Artificial intelligence could help deliver major breakthroughs in physics, biotechnology, chemistry and materials science within the next two to three years, according to TikTok CEO Shou Zi Chew.

Speaking at the DBS Asia Future Forward Conference in Singapore, Chew described scientific research as one of AI’s most promising applications. His outlook reflects growing investment in systems designed to help researchers understand biological structures, explore new materials and accelerate discovery.

The timeline remains a prediction. Turning advances in AI models into validated scientific results will depend on experimental evidence, reliable research processes and the ability to translate computational findings into practical applications.

AI research moves beyond consumer products

Chew’s comments come as ByteDance, TikTok’s parent company, expands its AI research beyond content platforms and consumer tools.

In April, its Seed research team introduced Protenix-v2, a model focused on biomolecular structure prediction and design, with potential applications in drug discovery. Chew also pointed to efforts to collaborate with scientists, describing scientific progress as an increasingly important area of exploration.

These initiatives illustrate a broader ambition for AI: helping researchers investigate problems where the number of possible structures, interactions or outcomes makes conventional exploration difficult and time-consuming.

For drug discovery, for example, better predictions could help researchers identify promising candidates for further investigation. Experimental validation would still be essential to establish whether those predictions lead to useful results.

A rapid pace of progress, an uncertain timeline

Chew argued that AI’s immediate impact may be overestimated while its longer-term potential remains underappreciated. In his view, the pace of progress in scientific applications suggests that the technology is still at an early stage of what it could eventually achieve.

That distinction matters for business and technology leaders. A promising research model, a validated discovery and a production-ready application represent different stages of development. Progress at one stage does not guarantee an equally rapid transition to the next.

Organizations exploring AI therefore need to assess both its potential and the work required to make its outputs dependable. Data quality, evaluation methods, integration and human expertise all influence whether a new capability produces meaningful results.

Greater autonomy brings greater responsibility

Alongside his optimism, Chew raised concerns about aligning increasingly capable AI systems with human expectations and ethical boundaries.

One immediate issue is the level of access users grant to AI agents. Systems authorized to take actions on a user’s behalf can create unintended consequences when their permissions exceed the scope of the task.

Chew used autonomous purchasing as an example: an agent trying to complete an assignment could buy more than the user intended.

For engineering teams, this makes authorization a central design decision. An agent needs clearly defined limits on what it can access, change or spend, along with approval requirements for consequential actions. Its activity also needs to be traceable so that teams can investigate errors and intervene when necessary.

Cybersecurity is part of the same discussion

Chew also warned that more capable AI models are reducing the cost of fraudulent activity, creating additional security challenges for digital platforms.

As malicious actors gain access to cheaper ways of generating deceptive content and scaling attacks, companies need to reassess how they detect abuse and protect their systems. Adopting AI introduces another layer of responsibility when models connect to internal data, external services or operational workflows.

Security decisions must therefore accompany decisions about capability and access from the beginning.

What this means for software teams

For Control F5 Software, the practical question is how to turn promising AI capabilities into systems that organizations can trust, operate and improve over time.

That requires clear use cases, appropriate architecture, measurable evaluation and explicit ownership of outcomes. When AI can take action, teams also need permission boundaries, monitoring and ways to recover from mistakes.

Chew’s forecast points to an ambitious future for AI in science. The engineering work needed to support that future is already relevant today: building systems whose outputs can be checked and whose actions remain under control.

Source