Top News
The AI drama is heating up
The AI industry is seeing increasing tension, highlighted by a recent clash between Elon Musk and Yann LeCun on social media. LeCun criticized Musk’s leadership at xAI, calling him an erratic megalomaniac, following Musk’s announcement of a $6 billion funding round for xAI. This exchange reflects the fierce competition for top AI talent, with Meta feeling insecure about losing researchers and both Meta and xAI vying for partnerships with startups like Character.ai. The high stakes in the AI race are creating a dramatic environment, driven by large sums of money and the limited pool of influential AI researchers.
OpenAI is also embroiled in controversy, with former board member Helen Toner accusing CEO Sam Altman of dishonesty and manipulation during a failed coup attempt. Toner alleges that Altman withheld important information from the board and mishandled safety processes. Her claims, along with those of fellow ex-board member Tasha McCauley, prompted a rebuttal from current board members who defended Altman and stated that an independent investigation cleared him. As former employees begin to speak out, it appears the internal conflicts at OpenAI are far from over.
Ex-OpenAI board member reveals what led to Sam Altman’s brief ousting
AI firms mustn’t govern themselves, say ex-members of OpenAI’s board
Vox Media and OpenAI Form Strategic Content and Product Partnership , PwC agrees deal to become OpenAI’s first reseller and largest enterprise user
Vox Media has announced a strategic partnership with OpenAI, aiming to leverage AI technology to enhance its content and product offerings. The collaboration will see Vox Media’s portfolio, including Vox, The Verge, Eater, and New York Magazine, contribute to the information provided to ChatGPT’s 100 million users. OpenAI’s technology will be used to develop new products for Vox Media’s consumers and advertising partners, and to enhance Vox Media’s affiliate commerce product, The Strategist Gift Scout. The partnership will also extend the capabilities of Forte, Vox Media’s first-party data platform, improving creative optimization and audience segment targeting for advertisers.
OpenAI has also partnered with consulting giant PwC to provide ChatGPT Enterprise, the business-oriented version of its AI chatbot, to PwC employees and clients. This collaboration will grant access to OpenAI’s latest tools, including the ChatGPT-4o model and new voice and image capabilities. Notably, this is the first time OpenAI has adopted a resale model for its AI products, as the company seeks new revenue streams to offset its losses.
Journalists “deeply troubled†by OpenAI’s content deals with Vox, The Atlantic
Google Rolls Back A.I. Search Feature After Flubs and Flaws
Google’s new artificial intelligence feature for its search engine, A.I. Overviews, has been significantly rolled back after it produced a series of errors and false information. Introduced by CEO Sundar Pichai, the feature was designed to generate comprehensive information summaries above traditional search results. However, tests by The New York Times revealed that only one of the six queries initially demonstrated still yielded an A.I. Overview. The rollback appears to be a response to the feature’s numerous mistakes, including recommending glue in a pizza recipe and suggesting the ingestion of rocks for nutrients, which led to widespread user complaints and mockery on social media.
Google curbs AI search tool after it told people to eat rocks
Google CEO Sundar Pichai on AI-powered search and the future of the web
Google’s A.I. Search Leaves Publishers Scrambling
Other News
Tools
Opera is adding Google’s Gemini AI to its browser – Opera integrates Google’s Gemini AI models into its Aria AI browser assistant, providing users with the most current information and experimental features.
Mistral announces Codestral, its first programming focused AI model – Mistral has launched Codestral, a 22B parameter AI model specializing in coding tasks for over 80 programming languages, outperforming previous models and being utilized by industry partners.
OpenAI has a has a new version of ChatGPT just for universities – OpenAI introduces ChatGPT Edu, a version of ChatGPT designed for universities, allowing students and faculty to deploy AI more broadly for teaching and research purposes.
Anthropic’s AI now lets you create bots to work for you – Anthropic’s AI chatbot Claude introduces a new feature allowing users to create personalized AI assistants for tasks like email management, purchasing, and image analysis, with applications in customer service and virtual interior design.
