Introducing Devin, AI Worms Pose New Threats, Google Unveils AI Extraction Technique, & Together AI Raises $106 Million

Cognition introduced their AI software engineer Devin, researchers created a generative AI worm, Google shared how to steal production models, and Together AI is now a unicorn.

Cognition Unveils Devin: The Revolutionary AI Software Engineer

Cognition has emerged from stealth with a groundbreaking announcement: the launch of Devin, an AI software engineer designed to revolutionize the coding landscape.

Funded by Peter Thiel's Founders Fund and supported by Elad Gil and Tony Xu, Devin distinguishes itself from existing coding assistants by autonomously managing entire development projects—from code writing, bug fixing, to execution. Unlike Github Copilot and similar tools that offer piecemeal assistance, Devin promises a comprehensive solution, capable of undertaking complex tasks and making thousands of decisions within a sandboxed environment.

Through a chatbot-style interface, users can initiate projects with natural language prompts, enabling Devin to plan, execute, and communicate progress in real-time, offering a hands-off approach to software development. While access is currently limited to select customers, Devin's capabilities hint at a future where AI could dominate software development, offering a blend of autonomy and collaboration that could redefine creativity and efficiency in engineering tasks.

This engineer has a solid thread of using Devin so far. The feedback has been excellent, with the biggest drawbacks being the speed and lack of collaboration.

GenAI Worms: A New Frontier in Cyberattacks

Researchers created "Morris II," a generative AI worm capable of autonomously spreading from one system to another, potentially hijacking generative AI systems like OpenAI's ChatGPT and Google's Gemini to steal data or deploy malware.

This AI worm exploits generative AI's capabilities to understand and act upon complex prompts, demonstrating a novel form of cyberattack by using "adversarial self-replicating prompts." These prompts can compel an AI system to produce further instructions that propagate the worm's malicious activities.

This research not only highlights the worm's ability to breach security measures of leading AI models while also warning about the potential vulnerabilities within the rapidly evolving AI ecosystem.

Google Unveils AI Extraction Technique

Google researchers shared how they can partially extract information from leading AI models, such as OpenAI's ChatGPT and Google's PaLM-2, by intelligently querying their public APIs.

This method, detailed in their recently published paper, successfully recovers the final layer of these models for a cost ranging between a few hundred to a few thousand dollars, depending on the layer's size. For under $20 USD, their attack extracted the entire projection matrix of OpenAI’s Ada and Babbage language models. They estimate it would cost under $2k in queries to recover GPT-3.5 turbo’s entire projection matrix.

The simplicity of the attack, based on linear algebra and transformer architecture insights, underscores the critical importance of API design in safeguarding AI technologies against emerging vulnerabilities.

Together AI Hits $1.25 Billion Valuation in Latest Salesforce-Led Round

Salesforce Ventures led a $106 million round in Together AI, valuing them at $1.25 billion.

Together AI plans on broadening their offerings beyond GPUs to become a comprehensive platform for developers looking at either building or tailoring open source AI models. With this latest round, the startup plans on augmenting their computing power and recruit additional researchers to develop new model architectures, signaling its ambition to become a go-to hub for deploying scalable AI applications.