Pssst š Weāre launching the HackerPulse platform soon. Itās your go-to AI-powered platform to help you showcase your skills. Wanna be a beta tester? Sign up!
Hey there! Join us for another edition of AI Fridays, showcasing the recent breakthroughs and cutting-edge advancements in the funky world of AI.Ā
ā”ļøAs the week draws to a close, we highlight OpenAIās disruption of five state-affiliated malicious actors, AI empathetic assistant BUD-E, Deep Fake 3D Avatars, and a few other stories.
OpenAI Disrupts State-Affiliated Malicious Actors (š Read Paper)
OpenAI, in collaboration with Microsoft Threat Intelligence, announces the disruption of five state-affiliated actors exploiting AI services for malicious cyber activities. While OpenAI emphasizes its commitment to building AI tools that improve lives, it acknowledges the risk of malicious abuse and outlines measures to counter such threats.
OpenAI terminated accounts associated with 5 state-affiliated threat actors, affirming its dedication to ethical AI use. The models provided by OpenAI offer limited, incremental capabilities for malicious cybersecurity tasks.
The five state-affiliated actors were: two China-affiliated threat actors known as Charcoal Typhoon and Salmon Typhoon; the Iran-affiliated threat actor known as Crimson Sandstorm; the North Korea-affiliated actor known as Emerald Sleet; and the Russia-affiliated actor known as Forest Blizzard.These actors sought to use AI services for tasks such as researching cybersecurity tools, intelligence agencies and regional threat actors along with common ways processes could be hidden on a system; debugging code and generating scripts; creating content likely for use in phishing campaigns; researching common ways malware could evade detection, and more.
OpenAI adopts a multi-pronged approach to combat malicious state-affiliated actors' use of its platform. This approach includes monitoring and disrupting malicious actors, collaborating with the AI ecosystem for information exchange, iterating on safety mitigations, and promoting public transparency. The organization's proactive measures aim to foster awareness, preparedness, and collective defense against evolving adversaries.
ImplicitDeepfake1 3D Avatars (š Read Paper)
ImplicitDeepfake1 seamlessly merges deepfake technologies with Neural Radiance Fields (NeRFs) and Gaussian Splatting (GS) for next-gen 3D avatar creation and gaming. This involves modifying training images separately using deepfake and then training NeRF and GS on the modified faces, showcasing a relatively simple yet effective strategy.
Numerous emerging deep-learning techniques, including Neural Radiance Fields (NeRFs) and Gaussian Splatting (GS), significantly impact computer graphics.
NeRFs encode object shape and color in neural network weights, generating novel views from a few known images with camera positions. GS accelerates training and inference without compromising rendering quality, using Gaussian distributions to encode object characteristics.
Applications of NeRFs and GS extend to spatial computing and various domains, highlighting their versatile impact.
Despite controversy, deepfake technology, generating AI videos closely mimicking authentic footage, holds potential for next-gen solutions in avatar creation and gaming of desirable quality.
Laionās BUD-E: AI Assistant With Natural Voice and Empathy (š Read Paper)
LAION launches BUD-E, an AI assistant designed to respond to user requests in real time with natural voice, empathy & emotional intelligence. BUD-E stands for Buddy for Understanding and Digital Empathy.Ā
LAION collaborates with the ELLIS Institute TĆ¼bingen, Collabora, and the TĆ¼bingen AI Center to develop AI open voice assistant BUD-E.Ā
BUD-E aims for real-time responses with natural voices, empathy, and emotional intelligence, maintaining long-term context in conversations, and handling multi-speaker interactions locally on consumer hardware.
BUD-E runs on consumer devices, with latencies of 300 to 500 ms, a fast response to user requests. The roadmap includes reducing latency, minimizing system requirements, and enhancing naturalness of speech and responses.Ā
Open call by LAION for contributions to tackle immediate challenges, including latency reduction, naturalness enhancement, conversation memory, and extending functionality, aiming to build a collaborative future in conversational AI.
Stable Cascade by Stability AI (š Read Paper)
Stability AI unveils Stable Cascade, a revolutionary text-to-image model built upon the WĆ¼rstchen architecture, now available in research preview. This non-commercially licensed model boasts a three-stage approach, promising ease of training and fine-tuning on consumer hardware.
Stable AIās text to image model introduces a three-stage approach for text-to-image generation, setting new standards for quality, flexibility, fine-tuning, and efficiency. Itās very easy to train and finetune on consumer hardware, ensuring an accessible experience for users.
Alongside providing checkpoints and inference scripts, the release includes finetuning, ControlNet, and LoRA training scripts available on the Stability GitHub page.
Comprising three models, Stage A, Stage B, and Stage C, Stable Cascade achieves remarkable image outputs through hierarchical compression. Stages A and B efficiently compress images, while Stage C generates small 24 x 24 latents from text prompts, offering versatility in image creation.
The model comes with varying parameters for each stage, with Stage C offering 1 billion and 3.6 billion parameter versions. While Stage B boasts 700 million and 1.5 billion parameter options, Stage A remains fixed with 20 million parameters. Optimal results are achieved by selecting larger parameter variants for each stage.
New Memory for ChatGPT (š Read Paper)
OpenAI is testing a new ChatGPT memory feature, aiming to enhance its conversational capabilities by reducing repetition and increasing the quality of responses. With ChatGPTās new capability to remember and recall specific details from past chats, users are now in control of a personalized and efficient dialogue.
Users are in control of ChatGPT's memory. They can explicitly tell it to remember something, ask it what it remembers, and tell it to forget conversationally or through settings. They can also turn it off entirely.Ā
The memory feature is being introduced to a select group of ChatGPT free and Plus users for testing, with plans for a broader rollout in the near future.
ChatGPT's memory improves with increased usage, becoming more context-aware and adaptive over time.
The introduction of memory raises privacy and safety considerations, prompting proactive steps to assess and mitigate biases. However, ChatGPT is designed to avoid proactively remembering sensitive information, unless explicitly requested by the user.
For Enterprise and Team users, memory proves invaluable in streamlining work processes. ChatGPT can learn and apply style preferences, coding languages, and frameworks, saving time and enhancing productivity in professional tasks.
Memory can be turned off at any time by Enterprise account owners, ensuring control over organizational data.
GPTs will also have their own distinct memory feature, enabling builders to enhance the capabilities of their models. Memories are not shared with builders, ensuring user privacy. Each GPT will have its own memory, requiring users to repeat specific details when interacting with different GPTs.Ā
Hope you enjoyed this weekās adventures in AI. Refer a friend ā the more the merrier! š