You.com empowers users to create personalized AI assistants with top language models – You.com launches Custom Assistants, allowing users to create personalized AI assistants using top language models, aiming to make powerful language models more accessible and adaptable to individual needs.
Telegram gets an in-app Copilot bot – Microsoft has added an official Copilot bot within the messaging app Telegram, allowing users to search, ask questions, and converse with the AI chatbot.
Udio introduces new udio-130 music generation model and more advanced features – Udio introduces a new udio-130 music generation model that generates 2 minutes of audio to help create more coherent and structured tracks, along with advanced features like random seed settings and cue word intensity adjustment.
ElevenLabs’ AI generator makes explosions or other sound effects with just a prompt – ElevenLabs introduces an AI generator for sound effects, allowing users to create up to 22 seconds of audio based on prompts, with options for commercial use and a character count limit for free users.
Perplexity AI’s new feature will turn your searches into shareable pages – Perplexity AI introduces a new feature called Perplexity Pages, allowing users to create visually appealing and shareable web pages with AI-generated content, including text, media, and customizable sections.
Business
Iyo thinks its gen AI earbuds can succeed where Humane and Rabbit stumbled – Iyo is introducing Bluetooth earbuds with generative AI technology, building on consumer familiarity with voice assistants and aiming to provide immediate value as a premium headphone product.
Anthropic hires former OpenAI safety lead to head up new team – Former OpenAI safety lead Jan Leike joins Anthropic to head a new “superalignment” team focused on AI safety and security, following his public criticism of OpenAI’s approach to the same.
OpenAI Says It Has Begun Training a New Flagship A.I. Model – OpenAI has started training a new flagship AI model to succeed GPT-4, aiming to achieve “artificial general intelligence” and address safety concerns.
OpenAI is restarting its robotics research group – OpenAI LLC, known for ChatGPT, is restarting its robotics research group after a three-year break, hiring a robotics engineer and investing in humanoid developer Figure AI’s Series B fundraising.
China’s capital city is opening a robotaxi route to a major train station, startup Pony.ai says – Beijing is expanding its robotaxi operation, allowing tests with human staff inside to run a route between a major train station and a suburban area, with plans to go fully driverless by the end of the year or early next year.
Arm Partners with Samsung & TSMC for Advanced AI Chip Designs – Arm partners with Samsung and TSMC to launch new AI-centric CPU and GPU designs, software tools, and a shift in delivery strategy to expedite AI chip development for smartphones and PCs.
Tech giants form an industry group to help develop next-gen AI chip components – Tech giants form an industry group to design next-gen interconnects for AI accelerator hardware, aiming to create an open standard for linking AI accelerator chips and reduce dependency on dominant vendors like Nvidia.
Apple Plans AI-Based Siri Overhaul to Control Individual App Functions – Apple is planning to overhaul Siri with more advanced AI, allowing users to control individual app functions with their voice.
Tribeca Festival to Debut Short Films Made Using OpenAI – Tribeca Festival will debut five short films made using OpenAI’s Sora model, a text-to-video technology, with filmmakers educated about the model and required to follow specific agreements.
Saudi Fund Joins $400 Million Financing for China AI Firm Zhipu – Saudi fund invests $400 million in Chinese AI firm Zhipu, marking the first known foreign backing of a major Chinese player in generative artificial intelligence.
The Very Slow Restart of G.M.’s Cruise Driverless Car Business – G.M.’s driverless car future faces setbacks after a pedestrian accident, causing delays in the aggressive expansion of its robot taxi services.
Dell stock plummets 20% because its AI play is falling short – Dell’s stock plummets 20% due to concerns about its not-yet-profitable AI servers, but analysts maintain confidence in the company’s AI strategy.
