DeepMind takes a novel approach to AI optimization

Pivot 5: 5 stories. 5 minutes a day. 5 days a week.

1. DeepMind takes a novel approach to AI optimization

DeepMind has recently published a paper introducing a groundbreaking method called Optimization by PROmpting (OPRO). Unlike traditional optimization algorithms that rely on mathematical definitions, OPRO uses large language models as optimizers and defines the optimization task in natural language. This approach allows for a more intuitive and adaptable optimization process.

The adaptability of OPRO is one of its most compelling features. By simply modifying the problem description or adding specific instructions, the LLM can be guided to solve a wide array of problems. Researchers found that LLMs could generate effective solutions through prompting alone, sometimes even surpassing the performance of expert-designed heuristic algorithms.

The true potential of OPRO lies in its ability to optimize LLM prompts for maximum accuracy. This is particularly relevant for models like OpenAI's ChatGPT and Google's PaLM. The method provides a systematic way to explore the vast space of possible LLM prompts, thereby finding the one that works best for a specific type of problem. While the real-world applications of OPRO are yet to be seen, this research marks a significant step forward in our understanding of how LLMs work.

Read the full story here 

2. LastMile AI secures $10M to streamline AI app development for software engineers

LastMile AI, a platform aimed at helping software engineers integrate generative AI models into their applications, has successfully raised $10 million in a seed funding round. Led by Gradient, Google's AI-focused venture fund, the funding will be used to expand LastMile's product offerings and team. The startup aims to democratize generative AI by simplifying the tooling and workflows, thus reducing the need for deep technical expertise in AI.

The startup's co-founder and CEO, Sarmad Qadri, highlights the challenges businesses face in adopting AI. According to a recent S&P Global survey, around half of IT leaders believe their organizations are not ready to implement AI. LastMile offers a suite of developer tools that includes AI Workbooks for experimenting with different models, AI Workflows for building complex workflows, and AI Templates for reusable development setups. These tools aim to make the AI developer journey less fragmented and more accessible.

Despite the challenges, business leaders are optimistic about AI adoption. A 2022 Gartner survey revealed that 80% of executives believe automation can be applied to any business decision. LastMile aims to solve the "last mile" issues that prevent businesses from fully incorporating AI into their operations. With the market for AI model operations expected to grow to $16.61 billion by 2030, LastMile is well-positioned to capitalize on this burgeoning sector.

Read the full story here 

3. Revolutionizing 3D model customization with Style2Fab

MIT researchers have developed Style2Fab, an AI-driven tool designed to simplify the customization of 3D models. The tool addresses a significant hurdle for amateur artisans who often struggle with CAD software. Style2Fab employs generative AI algorithms that allow users to add custom design elements using natural language prompts, thereby democratizing the 3D design process.

The tool's capabilities extend beyond mere customization. Style2Fab is engineered to partition 3D models into functional and aesthetic segments automatically. This feature is particularly beneficial in the emerging field of medical making. For instance, a user could customize a thumb splint's appearance without compromising its functionality, making the tool invaluable for DIY assistive technology.

The research behind Style2Fab is as intriguing as the tool itself. It employs machine learning to analyze a model's topology and divides it into segments based on functionality. A "human-in-the-loop" approach ensures that the AI's recommendations align with human expertise. The tool has already shown promise in studies involving makers with varying levels of experience, proving its versatility and potential for widespread adoption.

Read the full story here

4. Microsoft's EvoDiff provides a game-changer in protein design through AI

Microsoft has unveiled EvoDiff, a groundbreaking AI framework designed to generate high-fidelity proteins without requiring detailed structural information. Traditional protein design is a labor-intensive process, involving the creation of a protein structure and then finding a sequence of amino acids that will fold into that structure. EvoDiff eliminates this cumbersome step, streamlining the design process.

The framework is not just a theoretical construct; it has practical applications. According to Microsoft senior researcher Kevin Yang, EvoDiff could be employed to create enzymes for new drug delivery methods and industrial chemical reactions. The model is trained on a 640-million parameter set sourced from various species and functional classes of proteins, making it incredibly versatile.

While EvoDiff shows promise, it's worth noting that the research has not yet been peer-reviewed. The team acknowledges that more work is needed to scale the framework for commercial use. Nevertheless, EvoDiff represents a significant leap in protein engineering, offering a new paradigm that focuses on sequence-first design. The next steps involve lab testing to validate the generated proteins, setting the stage for future advancements in the field.

Read the full story here 

5. AI voice Cloning becomes a tool for sophisticated scams

AI voice cloning technology, once a marvel of innovation, is now a growing concern as it becomes a tool for sophisticated scams. The technology can capture the uniqueness of a person's voice, making it increasingly difficult to distinguish between genuine and fraudulent communications. Chris Carlis, a security consultant, warns that scams involving voice cloning can sound exactly like a familiar voice, making them harder to detect.

Federal agencies like the FBI and the Federal Trade Commission have issued warnings about the capabilities of AI in facilitating scams. According to the FBI's internet crime report, Americans lost over $10 billion last year due to internet scams involving AI and voice cloning. The technology is not just limited to voice mails and audio recordings; it's advancing towards real-time voice cloning, adding another layer of complexity to the issue.

To protect oneself from these advanced scams, experts recommend double-checking any suspicious communications and ensuring that social media accounts are set to private to prevent voice theft. As AI technology continues to evolve, so does the sophistication of scams, making vigilance and awareness more crucial than ever.

Read the full story here 

Listen to the daily Pivot 5 podcast on Spotify, Apple Podcasts, and more