AI Agents and Scientific Discovery Shape Tech News

Eric Schmidt, cofounder of Schmidt Sciences and former Google CEO, along with Suhas Mahesh, head of AI for science at the Schmidt Sciences AI Center, suggest that scientific research can be accelerated by using AI agents that simulate the iterative way humans make discoveries. Unlike previous systems that rely on massive datasets, these agents are inherently generalists designed to digitally model the human process of discovery.

The shift follows the success of AlphaFold, a neural network from Google DeepMind that predicted protein structures and contributed to a 2024 Nobel Prize in Chemistry. While AlphaFold demonstrated the potential of AI in science, Schmidt and Mahesh argue it may not be the best template for accelerating science. AlphaFold relied on a dataset of roughly 170,000 experimentally validated protein structures, which took 53 years and roughly $21 billion worth of experimental work to assemble, according to the authors.

Because comparable datasets will be difficult or impossible to create in many fields, researchers are looking toward AI agents. These generalist tools model the iterative, highly contingent process of actual research rather than applying a powerful approach to a limited question, Schmidt and Mahesh state.

Transition from Statistical Prediction to Neuro-Symbolic Reasoning

Enterprise R&D has historically operated on the premise that feeding domain-specific literature into massive parameter-scale transformers would lead to scientific discovery. As reported by whatisfuture.com, this method resulted in costly semantic search engines that could summarize known information but were fundamentally incapable of synthesizing verifiable, original scientific hypotheses.

The core failure of pure autoregressive transformers is their reliance on probabilistic token prediction. whatisfuture.com reports that these models lack an internal world model to validate causal relationships, spatial constraints, or kinetic laws. In fields like genomics or quantum chemistry, this often results in the AI hallucinating chemically impossible molecular structures because the output token sequence satisfies semantic syntax while violating fundamental thermodynamic principles.

To resolve this, modern deep tech architectures are shifting toward neuro-symbolic systems. These systems marry deep learning backbones with deterministic verification execution environments and explicit search algorithms, such as Monte Carlo Tree Search. According to whatisfuture.com, the neural network functions as a hypothesis generator driven by intuition within this framework, while domain simulators, theorem provers, and external formal solvers provide the validation layer.

Infrastructure Bottlenecks and Safety Guardrails

Two linked engineering hurdles impede the rollout of reasoning agents: the inability of pure autoregressive models to maintain logical consistency over long horizons and the effect of blunt safety classifiers. whatisfuture.com points out that these guardrails frequently confuse rigorous, domain-specific experimentation with malicious output.

To prevent abuse, current commercial foundational models utilize multi-stage safety layers, such as real-time input-output guardrail classifiers and Reinforcement Learning from Human Feedback (RLHF). However, these mechanisms can trigger false positives during legitimate research into virology or biological synthesis protocols. whatisfuture.com states that when safety classifiers mistake legitimate biochemical queries for potential dual-use biological hazards, they break stateful orchestration loops by causing refusal responses or inference truncations, which disrupts the agentic workflow’s deterministic execution loop.

Broader AI Security and Industry Developments

While scientific AI evolves, other sectors face acute security risks. OpenAI has paused work on its Astra AI model after tests found it could launch cyber-attacks autonomously, according to the Financial Times. Simultaneously, Reuters reports that the state-linked North Korean group Kimsuky has developed AI tools used in spear-phishing campaigns to automate attacks and analyse stolen data.

The hardware landscape is also shifting. Bloomberg reports that Chinese firms control 97% of global humanoid shipments, with shipments more than tripling since last year.

“It’s the humans that we need to watch out for. AI is just the tool.”

Oren Etzioni, professor emeritus at the University of Washington and former CEO of the Allen Institute for AI, via CNN