OpenAI’s new safety committee is made up of all insiders – OpenAI forms a new safety committee staffed with company insiders, including the CEO, to oversee critical safety and security decisions, amid criticism and high-profile departures from the safety side of its technical team.
AI training data has a price tag that only Big Tech can afford – AI training data is becoming increasingly expensive, putting it out of reach for all but the wealthiest tech companies, leading to concerns about centralization and lack of independent scrutiny in AI development.
Research
Google AI Introduce AGREE : A Machine Learning Framework that Enables LLMs to Self-Ground the Claims in their Responses and to Provide Precise Citations – Google AI introduces AGREE, a framework that enables LLMs to self-ground their responses and provide accurate citations, effectively addressing the issue of “hallucination” and improving reliability in domains requiring high factual accuracy.
The future of financial analysis: How GPT-4 is disrupting the industry, according to new research – GPT-4, a large language model, has been shown to outperform human analysts in predicting corporate earnings, potentially transforming the role of financial analysts.
Transformers Can Do Arithmetic with the Right Embeddings – Transformers can achieve high accuracy in solving large and complex arithmetic problems by adding position embeddings to each digit, which also leads to improvements in other multi-step reasoning tasks.
Combining the Best of Both Worlds: Retrieval-Augmented Generation for Knowledge-Intensive Natural Language Processing – Combining parametric and non-parametric memory in Retrieval-Augmented Generation (RAG) models significantly improves performance in knowledge-intensive NLP tasks, setting a new benchmark in the field.
FIFO-Diffusion: Generating Infinite Videos from Text without Training – A new inference technique called FIFO-Diffusion uses a pretrained diffusion model for text-conditional video generation, capable of generating infinitely long videos without training by iteratively performing diagonal denoising and introducing latent partitioning and lookahead denoising to reduce the training-inference gap.
Contextual Position Encoding: Learning to Count What’s Important – A new position encoding method, Contextual Position Encoding (CoPE), allows for more general position addressing by conditioning positions on context, improving performance on language modeling and coding tasks.
LLaMA-NAS: Efficient Neural Architecture Search for Large Language Models – Efficient neural architecture search for large language models is proposed, using one-shot NAS to find smaller, less computationally complex network architectures with a 1.5x reduction in model size and 1.3x speedup in throughput for certain tasks.
Grokked Transformers are Implicit Reasoners: A Mechanistic Journey to the Edge of Generalization – Transformers can learn implicit reasoning through extended training, with varying levels of generalization for different reasoning types, and potential improvements to the transformer architecture are suggested.
The Road Less Scheduled – A new approach to learning rate schedules in AI outperforms existing methods without the need for specifying an optimization stopping step.
From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step – Learning to internalize the concept of CoT is a step-by-step process, transitioning from explicit to implicit understanding.
Is In-Context Learning Sufficient for Instruction Following in LLMs? – In-context learning (ICL) in long-context LLMs shows promise for instruction following, but still underperforms compared to instruction fine-tuning, with additional ICL examples not consistently improving performance.
Nearest Neighbor Speculative Decoding for LLM Generation and Attribution – A new semi-parametric language modeling approach called Nearest Neighbor Speculative Decoding (NEST) enhances generation quality, attribution, and speed, outperforming conventional methods and competing with in-context retrieval augmentation.
Phased Consistency Model – A new Phased Consistency Model (PCM) is proposed to address limitations in the current design of text-conditioned image generation, outperforming previous methods across multiple step generation settings and showing versatility in video generation.
GPT-4 didn’t ace the bar exam after all, MIT research suggests — it barely passed – New research suggests that GPT-4’s claim of scoring in the top 10% on the bar exam was skewed, and it actually scored much lower when compared to first-time test takers.
First ‘bilingual’ brain-reading device decodes Spanish and English words – A bilingual person with a brain implant can communicate in both Spanish and English using an AI system that decodes their neural patterns, providing insights into language processing and potential for restoring multilingual speech.
