The Accelerating Impact of AI in Science and the Road to Full-Scale Deployment

Fifteen months ago, I indicated that AI now has the attention of CEO’s, but no-one could’ve predicted the speed with which AI's transformative power in science would reach new milestones. The 2024 Nobel Prize in Chemistry was awarded for pioneering work using AI to predict protein structures and design novel proteins, revolutionizing research across biology and medicine.

One thing is certain: technology is improving every day, and the pace of innovation is accelerating rapidly. All the predictions we read at the beginning of the year have come to fruition.

• Models have become multimodal, enabling the generation of text, voice, images, and even molecules and proteins in the life sciences. ?The announcement of AI models with enhanced reasoning capabilities has sparked discussions about the future of this technology and its potential impact across various sectors.
• Smaller, more specialized models (BioBERT, ClinicalBERT…) have complimented the larger ones, which have themselves seen significant advancements in the quality and speed of their outputs.
• Additionally, we’ve witnessed the emergence of a combination of Generative AI with traditional AI methods like Machine Learning. Agentic AI architectures are now commonplace, paving the way for the creation of new products and high-quality AI-powered commercial interfaces. During my visit to Biotech X, I noticed that most large organizations had not only internalized the potential of Generative AI, but also developed a deep understanding of its implications. Many had already built architectures and invested in numerous use cases, particularly in Data Engineering and conversational AI.

However, many of these solutions are still at the Proof of Concept (POC) or Minimum Viable Product (MVP) stage, with scalability yet to be achieved. Few solutions are in full production, and there’s not yet a large-scale deployment of such technologies. To ensure high-quality output from Generative AI, it’s crucial to have a robust data layer, with metadata ideally managed within a graph structure. Recent presentations at BiotechX and Phuse EU highlighted the growing importance of graph structures and GraphRAGs (Graph-based Retrieval-Augmented Generation). These approaches enhance the traditional RAG (Retrieval-Augmented Generation) method used for document processing, improving the accuracy and reliability of Generative AI by providing more structured and context-rich data connections..

Vision without action is hallucination, but action without vision is confusion. Joel A. Barker

Generative AI Investment Surges Amid ROI Concerns

Despite skepticism about generative AI’s return on investment, investors continue to place big bets on the technology. In Q3, 2024 alone, venture capitalists poured $3.9 billion into generative AI startups across 206 deals, according to PitchBook. Of that total, US-based companies attracted $2.9 billion spread across 127 deals. Tech giants are also continuing to invest in large AI models to transform healthcare through innovations like diagnostics and smart consultations, but challenges remain in achieving real-world integration and identifying impactful, value-generating applications.

An example of concrete impact is the announcement from Google's CEO revealing that AI systems now generate more than a quarter of new code for its products, with human programmers overseeing the computer-generated contributions. The statement, made during Google's?Q3, 2024 earnings call, shows how AI tools are already having a sizable impact on software development. There is indeed a potential to redefine software engineering, and for Pharma companies stat programming, and generativeAI has a sizeable impact on software engineering.

We're also using AI internally to improve our coding processes, which is boosting productivity and efficiency," Sundar Pichai?said?during the call.

Adding to the momentum, renowned AI researcher Fei-Fei Li, known for creating ImageNet, has raised $230 million for her latest venture,?World Labs, aimed at advancing AI's understanding of the 3D physical world. This startup focuses on "spatial intelligence," which could lead to significant breakthroughs in fields such as AR/VR and robotics, further demonstrating the sector's potential for innovation and investment.

AI Leaders Pivot to User-Focused AI Products for Sustainable Growth

AI leaders like OpenAI, Google, and Anthropic are shifting from model-centric development to building user-focused, revenue-generating products. This move reflects the need for ROI due to the high cost of maintaining advanced models. Indeed, AI companies are estimated to exhaust publicly available data for LLM training at the turn of the decade. It is unclear how they plan to keep up with the rising demand for data, the building blocks of an LLM’s consciousness. Financially, training a single large language model (LLM) costs tens of millions of dollars. For instance, OpenAI CEO Sam Altman said the company spent over $100 million to train GPT-4. Google spent an estimated $191 million on computing to train Gemini Ultra. Having deep pockets and a propensity to splurge helps. Instead of investing solely in larger models, they are focusing on practical products like Google’s Notebook LLM, which turns documents into podcasts, and OpenAI’s SearchGPT for enhanced search experiences.

Understanding the difference between AI tools and AI products is crucial: AI tools enhance workflows by automating tasks, while AI products are standalone solutions providing direct value. The rise of Agentic AI - autonomous systems capable of complex, multi-step tasks, is transforming AI products, enabling them to adapt and learn autonomously, and making user interactions more dynamic and efficient. This shift from simple generative AI to specialized, autonomous capabilities is seen as a game-changer for driving productivity and operational efficiency across industries. If you’re think of building your own, you may be interested in those lessons learnt.

