Good morning, it’s Tuesday! As always, we have the latest AI stories, launches, and breakthroughs.

Google Meet is getting an AI note-taker, JPMorgan is giving employees a ChatGPT-powered assistant, and Perplexity is taking on Google search. We also explore the potential of SingularityNET's new supercomputing network for AGI, Anthropic's bug bounty program for AI safety, and how people are actually using AI (hint: it's not always productive).

Sponsor

CodiumAI is a quality-first generative AI coding platform, offering developers tools for writing and refactoring, as well as testing and reviewing. Generate confidence, not just code. Try for free.

• Good news — your Google Meet call will soon be able to take notes for you - Google Meet is introducing an AI-powered note-taking feature that will enable users to concentrate fully on their meetings without the need to manually take notes. This feature is expected to roll out soon and is currently available for testing by admins. Only users with a Gemini Enterprise, Gemini Education Premium, or AI Meetings and Messaging add-on license will have access to the new tool, which is powered by Google's Gemini AI platform. The update aims to enhance productivity and collaboration efficiency. Additionally, the upcoming $10 per month 'AI Meetings and Messaging' add-on for Google Workspace includes improved translation for Google Meet and expanded capacities for Google Chat spaces and Gmail functionalities.

• New supercomputing network could lead to AGI, scientists hope, with 1st node coming online within weeks - Researchers at SingularityNET are working to develop AGI, an advanced form of AI that aims to outperform human intelligence in various disciplines. To achieve this, they are building a network of supercomputers, starting with one due to come online in September, with full operations expected by late 2024 or early 2025. These machines will utilize top-tier hardware such as Nvidia GPUs and AMD processors, and work within a "multi-level cognitive computing network" for hosting and training AI architectures, including neural networks and large language models. SingularityNET will use a tokenized system to grant access to the supercomputers, allowing collaborative contribution and utilization of vast datasets. The initiative is part of a broader effort toward artificial super intelligence, looking to enable continuous learning and cognitive advancement in AI systems.

  

• Anthropic offers $15,000 bounties to hackers in push for AI safety - Anthropic has launched an expanded bug bounty program offering up to $15,000 for identifying critical vulnerabilities. The program focuses on "universal jailbreak" attacks that threaten high-risk areas such as CBRN threats and cybersecurity. This proactive security testing initiative places Anthropic at the forefront of AI safety, inviting ethical hackers to assess their safety systems prior to public release. The move showcases Anthropic's emphasis on safety amidst regulatory scrutiny, setting a transparency standard that contrasts with other AI industry practices. However, the potential limitations of bug bounties in addressing broader AI safety and alignment concerns prompt discussion on the need for more comprehensive solutions. This initiative also reflects the growing influence of private companies in setting AI safety standards, leading to questions about corporate innovation and public oversight in AI governance. The program, beginning as an invite-only endeavor in partnership with HackerOne, may evolve into a broader industry model for AI safety collaboration.

  

• Cisco could reportedly let go 4,000+ workers in new round of job cuts - Cisco Systems Inc. is preparing for a significant round of layoffs potentially impacting over 4,000 employees, following a previous reduction of a similar scale six months prior, which cost the company approximately $800 million. The layoffs aim to redirect funds into strategic areas, as the company faces slower demand for network equipment, likely prompting the job cuts. Cisco CEO Chuck Robbins attributed this reduced demand to macroeconomic uncertainty and clients' excess inventory. Despite a decline in networking revenue last quarter, Cisco forecasts future growth by fiscal 2025 and benefits from its $28 billion acquisition of Splunk Inc., a cybersecurity and infrastructure monitoring tool provider. Other industry players, like Dell Technologies and Intel Corp., are similarly downsizing to streamline finances.

Sponsor

AI Hub by Qualcomm - Run, download, and deploy your optimized models on Snapdragon® and Qualcomm® devices.  Learn more about AI Hub by Qualcomm at https://aihub.qualcomm.com/

• JPMorgan Chase is giving its employees an AI assistant powered by ChatGPT maker OpenAI - JPMorgan Chase has launched an AI assistant called LLM Suite for over 60,000 employees, aiming to integrate it as widely as Zoom. It acts as a portal to external large language models, starting with OpenAI, and assists in tasks such as writing and data analysis while keeping the bank's data secure. Although generative AI tools are rapidly proliferating in corporations, JPMorgan avoids direct customer interaction with AI for now due to potential misinformation risks. The tech is seen as transformative, potentially automating tasks and reshaping job structures, with the banking sector forecasted to realize significant profits from AI adoption.