Concerns
Viral ‘All Eyes on Rafah’ image inspires wave of AI-generated Instagram posts about Israel-Hamas war – AI-generated images related to the Israel-Hamas war, including pro-Israel and pro-Palestinian content, have gone viral on Instagram, sparking a debate about the use of AI in political speech online.
OpenAI Says Russia and China Used Its A.I. in Covert Campaigns – Russia and China used OpenAI’s A.I. in covert campaigns to manipulate public opinion and influence geopolitics, raising concerns about the impact of generative A.I. on online disinformation.
Meta says it removed six influence campaigns including those from Israel and China – Meta removed six influence campaigns, including those from Israel and China, which used AI to create the illusion of wider support for certain viewpoints and influence political discourse on its platforms.
Deepfake of U.S. Official Appears After Shift on Ukraine Attacks in Russia – A deepfake video of a U.S. official discussing Ukraine’s potential strikes in Russia has surfaced, raising concerns about the use of AI-powered disinformation.
Most Americans expect AI abuses to impact 2024 election: survey – Most Americans expect AI abuses to impact the 2024 election, with concerns about manipulation through social media, fake information, and voter persuasion, as well as a lack of confidence in detecting faked media.
A Small Army Combating a Flood of Deepfakes in India’s Election – A small army in India is combating a flood of deepfakes during a high-stakes election, using A.I.-powered mimicry, ventriloquy, and deceptive editing effects.
Godmode’ GPT-4o jailbreak released by hacker — powerful exploit was quickly banned – A hacker released a jailbroken version of GPT-4o, allowing the AI chatbot to bypass restrictions and perform dangerous tasks, but it was quickly banned by OpenAI.
Hugging Face says it detected ‘unauthorized access’ to its AI model hosting platform – AI startup Hugging Face detected unauthorized access to its platform, leading to the potential compromise of Spaces secrets, prompting the revocation of tokens and a security investigation.
Meta’s AI is summarizing some bizarre Facebook comment sections – Meta’s AI is summarizing Facebook comment sections, injecting itself into the wild and often ridiculous discussions, prompting privacy concerns and raising questions about the usefulness of the generated summaries.
Policy
With the EU AI Act incoming this summer, the bloc lays out its plan for AI governance – The EU is establishing the AI Office to regulate AI risks, foster innovation, and influence global AI governance, with five units focusing on regulation and compliance, AI safety, excellence in AI and robotics, AI for social good, and AI innovation and policy coordination.
Big Tech goes on AI charm offensive in Europe as regulators circle – Big Tech promotes the benefits of AI at a European conference while regulators aim to address potential harms, with discussions on AI’s potential for solving global problems and the need for responsible innovation.
US Is Slowing AI Chip Exports to Middle East by Nvidia, AMD – US officials have slowed the issuing of licenses to chipmakers such as Nvidia Corp. and Advanced Micro Devices Inc. for exporting AI chips to the Middle East.
Analysis
New AI products much hyped but not much used, study says – AI products like ChatGPT are much hyped but not widely used, with only 2% of British respondents using such tools on a daily basis, according to a study surveying 12,000 people in six countries.
Inside the Google algorithm – Google’s algorithm, which powers its search engine, has been better understood thanks to a recent leak of API documentation, revealing new insights into how Google ranks content and operates.
AI Is Making Economists Rethink the Story of Automation – AI’s impact on the labor market and economists’ evolving understanding of technology’s effects on jobs and wages.
The New ChatGPT Offers a Lesson in A.I. Hype – OpenAI’s latest version of ChatGPT, featuring humanlike inflections and emotions, was released without most of its new features, including the improved voice and real-time analysis capabilities.
Expert Opinions
Will Scaling Solve Robotics? – Debate on whether training large neural networks on extensive datasets is a feasible solution to robotics, with arguments for and against scaling, and suggestions to explore other approaches and focus on real-world mobile manipulation and easy-to-use systems.
Source: Read MoreÂ
