- AI Breakfast
- Posts
- AI Is Building Highly Effective Antibodies That Humans Can’t Even Imagine
AI Is Building Highly Effective Antibodies That Humans Can’t Even Imagine
Good morning. It’s Friday, August 11th.
Did you know: In the 1960s, there was an early AI program called "ELIZA" developed by Joseph Weizenbaum at MIT. It functioned as a rudimentary chatbot that mimicked a psychotherapist. While its responses were basic, some users believed they were conversing with a real human.
In today’s email:
AI Model Launches & Developments
AI in Healthcare & Biology
Cybersecurity & AI
AI & Environment
Controversies and Incidents
5 New AI Tools
Latest AI Research Papers
You read. We listen. Let us know what you think of this edition by replying to this email, or DM us on Twitter.
Today’s edition is brought to you by:
Imagine if making videos was as simple as writing emails.
Enter DeepReel – the game-changer in video creation. With DeepReel, you can bring your ideas to life by creating a lifelike avatar from just text. No more time-consuming and expensive processes, just pure creativity!
Join the waitlist today to be among the first to experience the magic of DeepReel.
As an AI Breakfast reader, you'll even get an exclusive 50% off on your avatar creation cost. Unleash your potential and unlock the power of Gen-AI with DeepReel.
![](https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/c6e8c55b-385b-449a-b263-ab02f7f3aab3/Screenshot_2023-05-26_at_7.38.27_AM.png)
Today’s trending AI news stories
AI Model Developments & Features
Anthropic launches improved version of its entry-level LLM: AI startup Anthropic, co-founded by former OpenAI executives, has introduced an updated iteration of its rapid and cost-effective text-generating model, accessible via an API named Claude Instant 1.2 capitalizes the strengths of Anthropic’s flagship model, Claude 2, exhibiting substantial advancements in domains like mathematics, coding, reasoning, and safety. Notably, the version 1.2 boasts improved quote extraction, multilingual capabilities, question answering, and enhanced resistance to generating nonsensical content and security breaches. This development aims to compete with other entry-level AI text-generating models in the market..
GPT-4 dominates other LLMs in "real-world pragmatic missions," study finds: GPT-4 asserts its dominance in practical domains, as evidenced by the AgentBench benchmark. A comprehensive study involving 25 models, conducted by Tsinghua University, Ohio State University, and UC Berkeley, spotlights GPT-4’s exceptional performance across various sectors, including operating systems, databases, and creativity. Achieving an impressive score of 4.41, GPT-4 outshines its commercial and open-source counterparts.
Meet Jen, Futureverse's New Text-to-Music AI Generator: Futureverse has introduced Jen-1, its upcoming text-to-music generator. In a bid to surpass competitors such as Google’s MusicLM, Jen-1 seeks to produce more intricate and sonically rich musical compositions. Scheduled for an early 2024 release, Jen-1 is poised to generate compositions up to three minutes long, while also assisting music producers with incomplete works through its ‘continuation’ and ‘inpainting’ features.
AI in Healthcare & Biology
AI Is Building Highly Effective Antibodies That Humans Can’t Even Imagine: LabGenius, an AI-driven biotech firm, is transforming antibody engineering using automation and machine learning. Conventionally, humans laboriously design synthetic antibodies to combat diseases, searching through myriad amino acid combinations. LabGenius automates this process, employing a machine-learning algorithm that rapidly explores potential antibodies and their effectiveness. By automating antibody design, production, and testing, the firm reduces the time to develop new immunotherapies from years to just six weeks. This revolutionary approach discovers unique antibodies, potentially offering safer and more effective treatments, thus revolutionizing medical research.
New high-tech microscope using AI successfully detects malaria in returning travelers: Researchers have evaluated a fully automated system that combines AI detection software and an automated microscope for diagnosing malaria. The system achieved an 88% diagnostic accuracy rate relative to human microscopists, indicating its potential clinical usefulness. The AI-microscope system correctly identified 99 out of 113 malaria-positive blood samples. While the system offers benefits such as reduced workload for microscopists and reproducible results, it still falsely identified some samples as positive, demonstrating that further refinement is needed for clinical application.