This shift reshapes competition, moving beyond state-of-the-art models to developing impactful applications. The future lies in AI that blends practicality with adaptability, creating solutions that meet user needs and drive sustainable growth.

The Rise of Agentic AI

“The best way to think about an AI agent is as a digital twin of an employee with a clear role.”

To move beyond simple question-and-answer setups, we need agentic reasoning where LLMs can break down complex tasks into smaller components and use tools and iterate through a reasoning process. This approach enables the AI to tackle more intricate queries, like comparing documents or generating comprehensive reports - ultimately offering greater value to users.

AI agents come in various types, including simple reflex agents, model-based reflex agents, goal-based agents, utility-based agents, learning agents, and hierarchical agents, each designed to address different complexities and applications.Their core capabilities - perception, reasoning, learning, and interaction, enable them to streamline complex processes, support decision-making, and optimize operations in enterprise settings. Additionally, understanding key components of agent architecture, such as perception, knowledge representation, decision-making, learning, and communication, is essential for effectively implementing these systems.

AI agents are autonomous programs designed to perceive their environment and take actions to achieve specific goals. They’re now accessible to enterprises of all sizes, thanks to the proliferation of powerful platforms for building and deploying these agents. Among the top platforms,?crewAI, AutoGen, LangChain, Vertex AI Agent Builder, and Cogniflow?stand out for their unique capabilities, ranging from multi-agent collaboration and code generation to no-code interfaces that simplify development for non-technical users.

GraphRAG improves the accuracy and contextual understanding of conversational AI by integrating structured knowledge graphs with LLMs, reducing issues like hallucination and enabling complex question answering. However, challenges such as constructing, scaling, and maintaining knowledge graphs remain significant obstacles to its broader adoption. In text-heavy domains like research and customer support, specialized AI agents often manage tasks such as query execution and data retrieval to enhance workflow efficiency and contextual understanding. Integrating GraphRAG, which merges knowledge graphs with LLMs, can significantly benefit these agents by boosting accuracy, simplifying development, and providing explainable, fact-based outputs. Neo4j, renowned for its powerful graph database technology, plays a crucial role in facilitating this integration, making it a key contributor to the development and application of GraphRAG for GenAI. (See: A Tale of LLMs and Graphs — The GenAI Graph Gathering): However, as with other benefits of GraphRAG, there exist challenges too: the three main challenges with GraphRAG are constructing knowledge graphs due to complex entity extraction, ensuring scalability as graphs grow larger, and maintaining up-to-date information, which can be costly and time-consuming.

In conclusion, there are many opportunities to leverage Agentic AI, as depicted in this recent paper: ClinicalAgent: Clinical Trial Multi-Agent System with Large Language Model-based Reasoning

More deep dive:

• Mastering AI Agents: From Basics to Multi-Agent Systems
• An Argentine AI Architecture and framework is emerging
• Building Multi-Agent RAG Systems with LlamaIndex

Advancing LLM Models: OpenAI’s o1 Model Series, xAI’s Grok-2, and Anthropic’s Claude 3.5 Reshape Tech and Healthcare

While several companies are working on such advancements, OpenAI's introduction of their?o1 model series has brought this topic to the forefront of tech conversations. The o1 models use a "chain of thought" approach, where they think through problems step-by-step before providing a final answer. This process allows them to refine their thinking, try different strategies, and even recognize and correct their own mistakes. This level of sophistication opens up possibilities in Life Sciences and Healthcare where nuanced reasoning is crucial.

xAI has launched Grok-2 and Grok-2 mini, which mark significant advancements in chat, coding, and reasoning, outperforming notable models like Claude 3.5 Sonnet and GPT-4-Turbo. Grok-2's performance excels in real-time reasoning, tool use, and adaptive responses, setting new benchmarks across reading, math, and science. These models, available on ?? for premium users, are part of xAI's strategic expansion, with API in Beta access and plans for improving multimodal features to drive future innovations.

Anthropic's?Claude?3.5 Sonnet?is a powerful AI model that surpasses GPT in various domains, including coding, visual reasoning, and text transcription from imperfect images. It operates at twice the speed of its predecessors, introduces the?Artifacts?tool?for dynamic content interaction, and ensures privacy by not training on user data without permission. Additionally, a new?JavaScript-based?coding?sandbox?allows users to write and visualize code within the AI interface, enhancing interactive coding experiences. Recent research further underscores the capabilities of generative models, such as in the case of?ChessFormer, which outperformed the expert data it was trained on, showcasing generative AI's potential to exceed human expertise. The upgraded?Claude 3.5 Sonnet?introduces broad improvements, particularly excelling in coding tasks. Additionally, the new?Claude 3.5 Haiku?matches the performance of previous top models and operates at high speeds. A groundbreaking capability, “computer use,” is also now in public beta, allowing Claude to interact with interfaces like a human, navigating screens and executing tasks.