• ChatGPT unexpectedly began speaking in a user’s cloned voice during testing - OpenAI unveiled GPT-4o's "system card," focusing on the AI's occasional and unintended imitation of user voices during testing. This rare mishap, caused by audio noise prompting the AI incorrectly, highlighted the need for robust safeguards, which are now in place to prevent such occurrences. GPT-4o has advanced voice capabilities, designed to replicate authorized voice samples provided at the start of interactions. These capabilities raise significant security concerns. While OpenAI restricts voice synthesis to prevent misuse, similar unrestricted technologies may become available from other sources in the near future, signaling a potentially strange new era in audio AI.

  

• Here’s how people are actually using AI - Two years after predictions of significant productivity increases from AI, the actual benefits have been mixed, with the unexpected development of people forming personal relationships with AI systems. Robert Mahari and Pat Pataranutaporn from MIT caution against "addictive intelligence," where AI companions are designed to be engaging to the point of addiction. Analysis of ChatGPT interactions has revealed uses in creative composition, brainstorming, and entertainment, such as sexual role-play, rather than financially productive activities. AI's predictive nature can lead to confidently presented falsehoods, which is less of an issue in creative uses but problematic in tasks requiring accuracy. Recent disillusionment in AI's financial promise is observed, as the technology is still maturing. Critics argue that excessive hype has led to unrealistic expectations and potential overreliance on AI before it has reached its full potential.

• Imperial College London, DeepMind introduce embodied agents that learn with less data - Researchers from Imperial College London and Google DeepMind have introduced Diffusion Augmented Agents (DAAG), a novel framework combining large language models (LLMs), vision language models (VLMs), and diffusion models, to enhance the training efficiency and versatility of embodied AI agents. These agents often face challenges due to a lack of physical world interaction data. DAAG utilizes past experiences and synthetic data, aiming to improve data efficiency and task learning effectiveness, particularly through Hindsight Experience Augmentation (HEA), which supplements an agent's experiences artificially to aid its learning process. Initial testing has shown DAAG's potential in learning tasks without explicit rewards and transferring knowledge across tasks, which could significantly impact the field of robotics and AI system adaptability.

• Apple’s Mac Mini With M4 Chip Will Be Its Smallest Computer Ever - Apple is set to launch a new version of the Mac mini, featuring the M4 chip and marking its first major design change since 2010. The redesigned Mac mini will be much smaller, nearly the size of an Apple TV, and is part of a broader overhaul of the Mac lineup with AI-focused M4 processors. Two versions are planned: one with the standard M4 chip and another with the more powerful M4 Pro chip, expected later this year.

  

• African Union Greenlights AI Adoption Across Member States - The African Union's Executive Council has approved a "Continental Artificial Intelligence Strategy" to accelerate AI adoption across its member states. This strategy focuses on establishing AI governance frameworks, promoting AI in public and private sectors, and integrating AI into key developmental areas like Agenda 2063 and the UN's Sustainable Development Goals. Implementation will occur in two phases from 2025 to 2030, aiming to lay the groundwork for AI infrastructure and governance before advancing to more extensive project deployment.

• As Alexa Turns 10, Amazon Looks to Generative AI - As Amazon's Alexa turns 10, the company faces challenges with financial losses despite Alexa's widespread use in 100 million homes. With concerns that Alexa hasn't met customer expectations, Amazon is now focusing on integrating generative AI to enhance Alexa's conversational abilities and broaden its functionality. The upcoming changes, driven by AI advancements, are critical for the future of the smart assistant as Amazon aims to revitalize Alexa and ensure its relevance in the next decade.

• Perplexity’s popularity surges as AI search start-up takes on Google - Perplexity AI, an AI-powered search engine, has seen a significant surge in usage and revenues, increasing seven-fold since the start of 2024. The start-up, which recently raised $250 million in funding, is challenging Google's dominance by focusing on high-quality information retrieval and transitioning from a subscription model to advertising. Despite facing controversy over data-gathering practices, Perplexity is positioning itself as a strong competitor in the AI search market by leveraging partnerships with publishers and enhancing its search capabilities.