Cybersecurity & AI
The White House's 'AI Cyber Challenge' aims to crowdsource national security solutions: The US government is initiating an AI Cyber Challenge (AIxCC) in collaboration with DARPA, Anthropic, Google, Microsoft, and OpenAI. The competition aims to build AI systems capable of proactively identifying and rectifying software vulnerabilities to bolster cybersecurity. The two-year challenge offers a total prize pool of nearly $20 million, encouraging participants to develop AI solutions that can rapidly defend critical infrastructure codes from attacks. The competition aims to demonstrate AI”s potential to enhance society’s defense against cybersecurity threats.
Microsoft explains how it's Red Teaming GPT-4 and other AI models: Microsoft utilizes “red teaming” to identify vulnerabilities and enhance security in AI systems like GPT-4 and applications such as Bing Chat. Red teaming involves independent groups probing systems to find weaknesses, aiming to ensure system security. In AI, this process tests for vulnerabilities and unexpected behavior to improve the model’s robustness. Microsoft employs a two-tier red teaming approach for foundational models and application levels, considering risks arising from both malicious intent and normal model use. Red teaming in AI presents unique challenges due to probabilistic outcomes and rapidly evolving AI systems, necessitating a layered defense strategy.
AI Is Generating Security Risks Faster Than Companies Can Keep Up: As businesses embrace tools like Microsoft Copilot, concerns mount over managing potential cybersecurity threats. Tech leaders wrestle with comprehending these evolving risks, notably with the rapid advancement of generative AI. The push for a “software bill of material” similar to supply-chain inventories gains traction, aimed at ensuring security and quick response to vulnerabilities. Emerging startups like Protect Ai are stepping in to tackle the complexities, while CIOs navigate the surge of AI-driven features, grappling with code security, and AI’s impact on software development.
Legions of DEF CON hackers will attack generative AI models: DEF CON’s 31st annual event will witness throngs of hackers engaging in the Generative Red Team (GRT) Challenge, a comprehensive endeavor aimed within top-tier large language models. Leading Chinese tech firms, including Alibaba, Baidu, ByteDance, and Tencent, will assess models from Anthropic, Cohere, Google, Hugging Face, Meta, Nvidia, OpenAI, and Stability. This pioneering public foray into generative AI red teaming aligns with the Biden-Harris administration’s AI Bill of Rights and NIST Risk Management Framework.
AI & Environment
Google AI is helping airlines mitigate the climate impact of contrails: Google Research, in collaboration with American Airlines and Breakthrough Energy, utilized AI and satellite imagery to develop contrail forecast maps that help pilots avoid flight paths likely to create contrails. Contrails, created when planes fly through humid layers, contribute significantly to aviation’s global warming impact. Through AI predictions and open-source contrail models, pilots reduced contrail formation by 54% during test flights. While contrail avoidance may slightly increase fuel consumption, it remains a cost-effective solution to mitigate aviation’s climate impact. Further research aims to automate contrail avoidance and improve verification methods using AI.
AI Is Now Protecting The Planet From Asteroids: A groundbreaking AI algorithm, HelioLinc3D, has revealed a potentially perilous asteroid, named 2022 SF289, escaping human observation. The Vera Rubin Observatory in Northern Chile harnessed HelioLinc3D’s power, designed for synoptic astronomical surveys. The 600-foot asteroid, set to pass within 140,000 miles of Earth, underscores the algorithm’s capability to enhance early detection. HelioLinc3D’s efficiency and ability to function amid traditional systems could reshape celestial monitoring, offering improved insights into space and ensuring Earth's safety from potential cosmic threats.
Controversies and Incidents
Supermarket AI meal planner app suggests recipe that would create chlorine gas: A New Zealand supermarket’s AI meal planner app, created to suggest recipes for using up leftovers, has generated unusual and potentially dangerous dishes. The app, which was launched by Pak ‘n’ Save, originally offered recipes like “Oreo vegetable stir-fry.” However, when users input a wider range of household items, it suggested concoctions like “aromatic water mix” that would create chlorine gas, “bleach-infused rice surprise, and “methanol bliss.” The supermarket stated that it would fine-tune the controls of the bot to ensure safety and usefulness and that users should exercise their own judgment when relying on the app’s recipes.
🎧 Did you know AI Breakfast has a podcast read by a human? Join AI Breakfast team member Luke (an actual AI researcher!) as he breaks down the week’s AI news, tools, and research: Listen here
![](https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/8dc7ff85-f73a-4795-a993-3e0092745d40/Screenshot_2023-05-26_at_7.37.43_AM.png)
5 new AI-powered tools from around the web
PI.EXCHANGE This AI & Analytics Engine enables creating machine learning models to predict future events and make informed decisions, without coding skills. It offers smart data protection, model recommendations, insights, and extensive integrations aiming to democratize access to machine learning.
