## **Solving Domain-Specific Problems (cybersecurity, healthcare), Google, 2025** ![[Google_Solving Domain-Specific Problems Using LLMs.pdf]] This whitepaper by Google explores the transformative role of large language models (LLMs) when fine-tuned to solve domain-specific challenges, focusing on two highly specialized areas: cybersecurity and healthcare. It presents the development and implementation of SecLM and MedLM - models built to address the complexity, sensitivity, and evolving nature of problems within their respective fields. The document emphasizes that while general-purpose LLMs provide a foundational base, domain-specific models offer superior performance by incorporating context-aware, authoritative, and high-fidelity reasoning. **Key Insights** - **Cybersecurity (SecLM)** - The cybersecurity landscape is hindered by ever-changing threats, operational fatigue, and a shortage of skilled professionals. - SecLM addresses these issues through a multi-layered architecture combining LLMs, traditional ML, Retrieval-Augmented Generation (RAG), and tool use. - Security tasks include threat identification, alert triaging, reverse engineering of malware, access policy optimization, and automated planning. - The model is trained using domain-relevant corpora while carefully managing sensitive data and enabling parameter-efficient tuning (PET) for user-specific adaptations. - The SecLM API acts as a central intelligence system capable of integrating real-time data, reasoning over multiple tools, and performing complex, multi-step analysis. - **Healthcare (MedLM / Med-PaLM 2)** - Medical QA tasks require not only data comprehension but also nuanced reasoning and empathy. - Med-PaLM 2, derived from Google's PaLM 2, achieved expert-level accuracy (86.5%) on USMLE-style questions, marking a major leap in AI’s capability in medical domains. - The model's design prioritizes not just diagnostic accuracy but also safety, helpfulness, health equity, and factual soundness. - Evaluation encompasses both quantitative (accuracy, similarity metrics) and qualitative (clinician-led review) dimensions. - MedLM integrates rigorous retrospective, observational, and interventional evaluation protocols to ensure safe real-world deployment. - Med-PaLM is evolving to support multimodal data (text, images, sensors), extending its utility across broader medical use cases like intake processing, triage, and scientific discovery. **Actionable Takeaways** - **For Cybersecurity Professionals:** - Integrate SecLM to reduce manual toil, support novice analysts, and enable scalable threat analysis. - Utilize SecLM's dynamic planning and tool orchestration to automate complex threat intelligence workflows. - Use PET and RAG to align models with internal environments and keep outputs grounded in fresh, authoritative data. - **For Healthcare Organizations:** - Adopt MedLM for clinical support tasks ranging from patient triage to documentation and decision support. - Prioritize responsible deployment by following structured evaluation protocols—retrospective, observational, and interventional. - Leverage fine-tuning and ensemble techniques (e.g., CoT, self-consistency, ER) to improve accuracy and reliability in sensitive use cases. - Plan for future adoption of multimodal and flexible conversational systems to enrich healthcare workflows. **Notable Quotes** - *“The combination of LLMs and human expertise has the potential to revolutionize the field of cybersecurity, achieving superior results with less effort.”* - *“Medicine revolves around caring for people, and needs to be human-centric… an ambitious goal would be a flexible AI system that can interact with people and assist in many different scenarios.”* - *“Med-PaLM was the first AI model to exceed the passing mark on USMLE-style questions… Med-PaLM 2 reached expert-level performance with 86.5% accuracy.”* - *“Evaluation extends to qualitative assessment of factuality, expert knowledge, helpfulness, equity, and harm—beyond just accuracy.”* - *“Technology alone is not enough. Collaboration with the clinical community and careful multi-step evaluations are crucial for successful application of LLMs in healthcare.”* This paper presents a compelling case for the future of AI as domain-specialized, context-aware, and collaborative—combining advanced reasoning with structured, real-world deployment strategies. ## **The AI in Banking – Best Practices Playbook, Euromoney, 2025** ![[Euromoney_The AI in Banking.pdf]] This report from Euromoney provides a deep dive into the evolving landscape of generative AI (gen AI) in the banking sector. Drawing insights from over 30 in-depth interviews with senior leaders at global and national banks, as well as AI-focused vendors and fintechs, the playbook outlines best practices, strategic frameworks, and lessons learned in implementing AI. It emphasizes that while AI offers transformational potential, particularly through large language models (LLMs) and emerging agentic systems, its integration must be deliberate, flexible, and rooted in strong governance. The report is structured around three main pillars: Preparing the Organisation, Using Vendors Strategically, and Embracing the Future. **Key Insights** - **Organisational Readiness**: Successful AI adoption starts with flexible strategies that accommodate the rapid pace of AI evolution. Banks should adopt an LLM-agnostic approach, centralize AI strategy while empowering divisions, and upskill the entire organization - not just technologists. - **Strategic Use of Vendors**: Choosing the right models and vendors involves balancing cost, performance, security, and adaptability. Building proprietary AI platforms or LLM wrappers allows for easier switching between providers and greater internal control, especially critical under regulatory scrutiny. - **Embracing Gen AI & Agentic Systems**: Low-risk internal use cases (e.g., call center support, marketing content, code generation) yield measurable productivity gains. While customer-facing applications are more complex and risk-prone, banks are cautiously experimenting, especially with agentic AI that can perform autonomous actions like dispute resolution and appointment scheduling. - **Cultural Shift & Talent Development**: Democratizing AI access across staff levels is key. Banks are rolling out tools like Microsoft Copilot and creating internal training ecosystems with mentors and reverse-mentoring. Programs such as UBS's Copilot rollout and DBS’s “Data Chapter” showcase efforts to embed AI literacy and talent retention. - **Customization over Scale**: Large LLMs are not always optimal. Smaller, tailored models can offer better control, lower costs, and improved compliance. Partnerships with regional LLM developers (e.g., Mistral, Cohere) show the value of local, domain-specific models. **Actionable Takeaways** - Build a central AI function with a clear mandate for strategy, cost control, governance, and training. - Maintain model and vendor flexibility via LLM-agnostic platforms that decouple the interface from underlying models. - Start with internal, low-risk use cases to build expertise and demonstrate ROI before expanding to customer-facing or agentic applications. - Invest in staff training at all levels, with structured programs and cultural incentives for AI adoption. - Use AI development as an opportunity to rethink legacy systems, improve data governance, and align with regulatory expectations. - For system-critical institutions, consider long-term investment in proprietary or regional foundational models. - Evaluate every AI initiative on strategic alignment, business value, and scalability across divisions. **Notable Quotes** - *“Everything you build in AI should allow you to be agnostic as to which LLM you use, and as to the vendor.” – Christoph Rabenseifner, Deutsche Bank* - *“You need to become very agile...but you also need to be decisive.” – Dirk Marzluf, Banco Santander* - *“There are business strategies, and there are places where AI can enable the strategy.” – Teresa Heitsenrether, JPMorgan Chase* - *“The big thing about generative AI is how you allow for everybody in the organisation to be able to use it.” – Nimish Panchmatia, DBS* - *“We bring security and trust behind the interface.” – Gavin Munroe, Commonwealth Bank of Australia* - *“We’re in the early days for its adoption in banking.” – Vitor Guarino Olivier, Nubank* - *“If the transformation is enduring, what is the rush to get it done? Let’s get it done right.” – Prem Natarajan, Capital One*