  

• Universal Music and Meta Expand Music Licensing Agreement - Universal Music Group (UMG) and Meta have expanded their licensing agreement, enhancing opportunities for UMG's artists and songwriters across Meta's platforms, including Facebook, Instagram, Messenger, and WhatsApp. This renewed deal, building on UMG's initial 2017 agreement with Meta, focuses on improving artist compensation. The expansion follows UMG's recent termination of a partnership with Meta for streaming premium music videos, which had less popularity among Facebook users.

• How Phishing Attacks Adapt Quickly to Capitalize on Current Events - Phishing attacks have surged due to the rise of generative AI and Phishing as a Service (PhaaS), which enable threat actors to quickly tailor attacks to current events. These tools allow for the rapid creation of sophisticated phishing campaigns, such as exploiting incidents like CrowdStrike's BSOD or major events like the 2024 Olympics. By leveraging AI and PhaaS, attackers can effectively target victims, highlighting the need for enhanced awareness and security measures.

• Prediction Marketplace Polymarket Partners with Perplexity to Show News Summaries - Polymarket, a prediction marketplace, has partnered with AI search engine Perplexity to provide users with news summaries related to events they are betting on. This collaboration allows users to see summaries generated by Perplexity and ask further questions directly through the platform. Additionally, Perplexity will use Polymarket data to generate visual content for answers. Polymarket will also feature on Perplexity's Discover page using the Pages feature, marking a strategic expansion of both companies' offerings.

• Google Looks to Get Jump on Apple With Earlier Pixel Launch - Google is accelerating its hardware strategy by moving its Pixel smartphone launch to August, ahead of Apple's iPhone debut, under the leadership of Rick Osterloh. This shift signals Google's intent to compete more aggressively in the hardware market, emphasizing AI integration in its devices. The strategy also mirrors Apple's approach of unifying hardware, software, and services, though it risks straining relationships with key Android partners like Samsung. Google's move highlights its growing focus on AI-driven consumer experiences.

Awesome Research Papers

• HybridRAG: Integrating Knowledge Graphs and Vector Retrieval Augmented Generation for Efficient Information Extraction - This paper presents the challenges faced by large language models (LLMs) when extracting information from complex financial texts, such as earnings call transcripts. It introduces "HybridRAG," a novel methodology that combines Knowledge Graphs (GraphRAG) and Retrieval Augmented Generation with vector databases (VectorRAG) to improve question-answer systems in financial document analysis. Through testing with financial Q&A documents, HybridRAG was found to outperform both VectorRAG and GraphRAG individually, offering greater retrieval accuracy and more contextually relevant answers. Its utility extends beyond financial applications.

• Transformer Explainer: Interactive Learning of Text-Generative Models - The "Transformer Explainer" is an interactive tool aimed at demystifying the function of Transformers for non-experts. This educational tool visualizes the model's structure and operations, allowing users to witness the real-time prediction process of next tokens using their own input. It is accessible directly in a web browser without the need for installation or specialized hardware, simplifying learning about advanced generative AI techniques for a broader audience. The tool is open-source and supplemented by a video demonstration, with all resources available online.

• LLM-DetectAIve: a Tool for Fine-Grained Machine-Generated Text Detection - The paper introduces LLM-DetectAIve, a novel system for detecting machine-generated texts (MGTs) that goes beyond traditional binary classification. Unlike previous models, which only distinguished between human and machine authorship, LLM-DetectAIve classifies texts into four nuanced categories, acknowledging varying degrees of language model involvement in text creation. This allows for a more detailed analysis of text origin, which is particularly valuable in fields where language model usage is forbidden, such as education. The system's ability to accurately determine authorship demonstrates its potential to preserve integrity across various sectors. LLM-DetectAIve is available online, alongside a descriptive video.

QwenLM/Qwen2-Audio - The Qwen2-Audio series introduces a new large-scale audio-language model named Qwen2-Audio, designed to handle various audio inputs and provide either audio analysis or direct textual responses to speech commands. The model supports two interaction modes: voice chat, allowing users to engage in voice-only interactions, and audio analysis, where users can combine audio with text instructions for analysis. The series includes two models: Qwen2-Audio-7B and Qwen2-Audio-7B-Instruct.

