A pair of posts about AI and science make adjacent arguments about what is actually slowing research down. Read together, they do not present a single conclusion so much as two different views of where current AI systems run into limits.
In one post, ResearchHub Foundation argued that "AI discovery and hypothesis generation in biology are no longer the bottleneck". Its claim is that models can already propose targets, designs, and mechanisms faster than researchers can check them. The scarcer layer, the post says, is verification: measurements such as protein-protein interactions, drug-target responses in cells, or tissue-level readouts that can confirm or reject what a model proposes.
That framing is specifically about capacity in biology. ResearchHub Foundation’s post argues that the layer needing more funding is wet-lab verification and the measurement systems that produce feedback models can learn from, not hypothesis generation by itself.
Prediction versus insight
A second post, from Zhiruo Wang, argues that scientific discovery is "very different from solving fixed-procedure tasks". Wang writes that agents are getting strong at brute-force search and optimization, but that science also involves finding mechanisms, connecting the dots, and uncovering insights that open new directions.
Wang points to EurekaBench as an example of that gap, writing that agents can beat humans at prediction while still struggling to generate "the insights that keep science moving forward". In that framing, doing well on a defined predictive task is not the same thing as identifying what is worth exploring next.
The two posts are focused on different constraints. ResearchHub Foundation’s post argues that, in biology, models can generate candidate ideas faster than labs can validate them. Wang’s post argues that scientific discovery includes kinds of reasoning and insight that are not captured by faster search or better prediction alone.
That leaves a more limited comparison than a broad claim about whether AI can or cannot "do science." Based on these two posts, one argument is about verification in the wet lab, while the other is about the difference between prediction and scientific insight.
Together, the posts show two distinct ways people are describing AI’s current limits in research, rather than a single agreed view of the field.
