Today:

OpenAI Introduces MEMORY and New Controls for ChatGPT | Meet your new personalized AI assistant.

OpenAI introduced a new feature for Chat GPT, allowing it to remember previous conversations to enhance future interactions. Users have control over this memory function, able to enable or disable it as desired. Memory improves over time with use, personalizing responses based on past discussions.

Users can manage memories through settings, including deleting specific ones. Additionally, Chat GPT can now provide custom instructions and offers temporary chats for privacy. 

WATCH THE VIDEO ON YOUTUBE

OpenAI Researcher Andrej Karpathy Departs

Andrej Karpathy, a founding member of OpenAI, has left the company, confirmed by a spokesperson. Karpathy, a notable AI researcher, was working on an AI assistant and collaborating closely with the research chief, Bob McGrew. 

While ChatGPT was successful, OpenAI aims to launch software capable of automating intricate computer tasks, like handling expense reports. Karpathy's departure marks a shift in focus towards developing advanced automation solutions.

THE INFORMATION

North Korea and Iran using AI for hacking, Microsoft says

Microsoft warns that nations like Iran and North Korea are employing AI for cyberattacks. The tech giant, along with OpenAI, detected and thwarted attempts to exploit generative AI. These techniques, though early-stage, pose a threat as adversaries leverage large-language models to breach networks and spread disinformation. Examples include North Korea targeting think tanks, Iran using AI for social engineering, and Russia and China exploring military applications. 

Despite OpenAI's claim of limited AI malicious capabilities, cybersecurity experts anticipate evolving threats. Concerns arise regarding the responsible development and security of AI, with calls for more secure models and defensive strategies.

THE GUARDIAN

Apple’s latest prototype AI tool can animate images using text descriptions

Apple unveiled Keyframer, an AI animation tool that adds motion to 2D images using text prompts. The tool, based on OpenAI's GPT4, converts SVG files into CSS animation code, simplifying the process for users without coding experience. It allows for batch production of animations with customizable properties. 

While promising, Keyframer is in early stages, with limited editing capabilities and a focus on web-based animations. Apple acknowledges its current limitations, stating it's not suitable for complex animations seen in movies or video games. Keyframer joins Apple's other generative AI projects, signaling advancements in creative technology.

THE VERGE

Nokia unveils AI assistant for industrial workers

Nokia revealed "MX Workmate," an AI tool aiding industrial workers with real-time alerts on machine issues and productivity tips. It enhances Nokia's communication tech, addressing labor shortages by streamlining data interpretation. Stephane Daeuble, Nokia's Enterprise Solutions Marketing Head, highlighted its value in preventing accidents and optimizing output. The tool, showcased at Mobile World Congress, ensures compliance with operational tech rules. 

While experts raise concerns about AI regulation, Daeuble assures its accuracy, transparency, and moderation. Initial deployment may take 1-1.5 years, pending thorough testing and industry adaptation. This innovation signals Nokia's commitment to enhancing industrial efficiency amid evolving technological landscapes.

REUTERS

Airbnb plans to use AI, including its GamePlanner acquisition, to create the ‘ultimate concierge’

Airbnb aims to revolutionize customer experience using AI, highlighted by its acquisition of GamePlanner.AI. CEO Brian Chesky envisions an adaptive, personalized interface, akin to an "ultimate concierge." Unlike building AI infrastructure, Airbnb focuses on enhancing user interaction with existing AI models. Chesky predicts a significant shift akin to the internet's impact. 

While the plan sounds promising, challenges loom, such as AI's current limitations and potential misinformation. Specific AI products remain undisclosed, but generative AI testing includes writing review summaries. The GamePlanner acquisition, led by Siri's co-founder, hints at accelerating Airbnb's AI integration, promising a transformative future interface.

TECHCRUNCH

🧠RESEARCH

BASE TTS: Lessons from building a billion-parameter Text-to-Speech model on 100K hours of data

BASE TTS is a groundbreaking text-to-speech model boasting 1 billion parameters and trained on 100,000 hours of speech data. It achieves top-notch speech naturalness by converting texts to "speechcodes" then to waveforms. Highlighting emergent abilities for complex sentences with its novel tokenization and compression methods, BASE TTS sets a new standard in speech synthesis, challenging other big models with its innovative approach and state-of-the-art results.

Mixtures of Experts Unlock Parameter Scaling for Deep RL

The paper reveals that by adding Mixture-of-Expert (MoE) modules to value-based networks in deep reinforcement learning (RL), models can scale up more efficiently. Unlike traditional RL approaches, where bigger models often perform worse, this strategy leads to significant performance boosts across various sizes and training setups. Essentially, it paves the way for developing reliable scaling laws in RL, backed by solid empirical evidence.

World Model on Million-Length Video And Language With RingAttention

The paper tackles the challenge of training AI on massive multimodal datasets, combining video and language to understand complex, long-form content. By curating a diverse video and book dataset and using RingAttention for efficient training on sequences up to 1 million tokens, it sets new benchmarks in retrieval tasks and video understanding. This approach, fully open-sourced with a family of 7B parameter models, represents a significant step toward AIs that can grasp human knowledge and the physical world in a more integrated way.

Lumos : Empowering Multimodal LLMs with Scene Text Recognition

Lumos is a pioneering end-to-end multimodal question-answering system enhanced with Scene Text Recognition (STR). This technology extracts text from images and integrates it with a Multimodal Large Language Model (MM-LLM), enriching the AI's understanding capabilities. The challenges of STR quality, latency, and model inference were addressed, offering insights into the system's architecture, design, and efficient modeling techniques. Comprehensive evaluations highlight its high quality and efficiency in text understanding from visual inputs​.

