AI Is Starting to Change How Scientific Research Works
Most of the attention around AI still goes to things that are easy to see. A new image model comes out and suddenly the images look more realistic. A coding model gets better and someone posts a video of it building an app in ten minutes.
I follow all of that because I work with AI and software myself, but lately I’ve been much more interested in what is happening in scientific research.
Some AI systems are starting to do more than search papers, summarize information or answer questions. They are beginning to take part in the research process itself.
Google’s Co-Scientist is a good example. It can generate scientific hypotheses, but generating ideas is not really the impressive part anymore. Modern language models can come up with plenty of plausible ideas if you give them a problem.
The difficult part is figuring out which ideas are actually worth testing.
Co-Scientist approaches this by having different parts of the system generate ideas, criticize them, compare similar hypotheses and gradually improve the stronger ones. I find this much more interesting than simply asking an AI for its best answer, especially when language models can still produce something completely wrong while sounding very confident.
In science, eventually someone has to test the idea and see if it actually works.
Researchers have already explored systems like this in areas including cancer research and drug repurposing. This doesn't mean AI has suddenly cured cancer, and I think headlines sometimes make these results sound much more dramatic than they really are.
What matters is that AI can help researchers search through possibilities.
There are an enormous number of drugs, proteins, mutations, cell types and combinations that scientists could investigate. Testing everything in a laboratory is impossible. Experiments cost money and take time.
If AI can help move an unusual but promising candidate higher up the list, researchers can spend more of their time testing better guesses.
That alone could be extremely useful.
Other systems are starting to go further.
Robin, for example, has been used in research where the AI doesn't simply suggest something to test. It can also work with experimental results, help write analysis code, create visualizations and use what happened in one experiment to help decide what should be investigated next.
This is where I think things get particularly interesting.
Scientific research is a loop. You start with an idea, run an experiment, collect data, analyze what happened and then decide what to try next.
A lot of time is spent between those steps.
If AI can shorten that process, researchers may be able to try more ideas in the same amount of time. An experiment that creates another question can quickly lead to another experiment, and several possibilities could potentially be explored at once.
The important development may not be an AI suddenly producing one brilliant scientific answer. It may simply be that researchers can move through this cycle much faster.
Another system I’ve been following is Mammal, which tries to learn relationships between different kinds of biological information.
That makes sense because biology doesn't really care about the categories humans use to organize it. A molecule interacts with a protein, that protein affects a cellular pathway, gene expression changes and the cell behaves differently.
Those relationships cross several areas of biology.
A model that can work across those layers may notice useful connections that would otherwise take researchers much longer to find.
Mammal has been tested on things like predicting how cancer cell lines respond to drugs and generating parts of antibody designs. Again, these are early research results. A drug working on cells in a laboratory is very different from that drug working safely in a human patient.
But that isn't really the point.
The interesting part is that AI can help researchers decide where to look.
When you put systems like these together, you can start to imagine how scientific research might change.
A researcher begins with a problem. AI searches existing research and suggests possible explanations. Another model helps rank compounds or generate potential designs. Researchers test the most promising candidates. The resulting data goes back into the system, which helps analyze what happened and suggests what might be worth trying next.
Then the process repeats.
Laboratory automation could eventually make that loop even faster.
I don't think this means scientists disappear. Someone still needs to understand whether a result makes sense, decide which questions are worth asking and recognize when something unexpected deserves further investigation.
AI is useful here because humans have limits. Researchers specialize, nobody can read every paper ever published, and there are simply too many possible experiments to test everything.
The machine doesn't need to become the scientist.
It can help scientists search a much larger space.
There are obviously risks as these systems become more capable. Models hallucinate, datasets contain biases and experimental results don't always reproduce. Something that works in cultured cells can fail in animals, and something that works in animals can still fail in humans.
Biology is also an area where increasingly capable AI systems need serious safety controls and human oversight.
So I don't think every new AI biology paper should be treated as a breakthrough.
But I do think the overall direction is worth paying attention to.
A few years ago, the interesting question was whether AI could answer scientific questions.
Now researchers are beginning to explore whether AI can help decide which question should be asked next.
For me, that is a much more important development.
The biggest impact of AI on science may not eventually be one famous AI-discovered drug. It may be thousands of smaller improvements to the way research is done, allowing scientists to test ideas, learn from the results and move on to the next experiment faster.
We are still very early, but AI becoming part of the scientific process isn't really a future prediction anymore.
It has already started.
Full video : https://www.youtube.com/watch?v=rwujwQn3wOg&t=1102s
I build and experiment with AI and software to help my Siamese Cat Cafe :https://siamesecat.cafe/ which could help you saving hundreds of dollar every year on SEO : https://www.youtube.com/watch?v=DNlpgX7MAv4&t=5s
I also created the [Siamese Cat Dev course at DJAI Academy : https://djai.academy/siamese_cat/dev/course, where I focus on learning how to actually build with AI by connecting models, code, data and tools instead of simply learning whichever AI product happens to be popular right now.
Top comments (0)