Recast transforms your reading list into bite-sized audio conversations using AI. Enjoy short podcasts of articles with co-hosts providing real dialogue explanations. Convert any article or explore community recasts. Learn on the go, and save screen time. Experience “reading” without reading with Recast.
Pixcap is the next generation of graphic design powered by 3D and AI. Craft unique designs using 10,000+ 3D elements, then let the AI stylist generate captivating variations.
Spinach.io is your AI Scrum Master. Gain an always-on-time teammate who excels in meeting notes, action items, and ticket suggestions. Seamlessly integrated with Slack, Zoom, Google Meet, and more.
Taylor AI simplifies open-source LLM fine-tuning, freeing developers from Python complexities and model upkeep. Empower experimentation, building better models with ease. Explore, customize, and take ownership of the latest in open-source models.
![](https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/6890d6ed-d5ac-4c00-b39e-028f834d7efd/Screenshot_2023-05-26_at_7.38.05_AM.png)
arXiv is a free online library where scientists share their research papers before they are published. Here are the top AI papers for today.
This paper presents the Slimbot Challenge on Embodied AI, where university teams compete to create versatile conversational AI agents in a simulated environment. These AI assistants operate in a Unity-powered office/lab setting, executing tasks through voice commands. Users interact with the Slimbots, experiencing real-time interactions with dynamic responses and visual feedback. This pioneering challenge pushes the boundaries of embodied AI, merging game mechanics and AI development for an immersive experience. The initiative marks a significant step towards AI agents navigating and acting in the physical world.
AudioDLM 2, a collaboration between the university of Surrey’s CVSSP and ByteDance’s SAMI, introduces a transformative audio generation framework. Unifying speech, music, and sound effects under a shared “language of audio” (LOA), this approach employs AudioMAE for self-supervised pre-training. LOA acts as an intermediary for generation, facilitating contextual learning and reusing AudioMAE and latent diffusion models. Impressively, AudioDLM 2 excels in text-to-audio, text-to-music, and text-to-speech benchmarks, offering competitive or leading outcomes.
In a breakthrough for robotics, a revolutionary system known as FAn (Follow Anything) has emerged, showcasing the power of open-set detection, tracking, and real-time following. Developed by a team led by Alaa Maalouf at MIT’s CSAIL and Harvard University’s SEAS, FAn employs cutting-edge multimodal models like CLIP, DINO, and SAM, enabling seamless detection and tracking of objects specified via text, images, or clicks. The system’s adaptability, overcoming occlusion challenges, and its autonomous re-detection capabilities mark significant strides in object-following deployment on micro aerial vehicle, achieving impressive throughput, promises transformative applications across sectors.
The paper presents JEN-1, a universal music generation model for text-guided music creation. Music generation, especially from text prompts, poses challenges due to intricate structures and high sampling rates. JEN-1 employs models, that combine autoregressive and non-autoregressive training, for tasks like text-guided music generation, inpainting, and continuation. It achieves high-quality results by directly modeling audio waveforms. The model’s bidirectional and unidirectional modes ensure comprehensive context capture. JEN-1 outperforms existing methods in text-music alignment and quality, offering an efficient and versatile framework for generating high-quality music aligned with textual prompts and melodic structures.
The paper introduces FocalFormer3D, a 3D object detection model designed to address the issue of false negatives in autonomous driving scenarios. False negatives can lead to dangerous situations by missing predictions of pedestrians, vehicles, or obstacles. FocalFormer3D employs the Hard Instance Probing (HIP) strategy, which identifies false negative samples in a multi-stage manner, allowing the model to focus on challenging instances. The model includes a multi-stage heatmap encoder to generate high-recall object candidates and a box-level deformable decoder for refinement. Experimental results demonstrate FocalFormer3D’s superior performance in reducing false negatives and achieving state-of-the-art results on 3D LiDAR detection and tracking benchmarks.
![](https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/7c9f1f61-d1c1-474e-ba34-71316436330a/paperclipcontent.png)
Thank you for reading today’s edition.
![](https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/fe4085f6-474e-4723-84ae-e082c991cc78/Screenshot_2023-05-17_at_8.39.57_AM.png)
Your feedback is valuable.
Respond to this email and tell us how you think we could add more value to this newsletter.