Hallucinations: How GSK is tackling one of AI’s biggest hurdles in the name of drug development

GSK is implementing advanced computational strategies to address AI hallucinations – when AI models generate incorrect information – in their drug development and healthcare applications.

The core challenge: AI hallucinations pose significant risks in healthcare settings, where accuracy and reliability are crucial for patient safety and drug development outcomes.

• Healthcare applications of AI require exceptional precision as errors can have serious consequences for patient care and drug development
• GSK applies generative AI to multiple critical tasks including scientific literature review, genomic analysis, and drug discovery
• The company has identified hallucinations as a primary technical challenge requiring innovative solutions

Key technological solutions: GSK employs two main computational approaches to minimize AI hallucinations during the inference phase.

• Self-reflection mechanisms allow AI models to critique and edit their own responses through iterative review
• Multi-model sampling uses different AI models or configurations to cross-verify outputs and ensure consistency
• Both strategies require significant computational resources but are deemed essential for maintaining accuracy

Technical implementation: Test-time compute scaling enables more complex operations during the inference phase of AI systems.

• The approach allows for iterative output refinement and multi-model aggregation
• GSK’s literature search tool demonstrates this by collecting data from internal repositories and evaluating findings through self-criticism
• The system compares outputs across different temperature settings and model configurations to confirm conclusions

Infrastructure requirements: Advanced hardware capabilities are essential for supporting GSK’s AI reliability strategies.

• Companies like Cerebras, Groq, and SambaNova are competing to provide the necessary computational infrastructure
• These specialized chips process thousands of tokens per second, enabling complex inferencing routines
• The “inference wars” between hardware providers are driving improvements in token throughput and reduced latency

Industry perspective: Kim Branson, SVP of AI and machine learning at GSK, provides insight into the company’s approach.

• Branson emphasizes the importance of increasing iteration cycles to accelerate research and development
• Self-criticism is identified as the single most important technique for improving AI reliability
• The competition in AI infrastructure is driving down costs per token while enabling new algorithmic strategies

Resource implications: Implementing these solutions requires careful balance of various operational factors.

• Higher compute usage increases operational costs
• Longer inference times can impact workflow efficiency
• GSK views these trade-offs as necessary investments in reliability and functionality

Future trajectory: The evolving landscape of AI compute infrastructure and reliability strategies points to continued advancement in healthcare applications.

• GSK’s approach provides a potential roadmap for other organizations in regulated industries
• The combination of enhanced infrastructure and sophisticated inference techniques sets the stage for future breakthroughs
• Success in managing hallucinations could accelerate progress in drug discovery and patient care

Critical analysis: While GSK’s approach shows promise in addressing AI hallucinations, questions remain about scalability and cost-effectiveness across broader healthcare applications. The success of these strategies may depend heavily on continued advances in computational infrastructure and the development of more efficient verification methods.