creative work. Ed Newton-Rex, who used to work at Stability AI, started this nonprofit called Fairly Trained to do the checking. He left his old job worried about AI taking advantage of creative folks. BLOOMBERG
Accenture to Open 10 AI Hubs Globally, Including One in London
Accenture's going big on AI, planning to open 10 new innovation hubs across the globe, with one in London. CEO Julie Sweet spilled the beans at the World Economic Forum in Davos, Switzerland. She's pumped about the UK market, hinting that Accenture's been on a bit of a shopping spree, buying up companies lately. This move is a major part of their AI strategy, showing they're serious about staying at the forefront of generative AI tech. BLOOMBERG
WHO releases AI ethics and governance guidance for large multi-modal models
The World Health Organization (WHO) dropped new rules on using big AI systems in health care. These AI systems can handle different data types like text and videos and are super fast at learning. The WHO's got 40 tips for governments, tech firms, and health pros to use these AIs right, aiming to make health care better and fairer. They're great for diagnosing, helping patients, doing paperwork, training medical staff, and researching drugs. But, there are risks like wrong info, bias, and security problems. The WHO says everyone involved, from governments to patients, should have a say in making and checking these AIs to keep them safe and effective. WHO
Farm-ng makes modular robots for a broad range of agricultural work
Farm-ng is shaking up farming with modular robots, making farm work less of a grind. These bots, called Amiga, can do different jobs like planting, weeding, and spreading compost. They're a hit in big California farms. The idea is to give farmers LEGO-like tools to build what they need without breaking the bank. Farm-ng, starting during the pandemic, just bagged $10 million in funding. They've already put out about 100 Amigas and are using the cash to make more in their Watsonville factory. Farmers using Amigas are saving big on time and money, cutting labor hours by half to 80%. While it's early days, the feedback's super positive. TECHCRUNCH
TikTok can generate AI songs, but it probably shouldn’t
TikTok's dabbling in AI music creation, letting users write songs with AI assistance. Users type lyrics, and the AI, powered by Bloom, adds music in various genres. It's still in test mode, with mixed results - some tunes are off-key, even with auto-tune. Despite this, AI music's not new on TikTok; there have been viral hits like the AI-generated Drake and The Weeknd song. YouTube's also trying something similar, letting users make songs from text or humming. TikTok's now being more open about AI in content, making it clear when songs are AI-made. THE VERGE
Sundar Pichai is taking a leaf out of Mark Zuckerberg’s book and warns Google staff ‘ambitious goals’ can only be met with job cuts
Google's CEO, Sundar Pichai, is signaling job cuts to focus on AI investments. This move mirrors what happened at Meta last year. Pichai's memo, titled ‘2024 priorities and the year ahead’, hints at tough choices and job cuts for some teams to simplify and speed up work. Although he reassures not all teams will be affected and the scale won't be as large as last year's layoffs. In 2023, Google chopped 12,000 jobs, a reaction to over-hiring during COVID. Google's been racing to keep up in the AI field, especially after the launch of ChatGPT by rival Microsoft-backed OpenAI. They had a rocky start with their AI bot, Bard, causing a huge stock drop due to misinformation. Yet, Google's stock soared 56% last year, showing resilience in tech. Like Meta's 2022 cuts for AI focus, Google's trimming the fat for a leaner, more AI-centric future. Big Tech's feeling the heat, with Amazon and Salesforce also slashing jobs recently. FORTUNE
Microsoft to embed AI tool for clinical documentation in Epic’s health records software
Microsoft is leveling up health care with its AI tool, DAX Copilot, by teaming up with Epic, a big name in health records software. This tool's all about making docs' lives easier by drafting notes on the go in Epic's mobile app, Haiku. It's a big deal 'cause note-taking is a major headache in healthcare. DAX Copilot's already in hundreds of hospitals, but linking with Epic means it could help way more clinicians. Epic's huge in the U.S., known for its top-notch patient record software. This move could really change the game in how patient info is handled. STAT
AI could help the water industry curb its thirst for energy
