Ideas in AI: How GSK is tackling the biggest challenge in drug development

Generative AI has become a key factor in many industries, and healthcare is no exception. However, as organizations want GSK pushing the boundaries of what generative AI can achieve, they face serious challenges – especially when it comes to reliability. It’s a nightmareor when AI models generate false or false information, a constant challenge for advanced applications such as drug discovery and healthcare. For GSK, overcoming these challenges requires improving the time frame of the experiment to improve the performance of gen AI. This is how they are doing it.

The problem of visualization in automated health care

Medical applications require a high degree of accuracy and reliability. Mistakes are not only disruptive; can have life-changing effects. This makes modeling in major languages ​​(LLMs) a critical issue for companies such as GSK, where gen AI is used for applications such as scientific literature review, genomic analysis and drug discovery.

To reduce speculation, GSK uses advanced statistical methods, including self-report methods, multivariate sampling and retrospective analysis. According to Kim Branson, SvP of AI and machine learning (ML) at GSK, these methods help ensure that agents are “robust and reliable,” while helping scientists generate actionable insights quickly.

Use to calculate test time

A time trial test refers to the probability of add readings during the inference phase of AI systems. This allows for more complex tasks, such as repeated output processing or combining multiple models, which is very important to reduce the simulation and improve the performance of the models.

Branson emphasized the transformative role in expanding GSK’s AI efforts, saying “we all want to increase GSK’s iteration — how fast we think.” By using methods such as self-reporting and collective modeling, GSK can improve these additional processes to produce more accurate and reliable results.

Branson also touched on how the industry is expanding, saying, “You’re seeing the battle going on with how much I can serve, my price per token and time per token. This allows people to bring in different algorithmic methods that were technically impossible, and this will also facilitate the deployment and adoption of agents.”

Ways to reduce hallucinations

GSK has identified hallucinations as a major problem in gen AI for healthcare. The company uses two main methods that require additional equipment for calculation. The use of best-in-class methods ensures that each solution is evaluated for accuracy and consistency before being administered in clinical or research settings, where reliability is critical.

Self-reflection and repeat results

One important process is self-evaluation, where LLMs criticize or modify their answers for the better. The model “thinks slowly,” analyzing its initial results, pointing out weaknesses and repeating solutions as needed. GSK’s document search tool provides an example of this: It collects data from LLM’s internal databases and memory, then self-critically analyzes the findings to reveal inconsistencies.

This iteration results in clearer and more detailed final answers. Branson emphasized the importance of self-criticism, saying: “If you can do one thing, do it.” Processing its own assumptions before delivering results allows the system to generate data that conforms to strict medical standards.

Multi-model sampling

The second GSK method relies on multiple LLMs or single gene mutations to confirm the results. Instead, the system can answer the same question at different temperatures to provide different answers, use similar versions of the same model set in other domains or order very different models trained on different datasets.

Comparing and contrasting these results helps to establish common or related points. “You can get results with different approaches to the same idea,” Branson said. Although this method requires a lot of computational power, it reduces guesswork and increases confidence in the final solution – an important benefit in high-end clinical settings.

Imaginary wars

GSK’s systems rely on infrastructure that can handle heavy loads. In what Branson calls the “imagination wars,” companies developing AI tools — such as The brainGroq and SambaNova – compete to provide hardware breakthroughs that accelerate tokenization, reduce latency and reduce cost per token.

The unique chips and architecture support complex control systems, including multi-mode sampling and self-replicating, at scale. Cerebras’ technology, for example, processes thousands of tokens per second, allowing advanced processes to work in real-time situations. “You see the impact of this innovation on how we can better use clinical samples,” Branson said.

When hardware meets software requirements, solutions emerge to be more accurate and efficient.

Problems still exist

Even with these advancements, the expansion of computing resources presents obstacles. Long lead times can slow down the workflow, especially if clinicians or researchers need results quickly. Extensive use of computers also increases costs, which require careful management. However, GSK considers these trade-offs necessary for reliability and rich functionality.

“The more devices we support in the ecosystem, the more efficient the system is for people, and the more computers you can use,” Branson said. Balancing operational, financial and strategic capabilities enables GSK to maintain an efficient yet forward-looking approach.

What’s next?

GSK plans to continue to innovate its AI-driven healthcare solutions with time-to-trial testing as a key priority. The combination of self-representation, multi-dimensional models and robust architecture helps to ensure that the production models meet the needs of the clinical community.

The strategy also serves as a road map for other organizations, showing how to balance accuracy, efficiency and control. Keeping the future ahead of computer innovations and advanced referral systems not only addresses current challenges, but also lays the foundation for success in drug discovery, patient care and beyond.