AI News & Launches

NVIDIA Research released the Nemotron-4 340B model family with models designed to perform competitively on a range of evaluation benchmarks. Models include Nemotron-4-340B-Base, Nemotron-4-340B-Instruct, and Nemotron-4-340B-Reward, and are open access under the NVIDIA Open Model License Agreement. These models can be deployed on a single DGX H100 with 8 GPUs in FP8 precision and are particularly effective for generating synthetic data.

The article 'Turning the Tables on AI' from iA Writer discusses ways to use AI for enhancing thought rather than replacing it. It emphasizes using AI to ask questions that stimulate clear expression, avoiding plagiarism of AI suggestions, and properly editing and integrating AI-generated ideas. The author encourages creativity over mere editing and calls for a mindful engagement with AI in the writing process.

SoftBank Corp. introduced a new AI that can transform irate customer voices to calmer tones to reduce call center operators' stress. The 'emotion canceling' technology changes pitch and inflection without altering words. While keeping a trace of anger to help operators respond appropriately, it aims to shield employees from harassment and lengthy confrontations. The AI, trained with over 10,000 voice samples, will alert staff when calls are too abusive or long. SoftBank is set to commercialize the technology in fiscal 2025, looking to enhance the operators' mental wellbeing.

NVIDIA announced the release of Nemotron-4 340B, an open family of models for generating synthetic data to train large language models (LLMs). The Nemotron-4 340B offers an open model license and is optimized for NVIDIA NeMo and TensorRT-LLM. The suite includes base, instruct, and reward models to aid in data quality for robust LLMs across sectors such as healthcare and finance. It leads in the Hugging Face RewardBench leaderboard and can be fine-tuned using NVIDIA NeMo.

Dr. Ian Cutress discusses the challenge LLMs have with hallucinations and presents Lamini's new Memory Tuning method which claims to reduce hallucinations by 95%. Lamini, led by Sharon Zhou and Greg Diamos, has already attracted $25m in Series A funding and provides advanced AI services using AMD GPUs.

This YouTube video provides viewers with an understanding of the workings of the YouTube platform, including content creation, video sharing, and community engagement.

The blog post discusses the future of software engineering with the advent of Generative AI and draws an analogy to aircraft autopilot systems. It suggests that demand for software engineers will grow along with their salaries, and that generative tools will serve as another abstraction layer, not as a replacement. Accessible software engineering will encourage more individuals to enter the field, and a wider array of software is yet to be created, benefiting society at large. The future skill set for software engineers may evolve, with continuous importance of general skills like communication and problem-solving.

Apple's latest generative AI technology will be limited to their most powerful iPhone models, the 15 Pro and 15 Pro Max, leaving over 90% of current iPhone users without access. The technology will initially only support U.S. English and is set to be included in the new iOS 18 operating system. Critical advancements in Siri's functionality will also be exclusive to the latest devices, excluding earlier models of iPhones, HomePod speakers, and Apple Watches from the update.

NewRA is an AI-powered platform that allows enterprises to build chatbots utilizing their data to automate conversations and tasks. It offers cloud-based operations with no setup required, and features Retrieval Augmented Generation for enriched interactions.

Charity Majors’ essay, highlighted by Simon Willison, underlines the irreplaceable role of junior programmers, discussing issues like the 'software as apprenticeship' model, team composition, and the importance of a varied skill set for team dynamics.

Apple Intelligence demonstrates integration of AI into digital lives by connecting apps and simplifying tasks, promising privacy-conscious automation and information retrieval. The technical details include focus on personalization, performance, and on-device strategies, which may set a precedent in AI usability.

AI Top Rank is an AI tool launchpad that ranks and offers insights into various AI products. Users can vote for their favorite tools, see leaderboards, and discover new products weekly.

Humanize AI Text specializes in transforming AI-generated content into natural, human-like text, ensuring it bypasses AI detection tools. The platform offers features like content shaping, multilingual support, readability boost, integrated writing assistant, and provides a unique Human Score to measure the human feel of the text.

Lamini Memory Tuning is a novel technique that embeds facts into Language Learning Models (LLMs), achieving unprecedented accuracy and hallucination reduction. A Fortune 500 company attained 95% accuracy with this approach, a significant leap from the 50% accuracy of conventional methods. By leveraging expert adapters like LoRAs, Lamini Memory Tuning fine-tunes any open-source LLM, ensuring precise fact recall while preserving generalization capabilities. This research breakthrough promises high accuracy, speed, and cost efficiency without compromise.

Unity Catalog is an open and interoperable catalog for data and AI, supporting multi-format tables and assets with various engines. It provides unified governance and is fully open sourced under the Apache 2.0 license.

Adobe employees criticize the company for its handling of a controversy regarding updated Terms of Service that suggested creator content might train AI models. Despite two blog posts for damage control, employees express the need for a better long-term communication plan.

GitHub repository 'ask-fini/paramount' offers tools for agent accuracy measurements of Language Learning Models (LLMs). It focuses on quality assurance, capturing ground truth, and automated regression testing. The project stands under the GPL License but requires a commercial license for larger companies.

Google has partnered with NV Energy to launch a new clean energy initiative aimed at creating reliable and clean electricity. This partnership introduces a new Clean Transition Tariff to invest in technologies like enhanced geothermal to meet electricity demand with carbon-free energy.

Tesorio, a company specializing in finance technology, offers solutions like Accounts Receivable Automation and Cash Flow Management. They uphold a team-oriented culture with benefits like competitive salaries, unlimited PTO, and health insurance. Tesorio's 2023 Panama Offsite for team-building signifies their commitment to employee engagement.