Gaurav Gupta used AI to make a Coca-Cola plant in Abu Dhabi more efficient. Partnering with Canadian startup Pani Energy, they improved energy and water usage by 3% using AI software, saving costs and reducing damage risks. Pani's goal is to slash the water industry's greenhouse gas emissions significantly by 2032, focusing on energy efficiency and smarter water treatment. The approach is gaining traction, saving money, and contributing to a greener future. FORTUNE
AI PCs could make 16GB of memory the new normal
AI advancements are pushing 16GB of RAM to become the new standard for PCs. Microsoft suggests this for AI PCs, especially for features like Copilot in Windows. While Copilot works on lower specs, relying on remote servers, future AI tasks may need more local power. Intel and AMD are adding AI-focused hardware to their chips. Microsoft's move might push manufacturers to adopt 16GB RAM in new devices, making it a key upgrade for better performance, although upgrading RAM in laptops can be tricky due to soldered components. PCWORLD
OpenAI CEO Sam Altman says ChatGPT doesn’t need New York Times data amid lawsuit
Sam Altman, OpenAI's CEO, expressed surprise at The New York Times' lawsuit against OpenAI for allegedly using their content to train AI models. Speaking at the World Economic Forum in Davos, he clarified that while OpenAI was open to using the Times' content, it wasn't a priority and their AI doesn't rely on any single source. The Times accused OpenAI and Microsoft of copyright infringement, seeking substantial damages. OpenAI contends that the use of such content is rare and unintended. Altman emphasized OpenAI's commitment to reducing content regurgitation and exploring new ways to use and monetize news content, while balancing the challenges of avoiding direct copying. CNBC
Google DeepMind cofounder says AI can act like an entrepreneur and inventor in the next five years
Mustafa Suleyman, cofounder of DeepMind, predicts that AI will have the capacity to run a business like an entrepreneur or inventor within the next five years. Speaking at the World Economic Forum in 2024, he suggested that the future Turing test for AI could involve evaluating its ability to manage a project, market, manufacture, and sell a product profitably. Suleyman believes these capabilities will soon be widely available and affordable, potentially even open source, drastically changing the economy. He emphasizes the importance of focusing on AI's practical abilities, like interaction and planning, rather than getting caught up in the ambiguous concept of "intelligence." Suleyman's views highlight a shift towards AI as a labor-replacing tool that enhances productivity and creativity. BUSINESS INSIDER
Google News Is Boosting Garbage AI-Generated Articles
Google News has been found to promote AI-generated articles that often mimic or plagiarize content from legitimate news sources. This issue arises from the opaque and potentially exploitable ranking system of Google News. Google has stated that they do not differentiate between AI-written and human-written articles in their News service. This situation highlights a challenge in moderating content in an era where AI tools are easily accessible for mass content creation, often compromising the quality and originality of news. An example is Worldtimetodays.com, which published articles closely resembling those on Distractify and Heavy.com, including using the same author photo and failing to remove watermarked artwork. The site's author, Gary Graves, reportedly published over 40 articles in a single day, raising questions about the authenticity and integrity of the content. This development is a concern for readers like Brian Penny, who value original journalism involving real research and primary sources. 404 MEDIA
So you want to work in AI but don't know where to start? Follow these 5 steps
Looking to jump into the AI game but feel lost? This article is a guide for those looking to pivot into AI, inspired by a reader, Rick, who's transitioning from life sciences to tech.
Five Steps to Enter AI:
Identify Current Skills: Recognize existing skills that are useful in AI (e.g., leadership, writing).
Recognize Transferable Skills: Understand how current skills apply to AI, like prompt engineering for product managers.
Educate Yourself in AI: Learn about the AI industry, players, and technologies through reading and free courses.
Build AI Resume Points: Create AI-related work (e.g., blogs, project work) to demonstrate knowledge and interest.
Commit Time to Transition: Dedicate at least six months to learning and integrating AI into your skillset.
What'd you think of today's edition?
What are MOST interested in learning about AI?
What stories or resources will be most interesting for you to hear about?