We thought the problem was not having enough ideas. Somewhere along the way, it became not having enough time to test them.
Producing a hundred candidates is no longer difficult. The difficulty is choosing one, starting the equipment, using the materials, and waiting for a result. An idea that weighed almost nothing on a screen acquires weight as soon as it enters the laboratory.
Bojana Ranković and colleagues confront that weight in their work on GOLLuM. The question is not what the AI knows, but what it will try next after only a few observations. The study moves the place where we look for intelligence. [1]
Knowing a subject, knowing how sure you are
Change the conditions of a chemical reaction and search for a better outcome. One useful approach is Bayesian optimization. It uses earlier experiments to predict what might happen under conditions not yet tested. At the same time, it estimates how uncertain those predictions are.
Should we dig in a promising place, or investigate one we do not understand? Both kinds of value inform the choice of the next experiment.
Representing experimental conditions, however, often requires specialist ingenuity. Language models can accept those conditions as text. That sounds convenient, but knowing about something in words is not the same as being able to predict its experimental outcome with well-grounded confidence.
GOLLuM does not simply place these two approaches side by side. It connects a language model's numerical representations of conditions to a probabilistic model called a Gaussian process, adjusting them together through learning signals involving outcomes and uncertainty. The representation itself changes so that experiments with similar outcomes sit closer together inside the model.
This is different from asking a model to “be more cautious.” It is not learning a cautious tone. It is updating how it forms a forecast.
Starting from a few poor results
The published study compares methods across 23 tasks in chemistry, materials, and molecular design. Each search begins with ten low-performing observations and proceeds with a limited experimental budget. It is not a fortunate starting point. [1]
At a budget of 50 experiments, top-5% coverage, measuring the share of evaluated experiments whose outcomes fall in the top 5%, was 36.3% for GOLLuM and 29.7% for conventional Bayesian optimization using specialist representations. GOLLuM reached the conventional method's final performance with a median of 41% fewer iterations.
This does not mean laboratory costs fell by 41%. These were benchmark searches using existing data and other benchmark resources, not measurements of equipment queues, procurement, or washing up after an unsuccessful experiment. Experiment counts and research budgets are close relations, not the same person.
Even so, the comparison is clear. Not how many millions of ideas were generated, but how far a limited number of trials moved the search toward promising territory. Not the volume of candidates, but the yield of selection.
Returning a wrong prediction to the next decision
Moving from here to the organization of work is an association, not a finding of the study.
When a proposal fails, where does the record go? Good ideas stay in the presentation; unsuccessful ones disappear. The system producing the next proposal still has every reason to offer the same attractive idea again. The previous failure sleeps quietly somewhere in the minutes.
Learning requires more than a paragraph of reflection. What did we expect? What happened? How will the difference inform the next choice? A poor result is one piece of information. A poor result we had been confident would be good is another.
This research cannot simply be converted into a procedure for management decisions. Business has multiple criteria, and an experiment may change the market it is testing. Its candidates do not necessarily line up as neatly as the rows in a table of chemical conditions.
The technology has limits too. Standard Gaussian-process computation grows cubically with the number of data points, so scaling to large experiments is a separate challenge. Subjects such as complex three-dimensional structures also need representations that words alone may not adequately provide. This is not a declaration that anything describable in language can be solved.
What is interesting, rather, is that the language model was not made an all-purpose judge. It was connected to an existing optimization framework, allowed to change its representations through prediction errors, and sent back to choose again. Its role rests less on possessing knowledge than on changing in response to what the world returns.
Before asking for the next hundred ideas, does our AI know that the last one failed?