The latest Deeper Insights blogs

The Silent MVP: Computer Vision in Modern Sports

Explore how AI and computer vision transform sports through enhanced broadcasting, improved officiating, and revolutionary fan experiences with advanced analytics. [Read more]

Saving Marine Life with AI: A New Era in Ocean Conservation

Advanced AI technology aids The Ocean Cleanup in removing ocean plastics and protecting marine life, showcasing potential for broad applications in industries like healthcare and urban management. [Read more]

AI-Driven Agriculture: Agritech for a Sustainable Future

Discover how Artificial Intelligence is revolutionising agriculture, addressing pressing challenges like climate change, workforce shortages, and resource management to enhance efficiency and sustainability for a resilient agricultural future. [Read more]

LoRa to QDoRA: A Technical Analysis of Efficient LLM Fine-Tuning

Large language models (LLMs) need fine-tuning for specific tasks. Methods like LoRA and QLoRA improve efficiency. The latest innovation, QDoRA, further enhances performance and reduces computational demands. [Read more]

Featured GenAI news

AI Summit Proposes AI Kill Switch - Fortune.com

The AI Seoul Summit will discuss implementing a "kill switch" to ensure AI systems can be instantly shut down if they behave harmfully. Building on the Bletchley Park Summit's agreements, global leaders aim to enhance AI safety and foster responsible innovation. This initiative marks a significant step in regulating advanced AI technologies. [Read more]

EU's AI Act Incoming - TechCrunch.com

The EU is set to introduce the AI Act this summer, creating a comprehensive framework for AI regulation. This legislation aims to ensure ethical AI development and protect fundamental rights, impacting tech companies globally. The Act categorises AI systems by risk and sets strict guidelines for high-risk applications, marking a significant step in AI governance. [Read more]

Scarlett Johansson vs. OpenAI - TheVerge.com

Scarlett Johansson is suing OpenAI, claiming the company used a voice eerily similar to hers for their ChatGPT 4.0 system without her permission. Despite initially declining to voice the AI, Johansson noticed the similarity, leading her to hire legal counsel. OpenAI has since paused using the voice and apologised for the misunderstanding, asserting the voice was recorded by a professional actor. [Read more]

The New C-Suite AI Role - Forbes.com

As AI technology becomes integral to business strategies, the Chief AI Officer (CAIO) role is emerging as a crucial C-suite position. CAIOs oversee AI integration across departments, ensuring ethical use, driving innovation, and aligning AI initiatives with business goals. This role is gaining prominence across various industries, from tech startups to traditional businesses, highlighting the increasing importance of AI expertise in leadership. [Read more]