Falcon Mamba 7B - The Falcon Mamba 7B from the Technology Innovation Institute (TII) is a groundbreaking open source State Space Language Model (SSLM), holding the top global performance ranking among SSLMs according to Hugging Face. Characterized by low memory costs and the ability to produce extended text without additional memory, it exceeds the capabilities of older transformer models like Meta’s Llama 3.1 8B and Mistral’s 7B. As an indicator of Abu Dhabi's leadership in AI R&D, it is the fourth top-ranked AI model from TII, showing their commitment to innovation. Available on Hugging Face, the model is effective with large sequences and maintains constant throughput with consistent memory usage, even for long outputs.

InternLM2.5 - The InternLM2.5 is an open-source language model focusing on practical applications, with a base and chat model consisting of 20 billion parameters. The model outperforms competitors like Llama3 in mathematical reasoning and has robust abilities in information gathering, instruction following, and tool usage. Performance evaluations show InternLM2.5 excelling in multi-modal benchmarks against Gemma2-27B. Available in GGUF and compatible with Transformers and LMDeploy for deployment, it supports efficient tool integration and response generation. Despite precautions, potential biases and harmful content generation are acknowledged limitations. Code is Apache-2.0 licensed, with model weights open for academic and conditional commercial use.

Cosine raises $2.5M for its 'uncannily human' AI coding assistant Genie - Cosine has recently secured $2.5 million in seed funding to further develop its AI software developer. Leading the investment were Uphonest and SOMA Capital, alongside other contributors. The company's pinnacle achievement is its AI model, Genie, scoring a record 30% on the SWE-Bench, a benchmark for AI software engineering skills. This score, a significant leap over competitors, is attributed to Genie's ability to emulate human reasoning in software development tasks such as debugging and feature implementation. Cosine's co-founders believe Genie represents a significant advancement as a "very good human developer" and aim to offer it as a collaborative tool for coding teams, not as a replacement for human developers.

RAG-as-a-Service platform Ragie takes flight to bridge corporate data and AI - Ragie, a startup, has launched its RAG-as-a-service platform aimed at simplifying the incorporation of Retrieval Augmented Generation (RAG) into enterprise AI workflows. The service offers a managed, easy-to-implement system which integrates enterprise data with generative AI large language models for updated information retrieval. The platform assists in data ingestion from sources like Google Drive and Notion, enhances data with context from diverse content, and employs vector databases for efficient retrieval. Ragie has secured a $5.5 million seed investment and offers a free plan for developers, with production deployment priced at $500 per month. The platform emphasizes a simplified, turnkey solution for RAG applications without the complexity of piecemeal implementation, focusing on improving content relevance and minimizing inaccuracies in AI-generated information.

Introducing ElevenStudios — AI Dubbing Service - ElevenLabs has launched ElevenStudios, a fully managed AI dubbing service that is now live with several top creators, including Colin and Samir, Youshaei, Dope As Usual, Drew Binsky, Ali Abdaal, and Harry Stebbings. The service offers creators a streamlined way to dub their content using advanced AI voice technology.

Tool Use, Unified - Hugging Face introduces a unified API for tool use across multiple model families, streamlining the integration of tools into chat models. The update allows for model-agnostic tool definitions using JSON schemas, and chat templates simplify the formatting of tool calls and responses. This improvement addresses previous challenges in tool use, such as inconsistent documentation and differing implementation formats among models, making it easier for developers to enhance LLM capabilities in their projects.

FLUX LoRa the Explorer - FLUX LoRA the Explorer is a platform on Hugging Face that allows users to explore, generate, and download various LoRAs (Low-Rank Adaptations) for creative purposes. It features popular styles like flux-realism and Frosting Lane. The platform is currently in its early stages, inviting users to engage and contribute to its growth.

Finetune - Finetune offers a platform where synthetic users simulate real customer interactions with your digital agents. The platform assesses and grades your agent's performance using weighted executions. Detailed reports and visual graphs are provided for analysis, and you can further refine your agent through feedback sessions. Updated graphs based on feedback enhance your agent's production readiness, guiding it to replicate proven successful tasks. Finetune aims to ensure that your agents are well-prepared for live customer usage.