Learning Continuous 3D Words for Text-to-Image Generation

The paper discusses a novel method called Continuous 3D Words, allowing for fine-grained control over image generation attributes, such as illumination direction and object poses, through text-to-image models. It introduces input tokens that users can adjust continuously, akin to sliders, for detailed manipulation of images alongside textual prompts. This approach enables seamless integration of 3D-aware attributes into generated images without adding complexity to the generative process, showcasing practical applications like adjusting time-of-day lighting or object orientations directly from a single mesh and rendering engine​.

HeadStudio: Text to Animatable Head Avatars with 3D Gaussian Splatting

HeadStudio introduces a groundbreaking method for creating lifelike and animated head avatars from text prompts, utilizing 3D Gaussian splatting. This technique, leveraging the FLAME framework, allows for detailed and dynamic avatar animation that can be controlled in real-time with speech or video. It showcases impressive capabilities in rendering realistic avatars at high frame rates and resolutions, pushing the boundaries of digital avatar creation.

🛠️TOP TOOLS

Fora - offers an AI-driven approach for executives to address organizational challenges and nurture relationships.

Openart - AI art generator that allows users to create unique art and images for free, using AI. 

Snowpixel - lets you generate images, videos, music, audio, and 3D objects using text descriptions. 

Aragon - realistic AI headshot generator, enabling professional headshots in minutes.

TTSLabs -  provides a specialized AI text-to-speech service for Twitch streamers, enabling them to customize their text-to-speech experiences with unique voices, sound clips, and more.

📲SOCIAL MEDIA

🗞️MORE NEWS

Slack AI is here, letting you catch up on lengthy threads and unread messages

Slack introduces AI features for Enterprise users, offering thread summaries and channel recaps. Users can catch up on conversations, ask questions, and receive message summaries. The AI integrates with other apps, enhancing productivity. Slack assures data privacy. The feature is currently available in US and UK English. More languages and features are in the pipeline. THE VERGE

Meta Unmasks Hundreds Of AI Spies On Facebook And Instagram Made By Italian Surveillance Dealers

Meta exposed Italian surveillance firms creating fake AI profiles on Facebook and Instagram, targeting journalists and activists. Nearly 1,000 personas used AI-generated images to deceive users into revealing their IP addresses. Meta aims to disrupt surveillance attacks early to prevent further harm, emphasizing the importance of collective action against such threats. FORBES

Google quietly launches internal AI model named 'Goose' to help employees write code faster, leaked documents show

Google quietly launches an internal AI model named 'Goose' to aid employees in coding faster. Trained on 25 years of Google's engineering expertise, Goose helps with writing code using natural language prompts. Part of Google's efficiency drive, it's a collaborative effort between Google Brain, DeepMind, and internal infrastructure teams. BUSINESS INSIDER

Google’s Gemini AI now available on iOS and Android outside of the US

Google's Gemini AI, previously known as Bard, expands beyond the US to iOS and Android globally. Product lead Jack Krawczyk announces English version availability in other countries via a dedicated Android app or iOS toggle within Google app. Rollout in progress; expect Japanese and Korean support next, with more languages and countries to come. THE VERGE

New AI tool helps leverage database of 10 million biology images

A new AI tool, BioCLIP, developed by researchers at Ohio State University, harnesses a massive dataset of over 10 million biology images to enhance machine learning capabilities. The tool, which outperforms existing models by 17-20%, can classify species and promises to unlock biological mysteries, aided by diverse image sources. PHYS ORG

AI tool predicts function of unknown proteins

A new AI tool, DeepGO-SE, predicts unknown protein functions, aiding cellular understanding. Developed by KAUST researchers, it outperforms existing methods, utilizing logical inference similar to Chat-GPT. Applied to poorly understood proteins, it accurately forecasts functions and ranks top in prediction competitions. Promising implications for drug discovery and biotechnology. PHYS ORG

Nvidia is now worth more than Amazon thanks to the AI chip boom

Nvidia outpaces Amazon in market value, driven by soaring demand for AI chips. Its market cap hits $1.78 trillion, surpassing Amazon's $1.75 trillion. Nvidia's focus on server AI chips sees shares surge 246% in a year, while Amazon maintains strength post-earnings. The shift reflects ongoing dynamics among top companies. CNBC

Mindy gets backing from Sequoia to build an email-based AI assistant

Mindy, backed by Sequoia, pioneers email-based AI assistant. Founded by PayPal and YouTube engineers, it utilizes AI for tasks like event searches and data retrieval via email queries. Prioritizing email for its clarity and adaptability, Mindy aims for seamless integration into users' workflow. With funding secured, it eyes user growth and monetization avenues. TECHCRUNCH

Zylon launches to take away the pain of generative AI adoption for SMBs

Zylon, a Madrid-based startup, secures $3.2 million funding to simplify generative AI for SMBs. Co-founded by Iván Martínez Toro and Daniel Gallego Vico, Zylon offers a user-friendly AI workspace for non-tech professionals. Their platform, rooted in open source AI, prioritizes privacy and ease of use, aiming to revolutionize AI adoption. VENTUREBEAT

YC-backed Cambio puts AI bots on the phone to negotiate debt, talk to a bank’s customers

Cambio, a startup backed by Y Combinator, deploys AI bots to handle debt negotiations for consumers, boosting credit scores for 70% of users. Founder Blesson Abraham, a banking veteran, pivoted the company's focus from neobanking to debt management. Now, Cambio's AI aids banks in sales calls, navigating regulatory hurdles with caution.  TECHCRUNCH

AI Retail Analytics Enhance Inventory Management

Retailers are turning to AI retail analytics to streamline inventory management and enhance customer experience. Solutions like Shelfie use cloud-based software and machine learning algorithms to monitor stock, alert staff to low items, and provide valuable insights. With applications beyond retail, such solutions improve business processes and efficiency. INSIGHT.TECH