GenAI news snapshots - Industry report

• OpenAI has shared insights from its preview of the Voice Engine model, which generates realistic speech from text and a brief audio sample. Despite its success in creating natural-sounding voices, OpenAI is carefully considering a broader release due to potential misuse, aiming to foster discussions on responsible deployment. [Read more]
• FCC Proposes AI Disclosure Rule. The FCC has proposed a new rule requiring all AI-generated content in political ads on TV and radio to be disclosed. This measure aims to increase transparency and protect voters from potential misinformation, especially with the growing use of AI tools in election campaigns. If adopted, broadcasters must clearly label any AI-generated elements in political advertisements. [Read more]
• Mastercard is using generative AI to enhance credit card fraud detection, doubling the speed at which compromised cards are identified. This new technology scans transaction data across billions of cards and millions of merchants, significantly improving accuracy and reducing false positives. The advancements aim to protect cardholders and secure the payment ecosystem against emerging threats. [Read more]
• The U.S. Air Force successfully tested an AI-controlled F-16 fighter jet, with Air Force Secretary Frank Kendall aboard. This groundbreaking event at Edwards Air Force Base showcased the potential of AI in military aviation. The Air Force plans to develop a fleet of over 1,000 AI-enabled unmanned aircraft by 2028, aiming to enhance security and reduce costs. [Read more]
• Emad Mostaque steps down as CEO of Stability AI to pursue decentralized AI, with Shan Shan Wong and Christian Laforte appointed as interim co-CEOs. The board expresses gratitude for Mostaque's leadership and confidence in Wong and Laforte to guide the company forward, emphasizing a commitment to preserving Stability AI's vision and position as a leader in open multi-modal generative AI. [Read more]
• Apple AI has developed an on-device model, described in their recent paper "ReALM," which significantly outperforms GPT-4 by enhancing Siri's capabilities. This model not only attempts to describe images but also considers the context of on-screen content and active tasks, promising a substantial boost in the usefulness of the voice assistant. [Read more]
• Amazon has invested $2.75 billion in Anthropic, marking its largest venture investment to date, finalizing a total commitment of $4 billion to the OpenAI rival. Dr. Swami Sivasubramanian from Amazon Web Services highlights their history with Anthropic, emphasizing their collaboration in deploying advanced generative AI applications worldwide. [Read more]
• Drake has removed his diss track "Taylor Made," featuring AI-altered vocals of Tupac Shakur, after receiving a legal threat from Tupac's lawyers. The track, intended as a response in Drake's feud with Kendrick Lamar, was taken down following the ultimatum, though its impact on Drake's earnings and the broader implications of AI in music remain subjects of discussion. [Read more]
• Ray-Ban Meta Smart Glasses have been updated to include multimodal AI, allowing the AI assistant to process various types of information like photos, audio, and text. [Read more]
• DALL-E now integrates image editing tools within ChatGPT, allowing users to refine their AI-generated creations seamlessly. With preset style suggestions and improved user-friendliness, this update enhances the collaborative creative process and expands the possibilities of AI-generated imagery. [Read more]
• TikTok's AI-Powered Ad Boost. TikTok is enhancing its advertising business with the launch of the "TikTok Symphony" AI suite. This new toolset includes an AI video generator and script-writing assistant, aimed at helping brands create and optimize ad content efficiently. The suite is designed to streamline video production and improve campaign performance on the platform. [Read more]
• Meta has unveiled the next generation of its Meta Training and Inference Accelerator, a series of custom chips tailored for AI workloads. This advancement is set to boost the efficiency of Meta's ranking and recommendation ads models and represents a key investment in enhancing AI infrastructure to improve user experiences across Meta's products and services. [Read more]
• U.S. intelligence agencies are cautiously adopting generative AI to manage the massive influx of data and maintain a competitive edge. These AI tools assist in identifying key information and predicting potential threats, although concerns about data privacy and AI model security persist. This strategic move aims to counter adversaries' AI capabilities while ensuring human oversight remains integral. [Read more]
• Stable Audio 2.0 revolutionises AI-generated audio by producing high-quality, structured music tracks up to three minutes long at 44.1kHz stereo, and introduces audio-to-audio generation that transforms user-uploaded samples via natural language prompts. Trained on a fully licensed dataset, this model ensures fair compensation for creators and is available for use on the Stable Audio website. [Read more]

GenAI in Robotics

• Columbia engineers have developed Emo, a robotic face covered in silicon that enhances human-robot interactions by making eye contact and predicting and mimicking human smiles using two AI models. This technology represents a significant step forward in enabling robots to anticipate human facial expressions, potentially improving trust and engagement between humans and robots. [Read more]
• AI robots may require a body. Meta researchers suggest that for AI to reach human-level intelligence, it must have a physical form that allows it to interact with and navigate the real world, similar to how infants learn. [Read more]
• Sanctuary AI is set to deliver its humanoid robots to a Magna manufacturing facility in Austria, which assembles cars for leading European automakers such as Mercedes, Jaguar, and BMW.  [Read more]
• Tesla is set to unveil its new 'robotaxi,' a fully autonomous vehicle designed without pedals or steering wheel, in August. This vehicle, distinct from consumer models that can become robotaxis via software updates, is built specifically for autonomous driving. [Read more]
• Video: Robot doing housework. Over the past year, Google DeepMind has advanced its ALOHA 2 fleet under the initiative "ALOHA Unleashed," enhancing the scale and complexity of autonomous tasks. One highlighted achievement includes the development of a robot capable of autonomously hanging a shirt on a hanger. [Read more]

