Congratulations, You’ve Invented Something. Now What?

Consider the plastic tray left after lunch. Not the most promising place to look for the future of intelligence. There is probably a reluctant tomato in it.

Yet in April 2022, researchers reported an enzyme, engineered with help from machine learning, that could break down PET from 51 different postconsumer thermoformed products almost completely within a week. They also demonstrated making PET again from recovered chemical building blocks. The enzyme was called FAST-PETase. [1]

That is a proper achievement: a computational suggestion meeting a physical experiment, followed by a measurable result. It deserves considerably more respect than another chatbot confidently inventing a court case.

It also leaves a rather large question. How do you turn that result into a reliable recycling service?

Congratulations, you have invented something. The world would now like to know where it plugs in, what it costs and whether it still works on Thursday.

As computational intelligence makes more discoveries possible, the journey after discovery deserves just as much attention. The exciting prospect is that the same broad family of tools can help us interpret a result, test its limits, engineer a process and learn from its use. We should be building that journey, rather than merely accelerating the production of things waiting to enter it.

FAST-PETase was not an enzyme that ate every kind of plastic, anywhere, under any conditions. The paper concerns PET and specific experimental conditions. Its whole-water-bottle demonstration involved thermal pretreatment. That detail matters rather a lot if your sales presentation contains a photograph of the Pacific Ocean. [1]

The first translation job is therefore from a scientific result into a useful claim. What worked? Against which comparison? With what preparation? Which part remains a hypothesis?

A competent computational assistant could help map those conditions, compare earlier research and identify applications whose requirements actually fit. Perhaps a discovery suits one carefully sorted industrial waste stream before it suits the contents of everyone’s recycling bin. That is a question to investigate, not an embarrassment to conceal.

The valuable output would be an application brief with traceable evidence: a plausible customer, a measurable advantage, the conditions required and the next experiment that could kill the idea. A machine that generates twenty flattering markets for a discovery has produced a brochure. A machine that helps eliminate nineteen unsuitable ones may have saved a company.

Once we have a plausible use, we need to discover how the result fails. Change the feedstock. Repeat the measurement. Test the awkward batch. Check whether the improvement survives outside the conditions that made the original graph so attractive.

This is where the familiar objection appears: even if AI proposes discoveries faster, there are only so many people and laboratories available to test them. True. But it does not follow that the testing capacity must remain fixed.

In a 2020 Nature paper, a University of Liverpool team described a mobile robotic chemist that ran 688 experiments over eight days. A search algorithm varied ten experimental parameters, finding photocatalyst mixtures six times more active than the initial formulations. It was a bounded optimisation task, not an autonomous professor with tenure and an increasingly eccentric office. [2]

The important mechanism is the connection: propose an experiment, perform it, measure the outcome, use the result to choose the next one. That loop can bring computation into the work of testing, rather than leaving it outside the laboratory issuing increasingly impatient suggestions.

Nor does automation remove the need for independent checks. A fast loop connected to a misleading measurement can become exceptionally efficient at producing the wrong answer. The instruments, objective and success criteria deserve as much scrutiny as the intelligence choosing the next move.

For our recycling idea, a useful enzyme is one component of a process. Someone must obtain suitable waste, prepare it, produce the enzyme, control the reaction, recover useful material and establish what happens to everything left over. Reliability and economics belong in that sentence too.

To see the scale of the industrial task, consider Carbios, which is developing a separate enzymatic PET-recycling process. Its research lineage includes engineered leaf-branch compost cutinase; it is not the commercial continuation of FAST-PETase. [3]

In its 24 September 2026 half-year announcement, Carbios said financing for its Longlaville plant was still being finalised before construction could resume. Credit approvals represented progress, but other due diligence remained underway. In August, the company had already said a financing delay would push production beyond its previous first-half-2028 expectation. These are company reports and forecasts, not evidence of an operating plant. [4, 5]

There is no need to sneer at that gap. Building industrial capacity is substantial work. The lesson is simply that a genuine scientific advance can coexist with unresolved financing and construction. Nature does not issue a purchase order when it accepts your paper.

Computational intelligence could help here too: compare process configurations, explore operating conditions, analyse failures and test cost assumptions. But those proposals need engineering models grounded in measurements. A digital twin that has never met its physical counterpart is, at best, a promising pen pal.

The handovers are easy to underestimate. A scientist reports performance under particular conditions. An engineer designs around it. A commercial team estimates a price. Somewhere between those desks, “achieved in this experiment” can quietly become “guaranteed at scale”. Nobody quite remembers authorising the promotion.

I would give each promising discovery a living evidence record. Keep the original result, its limits, the application assumptions, the failed tests and the reasons for each design decision together. When a condition changes, identify which downstream conclusions need another look.

That is a more demanding job than summarising a paper. It requires access to the actual experiment records, versioned designs and operating results. It also gives an assistant something useful to do between bursts of inspiration: notice that the commercial forecast still assumes the yield the pilot stopped achieving three weeks ago.

At the point of use, the loop should continue. Measure whether the process meets the customer’s requirements in ordinary operation. Record interventions, failures and costs, including inconvenient ones. Feed those observations back into the next design.

The goal is a shorter distance between learning something and changing the right thing. The customer should not have to become an honorary research fellow merely to obtain the advertised result.

My expectation is that some of the most valuable uses of computational intelligence will sit between discoveries and their applications. Tools that connect scientific evidence to engineering decisions, and operating results back to research, could become more valuable as the supply of promising discoveries grows.

That is a forecast, not a claim that a universal automated invention factory already exists. Some delays are genuinely physical. Some require decisions about safety, investment or who bears the consequences. Others are slow because information, equipment and people are poorly connected. We should stop treating all three as the same immovable obstacle.

For a business, the practical starting point is one promising result and one intended use. Name the unanswered question that currently blocks progress. Decide what evidence would resolve it. Then ask whether computation can help obtain that evidence, or whether the immediate need is a test rig, a partner, an independent check or a customer willing to run a bounded pilot.

Measure time to the next trustworthy decision. Do not settle for counting generated ideas. We already possess a formidable global infrastructure for producing suggestions nobody implements. It is called the meeting.

The hopeful outcome is not an endless procession of miraculous announcements. It is more useful things reaching people who need them, with fewer good ideas stranded between a paper and a purchase order.

Back to the lunch tray. Breaking it down was a discovery worth celebrating. Building a dependable way to do something useful with millions of them is the next adventure.

The tomato can wait.