Help Scout is a customer service platform offering AI tools to boost productivity by 36%. Features include AI drafts, AI answers, AI summarize, and AI assist, aiming to reduce resolution times and improve customer interactions.

Nvidia has extended its dominance over machine learning benchmarks with its latest performance in MLPerf, showcasing the power of its GPU systems in AI training. The recent tests included fine-tuning of large language models (LLMs) and graph neural networks, demonstrating significant improvements in training times through software enhancements and efficient system scaling.

Apple Inc. integrates ChatGPT from OpenAI into Siri and new writing tools for its devices, counting on brand and technology distribution as compensation rather than monetary payments. As the partnership grows, it may evolve into a revenue-sharing model where Apple takes a cut from monetized AI chatbot services on its platforms, compensating for potential losses from its search engine deal with Google.

Google is extending its AI Overviews feature to a billion people, but the environmental and financial costs associated with its large language models, like Gemini, are significant. Generative AI like this consumes substantial energy, with estimates showing it's about 30 times more for generating text than extracting it. Big tech's push towards AI could double data centers' energy use by 2026, raising concerns over fossil fuel dependence and the mismatch of renewable energy supply.

The article questions whether the extensive 165-page paper on AI, written by Leopold Aschenbrenner, a 25-year-old ex-OpenAI employee, may stem from youthful naivety or contains significant insights on AI progress. Filip Dimitrovski's piece reflects on his own skepticism towards neural networks and AI, recounting past experiences and his reaction to AI Dungeon. The post also hints at wider implications for AI in society and references a notable dedication within the paper to Ilya Sutskever, a prominent figure in AI.

AMD's MI300X outperforms NVIDIA's H100 in large language model (LLM) inference tasks with 33% higher throughput. Extensive testing by TensorWave and MK1 shows the MI300X's superior performance both offline and online, marking AMD as a strong competitor in AI hardware.

Joor is a journaling app that encourages self-reflection and personal growth. It simplifies the journaling process, allowing users to capture their thoughts seamlessly.

PromptChains AI leverages advanced AI and blockchain technologies to provide innovative solutions for data processing and decentralized applications.

Icons8's Illustration Generator is an AI-powered art maker that can turn text prompts or images into professional illustrations. Ideal for creating visuals for various design needs, it offers unique art styles and ethical use of AI without infringing artists' rights.

TwoShot is a mobile app that offers an innovative way to share moments with friends by combining two photos into one. It encourages users to collaborate on a single picture despite being apart.

Mapify is an AI-powered tool that instantly transforms documents, websites, images, and audio into structured mind maps. It offers a variety of features, such as built-in AI templates, step-by-step creation, and mind map interaction through AI chat. Mapify is designed to break down complex content into clear presentations, assist with brainstorming, planning, learning, and analysis. It originated as Chatmind and is now rebranded as Mapify.

Dream Machine by Luma Labs is an innovative AI-driven solution that turns text and images into realistic and high-quality videos quickly. This tech facilitates speedy iteration and encompasses features like smooth motion, cinematic camera moves, and character consistency. It sets a foundation for a future universal imagination engine.

Joor is an AI-powered journal app designed to transform journaling into a journey of self-discovery. It acts as a personal digital mood ring, tracking and explaining your feelings over time. The app suggests personalized activities to promote well-being, adopts a scoring system for mental health, and even serves as a safe space for self-expression.

Sakana AI explores whether AI can automate AI research, specifically using LLMs in a process termed 'LLM²' for self-improving AI training methods. They combined evolutionary algorithms with LLMs' capabilities, leading to a new optimal loss function, 'DiscoPOP,' outperforming existing methods and offering potential for less resource-intensive discovery.

Hugging Face introduces 'abliteration,' a technique to uncensor any LLM without retraining, by removing built-in refusal mechanisms. This not only alters model responsiveness to all types of prompts but raises ethical questions. The technique aims to neutralize the refusal direction in activations, allowing open response to both harmful and harmless prompts. Maxime Labonne outlines the process with code and deep dives into further fine-tuning methods to recover performance drops post-abliteration.

The blog post discusses a novel application that matches conference attendees and provides personalized talking points to spark meaningful conversations. The AI-driven app aims to reduce the awkwardness of networking by suggesting relevant discussion topics based on attendees' profiles and shared interests.

Stability AI's release of Stable Diffusion 3 Medium (SD3), an AI image generation model, has been met with ridicule due to its poor rendering of human images, marking a significant regression from earlier models. Reddit threads illustrate SD3's failure at generating anatomically correct limbs, a byproduct of heavily censored training data lacking NSFW content.

Mihai Eric reflects on Alexa's failure to dominate the conversational AI market despite Amazon's resources and talent. He highlights the challenges faced within Alexa AI, including technical limitations, a fragmented organizational structure, and a misalignment between product and science. Eric emphasizes the importance of robust infrastructure, leveraging LLMs, and giving research initiatives their needed time.

AI text detectors inaccurately flag professional writers' work, causing job terminations. Freelancer Kimberly Gasuras was wrongly accused by WritersAccess of using AI-generated content due to 'Originality' tool's red flag, despite her protest and evidence. The misuse of AI detectors is a growing problem, impacting lives beyond academia, with a rise in false positives among students and professional writers.

The blog post discusses integrating ChatGTP into iOS, iPadOS, and macOS using Text Workflow and Shortcuts, regardless of official support in iOS 18, providing a solution for devices that aren't compatible with Apple Intelligence.

The 'dead internet theory' posits that the internet is dominated by AI-generated content and bots, which farm engagement. This creates artificial engagement cycles, potentially disguising sophisticated propaganda efforts and influencing public opinion through bot-fueled disinformation. Studies show that bots on social media platforms play a major role in spreading misinformation, with serious implications as social media often serves as the primary news source for many.