GenAI tools: LLM models

• Infinite Text Processing with Google's Infini-Attention. Google researchers introduce Infini-attention, a technique enabling LLMs to process text of infinite length by extending their context window while maintaining memory and compute efficiency. This innovation addresses limitations in traditional models, ensuring consistent performance even with longer texts. [Read more]
• AIDE Masters Kaggle: AI Achieves Human-Level Data Science. An AI-powered data science agent, achieves human-level performance in Kaggle competitions, marking a significant milestone in the field of data science. Kaggle competitions serve as a standard for evaluating data scientists' skills, and this achievement demonstrates AIDE's ability to solve complex real-world problems and develop high-performing machine learning solutions. [Read more]
• R2R: Streamlining RAG Deployment with FastAPI. R2R (RAG to Riches) provides a streamlined framework for delivering high-quality Retrieval-Augmented Generation (RAG) to end users. With customisable pipelines and a feature-rich FastAPI implementation, developers can efficiently deploy and scale RAG-based applications. [Read more]
• DreamDA: Enhancing Data Augmentation with Diffusion Models. DreamDA, is a classification-oriented framework that leverages diffusion models for generative data augmentation. DreamDA addresses issues in existing methods by generating diverse samples adhering to the original data distribution and introducing a self-training paradigm for accurate label generation. [Read more]
• Extreme Low-Bit Quantisation with HQQ+. In recent developments, extreme low-bit quantisation techniques such as BitNet and 1.58 bit have garnered significant attention for their potential to revolutionise the compute efficiency of large models. This blog post explores the direct quantisation of pre-trained models using HQQ+, adapting the HQQ technique with a low-rank adapter. Results demonstrate significant improvements in output quality, even with 1-bit quantisation, surpassing smaller full-precision models. [Read more]
• Cohere Toolkit: Streamlining AI Deployment Across Major Cloud Platforms. This open-source repository empowers developers to accelerate the creation of AI applications by providing production-ready models deployable across cloud platforms like AWS, Azure, and Cohere's own platform. With access to Command, Embed, and Rerank models, developers can deploy applications in their preferred environments, ensuring security and scalability while connecting to custom data sources. [Read more]
• Jamba, by AI21Labs, was built on top of an SSM-Transformer mixture-of-experts (MoE) architecture, designed to overcome the limitations of traditional Transformer architectures. [Read more + HF]
• DBRX, by Databricks, is a groundbreaking open, general-purpose LLM that sets new benchmarks in performance and efficiency among open LLMs. Offering advancements like a fine-grained mixture-of-experts (MoE) architecture, DBRX not only performs tasks up to twice as fast as similar models but also operates at nearly 40% smaller size while maintaining competitive quality with top models such as Gemini 1.0 Pro. [Read more + HF]
• Grok-1.5, from xAI, boasts enhanced reasoning abilities and an extended context length of 128,000 tokens, representing significant progress in understanding long contexts and problem-solving. Soon to be available on the 𝕏 platform for early testers and current users. [Read more + GitHub]
• Meta has launched Llama 3, an advanced AI model available in both 8B and 70B versions, integrated into Meta AI to enhance coding tasks and problem-solving capabilities. [Read more]
• OpenELM, by Apple, is a series of small, open-source AI models capable of running directly on devices. This initiative positions Apple alongside Google, Samsung, and Microsoft in advancing generative AI technologies that operate independently of cloud servers, enhancing on-device processing capabilities. [Read more + HF]

GenAI tools: Visual models

• SwapAnything is a new framework enabling personalised object swapping in images while preserving contextual integrity, offering precise control and faithful adaptation of personalised concepts. [Read more]
• SIMA, by Google DeepMind, is a Scalable Instructable Multiworld Agent, capable of understanding natural-language instructions to navigate and perform tasks across various video game environments. This milestone marks a shift towards developing a generalist AI agent for 3D virtual settings, aiming to leverage video games as training grounds for more versatile and helpful AI systems in real-world scenarios. [Read more]
• SegRefiner is a model-agnostic solution for improving object masks produced by segmentation models. By employing a discrete diffusion process to refine coarse masks, SegRefiner outperforms previous methods across various segmentation tasks, demonstrating enhanced segmentation metrics and the ability to capture fine details in high-resolution images. [Read more]

GenAI tools: Everything else models

• Gemini 1.5 Pro is now available for public preview on Vertex AI, Google's enterprise-focused AI development platform. With the ability to process up to 1 million tokens of context, it offers powerful capabilities for generating complex content. [Read more]
• Grok-1.5V, the latest multimodal model by xAI with enhanced visual processing capabilities, including documents, diagrams, charts, screenshots, and photographs. Competitive across various domains, Grok excels in understanding real-world spatial relationships, outperforming peers in the RealWorldQA benchmark. Available soon for early testers and existing Grok users. [Read more]

‍