Source

Generative AI and Life Sciences - Recent News

• Generative models, when trained on human-generated data, can exhibit "transcendence," surpassing the performance of the experts they were trained on, as demonstrated by a transformer model trained to play chess that outperformed all players in its training dataset.
• Medical Education Transformed: Harvard Medical School (HMS) is revolutionizing medical education with AI integration. A new curriculum features mandatory AI courses and a competitive AI in Medicine PhD track. Practical AI applications in patient documentation allow doctors to focus more on care. HMS prioritizes ethical considerations and critical thinking, preparing future leaders for an AI-enhanced healthcare landscape that balances technology with human-centric care.
• Cohere Launches Multimodal Embed 3: Multilingual & Multi-Modal Search: Embed 3, an industry-leading AI search model, is now multimodal. It is capable of generating embeddings from both text and images. This advancement enables enterprises to unlock real value from their vast amount of data stored in images.
• Generative AI is becoming a crucial tool in healthcare, driving transformation by enhancing care delivery, simplifying administrative tasks, and boosting clinical productivity through innovations such as improved interoperability, AI copilots for decision support, and the integration of multimodal data in research. For example, it enhances interoperability by interpreting unstructured data, allowing seamless data exchange across healthcare systems. In benefits verification, a traditionally manual and time-consuming process between providers and payers, GenAI can automate and speed up verification, reducing delays in patient care. In light of declining EU productivity and global competition, Europe must prioritize innovation in health and biopharmaceuticals by embracing AI technologies to accelerate drug discovery, ensure responsible AI development aligned with EU values, promote secure data sharing, and enhance AI literacy among healthcare professionals, positioning the region as a global leader in healthcare advancements.
• A survey found that one in five GPs in the UK use AI tools like ChatGPT for daily tasks, such as writing patient letters and suggesting diagnoses. While these tools help with administrative efficiency and potential diagnosis and treatment choice, concerns have been raised about patient privacy and the potential for inaccuracies
• ??A Medscape UK survey found that 47% of doctors are hesitant to use generative AI in patient treatment, though 48% would consider it. There’s more support for AI in diagnostics, especially medical imaging, with 57%expressing enthusiasm. Privacy concerns, misinformation, and self-diagnosis remain key issues, and 83% of doctors favor government oversight of AI in healthcare.
• ??Patients seem more bullish, a report by Upvio reveals that 97% of patients are confident AI will enhance patient care, with key benefits like improved efficiency and access to services such as telehealth. 53% of patients trust AI in healthcare, though many practitioners have yet to fully adopt AI-powered tools.
• ??Harrison.ai has introduced a radiology-specific vision language model, Harrison.rad.1, capable of conducting open-ended chats about X-ray images, detecting findings, and generating reports with a focus on clinical safety and accuracy.
• ??Innovaccer's new AI-powered Care Management Copilot, driven by its AI model Sara, automates patient record reviews, visit documentation, and care plans, saving care managers over 10 hours a week and enabling 30-40% more patient engagement.

And last, a hint for your own implementation:

When implementing AI in workflows, industry stakeholders should focus on automating tasks that are redundant and resource-intensive, start with small, targeted objectives, ensure high-quality data for model training, invest in AI development with realistic timelines, and consider the impact on both people and processes to optimize the technology’s potential. Diane Lacroix , VP eClinical Solutions.

Upcoming 2024 Conferences

• 6th GCC eHealth Workforce Development Conference, Dubai, November 6-7, 2024
• DigiHealthDay 2024, Pfarrkirchen, Germany, November 11-15, 2024
• GIANT Health 2024 , London, December 9-10, 2024
• World Summit AI Qatar + Intelligent Health AI, Doha, Dec 10 - 12 2024

Other Informational Newsletters on Healthcare and AI

Stay up-to-date with the latest insights and industry trends by subscribing to these informative newsletters from my esteemed colleagues and professionals in the field:

• Healthcare in the Metaverse by Philippe GERWILL
• DigitalhealthReadingList by Wolfgang Schleifer
• Future Tech Trends by Bernard Marr
• All about Blockchain Adoption by Kamlesh Nagware
• AI Frontier by Steve Nouri
• The Atlas by Generative AI
• Last Month in Digital Health by Gary Monk

AI Basics

DataCamp:?AI for Beginners

Google’s?AI Video Course

Matt Turck:?The AI/ML Landscape

MIT:?AI Basics

Nvidia: Jensen Huang on AI

SimpliLearn:?AI Bootcamp

WSJ:?AI Predictions and Future of Humanity